Skip to content
Stale · last refresh 45 h ago

Spend · My stack (Plan) · 89 complete rigs from 171 of 209 catalog parts · recomputed in 7577 ms

NVIDIA DGX Station GB300 (Grace Blackwell Ultra), 748GB is the most expensive machine this catalog can build, at $106,625 — first of 81 rigs ranked on the price floor (a lower bound on landed cost). No rig in this catalog meets “My stack (Plan)”.

NO RIG MEETS THE PLANNO CONFIDENCE LEVEL
8,739 price observations on record, 4 of them carrying a shipping charge. 81 rigs ranked on total cost, 8 listed without a rank. Every number below expands to the arithmetic that produced it.
Since yesterdayNo change — the same rig leads this ranking and not one of the 81 rigs costable on both days moved.compared with prices observed on or before Oct 19, 1:14 AM ET (24 h ago)
What moved
  • 81 of 89 rigs had a price floor on both days; the rest were unpriced on one of them and are not counted as unchanged.
  • No rig's cost moved in that window.
  • The leader 24 h ago was NVIDIA DGX Station GB300 (Grace Blackwell Ultra), 748GB at $106,625.

How these numbers are costed

81 of 89 at landed cost · 81 of 89 at the price floor

Price floor. You asked for the price floor. 81 of 89 rigs have a full landed cost; the floor excludes shipping entirely and is a lower bound on landed cost, not an estimate of it.

Cost basis

Price + NJ sales tax, with shipping EXCLUDED because no listing states one. Shipping is never negative, so this is a strict lower bound on landed cost, not an estimate of it. You chose this basis. Landed cost is computable for 81 of 89 rigs; the price floor for 81 of 89.

Rank by

Ranked by total cost, most expensive first — SPEC §15.11 target 1 ranks complete rigs by total landed cost descending.

Cost per tok/s measured on

The only decoding model in the active Plan. A rig that cannot load it has no cost per tok/s, and the cell says which band it fell in.

Complete rigs, ranked

81 ranked by total cost, largest first · 0 meet the Plan
Every complete rig the configurator can build from this catalog, ranked by total cost (largest first), with the cost basis, the three SPEC §8.4 value ratios and the provenance of each rig's price
# rankRigPlanPrice floor (the ranking key)+ 5 yr power$ / GB usable$ / GB/s b/w$ / tok/s decode
1NVIDIA DGX Station GB300 (Grace Blackwell Ultra), 748GBsealed appliance · 1 part · break down its parts4 BLOCKERS
$106,625 (derived)

DERIVEDNVIDIA DGX Station GB300 (Grace Blackwell Ultra), 748GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$106,625 = $106,625

NVIDIA DGX Station GB300 (Grace Blackwell Ultra), 748GB
10662499 ¢ — Central Computers · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
$111,921 (estimate)

ESTIMATENVIDIA DGX Station GB300 (Grace Blackwell Ultra), 748GB: purchase plus five years of powerArithmetic that also depends on a tunable assumption, listed below.

purchase_plus_5yr_power = purchase + five_year_power

purchase_plus_5yr_power = $106,624.99 + $5,295.68 = $111,920.67

purchase_cents
10662499 ¢ — rig_price_floor = Σ part price_floor × quantity
five_year_power_cents
529568 ¢ — five_year_power = annual_kwh × ¢_per_kwh × 5
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Watts are at the wall (PSU conversion losses included by the caller).
  • Load and idle draws are averages over their hours, not peaks.
  • Duty cycle (Settings → Power): 6 h/day at load, 18 h/day idle.
  • 365 days per year (leap days ignored, −0.07%).
  • Flat rate for the whole period: no escalation, no time-of-use or tiered pricing, no demand charges.
  • Rounded to the nearest cent, half-up, once from the unrounded annual cost.
$154.34 (derived)

DERIVEDNVIDIA DGX Station GB300 (Grace Blackwell Ultra), 748GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $106,625 / 690.84 GB = $154.34 /GB

price_floor
106624.99 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
690.84 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$15.02 (derived)

DERIVEDNVIDIA DGX Station GB300 (Grace Blackwell Ultra), 748GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $106,625 / 7,100 GB/s = $15.02 per GB/s

price_floor
106624.99 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
7100 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$7,556 (estimate)

ESTIMATENVIDIA DGX Station GB300 (Grace Blackwell Ultra), 748GB: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $106,625 / 14.11 tok/s = $7,555.89 per tok/s

price_floor
106624.99 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
14.11 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU HBM3e (1.58 ms), CPU LPDDR5X (0 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced fromCentral Computerscondition newaccess tier 1, route jsonld_product
2Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMDIY workstation · 7 parts · break down its parts4 BLOCKERS
$71,596 (derived)

DERIVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $49,900 + $19,192 + $75 + $192 + $704 = $71,596

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM
4990040 ¢ — Newegg · new
NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB
1919249 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: idle_watts, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$67.17 (derived)

DERIVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $71,596 / 1,065.92 GB = $67.17 /GB

price_floor
71595.81 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
1065.92 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$39.95 (derived)

DERIVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $71,596 / 1,792 GB/s = $39.95 per GB/s

price_floor
71595.81 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
1792 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$10,748 (estimate)

ESTIMATEHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $71,596 / 6.66 tok/s = $10,747.87 per tok/s

price_floor
71595.81 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
6.66 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
3Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMDIY workstation · 7 parts · break down its parts4 BLOCKERS
$65,124 (derived)

DERIVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $49,900 + $12,721 + $75 + $192 + $704 = $65,124

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM
4990040 ¢ — Newegg · new
NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB
1272101 ¢ — SabrePC · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: idle_watts, NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$61.10 (derived)

DERIVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $65,124 / 1,065.92 GB = $61.1 /GB

price_floor
65124.33 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
1065.92 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$36.34 (derived)

DERIVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $65,124 / 1,792 GB/s = $36.34 per GB/s

price_floor
65124.33 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
1792 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$9,776 (estimate)

ESTIMATEHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $65,124 / 6.66 tok/s = $9,776.38 per tok/s

price_floor
65124.33 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
6.66 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
4Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMDIY workstation · 7 parts · break down its parts4 BLOCKERS
$57,841 (derived)

DERIVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $49,900 + $5,438 + $75 + $192 + $704 = $57,841

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM
4990040 ¢ — Newegg · new
MSI GeForce RTX 5090 32G GAMING TRIO OC
543786 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: idle_watts, MSI GeForce RTX 5090 32G GAMING TRIO OC: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$57.62 (derived)

DERIVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $57,841 / 1,003.84 GB = $57.62 /GB

price_floor
57841.18 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
1003.84 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$32.28 (derived)

DERIVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $57,841 / 1,792 GB/s = $32.28 per GB/s

price_floor
57841.18 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
1792 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$9,685 (estimate)

ESTIMATEHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $57,841 / 5.97 tok/s = $9,685.05 per tok/s

price_floor
57841.18 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
5.97 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
5Hybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMDIY workstation · 7 parts · break down its parts5 BLOCKERS
$54,323 (derived)

DERIVEDHybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $49,900 + $1,919 + $75 + $192 + $704 = $54,323

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM
4990040 ¢ — Newegg · new
Sapphire Radeon AI PRO R9700 32GB (32358-01-20G)
191924 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: idle_watts, Sapphire Radeon AI PRO R9700 32GB (32358-01-20G): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$54.11 (derived)

DERIVEDHybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $54,323 / 1,003.84 GB = $54.11 /GB

price_floor
54322.56 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
1003.84 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$84.88 (derived)

DERIVEDHybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $54,323 / 640 GB/s = $84.88 per GB/s

price_floor
54322.56 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
640 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$9,292 (estimate)

ESTIMATEHybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $54,323 / 5.85 tok/s = $9,292.44 per tok/s

price_floor
54322.56 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
5.85 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
6Hybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMDIY workstation · 7 parts · break down its parts5 BLOCKERS
$54,228 (derived)

DERIVEDHybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $49,900 + $1,824 + $75 + $192 + $704 = $54,228

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM
4990040 ¢ — Newegg · new
GIGABYTE AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD)
182424 ¢ — SabrePC · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: idle_watts, GIGABYTE AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$54.02 (derived)

DERIVEDHybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $54,228 / 1,003.84 GB = $54.02 /GB

price_floor
54227.56 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
1003.84 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$84.73 (derived)

DERIVEDHybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $54,228 / 640 GB/s = $84.73 per GB/s

price_floor
54227.56 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
640 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$9,276 (estimate)

ESTIMATEHybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $54,228 / 5.85 tok/s = $9,276.19 per tok/s

price_floor
54227.56 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
5.85 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
7Hybrid · GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMDIY workstation · 7 parts · break down its parts4 BLOCKERS
$54,216 (derived)

DERIVEDHybrid · GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $49,900 + $1,813 + $75 + $192 + $704 = $54,216

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM
4990040 ¢ — Newegg · new
GIGABYTE GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD)
181261 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: idle_watts, GIGABYTE GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$54.86 (derived)

DERIVEDHybrid · GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $54,216 / 988.32 GB = $54.86 /GB

price_floor
54215.93 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
988.32 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$56.47 (derived)

DERIVEDHybrid · GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $54,216 / 960 GB/s = $56.47 per GB/s

price_floor
54215.93 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
960 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$9,360 (estimate)

ESTIMATEHybrid · GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $54,216 / 5.79 tok/s = $9,360.06 per tok/s

price_floor
54215.93 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
5.79 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
8Hybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMDIY workstation · 7 parts · break down its parts5 BLOCKERS
$54,216 (derived)

DERIVEDHybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $49,900 + $1,813 + $75 + $192 + $704 = $54,216

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM
4990040 ¢ — Newegg · new
ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT)
181261 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: idle_watts, ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$54.01 (derived)

DERIVEDHybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $54,216 / 1,003.84 GB = $54.01 /GB

price_floor
54215.93 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
1003.84 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$84.71 (derived)

DERIVEDHybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $54,216 / 640 GB/s = $84.71 per GB/s

price_floor
54215.93 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
640 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$9,274 (estimate)

ESTIMATEHybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $54,216 / 5.85 tok/s = $9,274.2 per tok/s

price_floor
54215.93 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
5.85 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
9Hybrid · GeForce RTX 5080 16G SHADOW 3X OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMDIY workstation · 7 parts · break down its parts4 BLOCKERS
$54,205 (derived)

DERIVEDHybrid · GeForce RTX 5080 16G SHADOW 3X OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $49,900 + $1,802 + $75 + $192 + $704 = $54,205

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM
4990040 ¢ — Newegg · new
MSI GeForce RTX 5080 16G SHADOW 3X OC
180195 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: idle_watts, MSI GeForce RTX 5080 16G SHADOW 3X OC: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$54.85 (derived)

DERIVEDHybrid · GeForce RTX 5080 16G SHADOW 3X OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $54,205 / 988.32 GB = $54.85 /GB

price_floor
54205.27 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
988.32 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$56.46 (derived)

DERIVEDHybrid · GeForce RTX 5080 16G SHADOW 3X OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $54,205 / 960 GB/s = $56.46 per GB/s

price_floor
54205.27 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
960 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$9,358 (estimate)

ESTIMATEHybrid · GeForce RTX 5080 16G SHADOW 3X OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $54,205 / 5.79 tok/s = $9,358.22 per tok/s

price_floor
54205.27 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
5.79 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
10Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMMDIY workstation · 7 parts · break down its parts4 BLOCKERS
$54,109 (derived)

DERIVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $32,414 + $19,192 + $75 + $192 + $704 = $54,109

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM
3241392 ¢ — Newegg · new
NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB
1919249 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: idle_watts, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$65.77 (derived)

DERIVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $54,109 / 822.72 GB = $65.77 /GB

price_floor
54109.33 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
822.72 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$30.19 (derived)

DERIVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $54,109 / 1,792 GB/s = $30.19 per GB/s

price_floor
54109.33 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
1792 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$8,123 (estimate)

ESTIMATEHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $54,109 / 6.66 tok/s = $8,122.83 per tok/s

price_floor
54109.33 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
6.66 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
11CPU · EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMDIY workstation · 6 parts · break down its parts5 BLOCKERS
$52,403 (derived)

DERIVEDCPU · EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $49,900 + $75 + $192 + $704 = $52,403

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM
4990040 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 6 of 6 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$53.87 (derived)

DERIVEDCPU · EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $52,403 / 972.8 GB = $53.87 /GB

price_floor
52403.32 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
972.8 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 6 of 6 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$170.58 (derived)

DERIVEDCPU · EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $52,403 / 307.2 GB/s = $170.58 per GB/s

price_floor
52403.32 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
307.2 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 6 of 6 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$9,228 (estimate)

ESTIMATECPU · EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $52,403 / 5.68 tok/s = $9,228.43 per tok/s

price_floor
52403.32 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
5.68 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 6 of 6 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from6 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
12Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMMDIY workstation · 7 parts · break down its parts4 BLOCKERS
$47,638 (derived)

DERIVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $32,414 + $12,721 + $75 + $192 + $704 = $47,638

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM
3241392 ¢ — Newegg · new
NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB
1272101 ¢ — SabrePC · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: idle_watts, NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$57.90 (derived)

DERIVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $47,638 / 822.72 GB = $57.9 /GB

price_floor
47637.85 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
822.72 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$26.58 (derived)

DERIVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $47,638 / 1,792 GB/s = $26.58 per GB/s

price_floor
47637.85 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
1792 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$7,151 (estimate)

ESTIMATEHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $47,638 / 6.66 tok/s = $7,151.33 per tok/s

price_floor
47637.85 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
6.66 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
133× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts4 BLOCKERS
$42,494 (derived)

DERIVED3× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $1,828 + $38,163 + $75 + $192 + $704 = $42,494

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM
182776 ¢ — SabrePC · new
NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB
3816303 ¢ — SabrePC · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: idle_watts, NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$81.32 (derived)

DERIVED3× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $42,494 / 522.56 GB = $81.32 /GB

price_floor
42493.71 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
522.56 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$23.71 (derived)

DERIVED3× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $42,494 / 1,792 GB/s = $23.71 per GB/s

price_floor
42493.71 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
1792 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$4,171 (estimate)

ESTIMATE3× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $42,494 / 10.19 tok/s = $4,170.69 per tok/s

price_floor
42493.71 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
10.19 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • layer split across 3 GPUs: capacity adds, per-token bandwidth does not (layers run sequentially)
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
14Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMMDIY workstation · 7 parts · break down its parts4 BLOCKERS
$40,355 (derived)

DERIVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $32,414 + $5,438 + $75 + $192 + $704 = $40,355

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM
3241392 ¢ — Newegg · new
MSI GeForce RTX 5090 32G GAMING TRIO OC
543786 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: idle_watts, MSI GeForce RTX 5090 32G GAMING TRIO OC: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$53.05 (derived)

DERIVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $40,355 / 760.64 GB = $53.05 /GB

price_floor
40354.7 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
760.64 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$22.52 (derived)

DERIVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $40,355 / 1,792 GB/s = $22.52 per GB/s

price_floor
40354.7 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
1792 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$6,757 (estimate)

ESTIMATEHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $40,355 / 5.97 tok/s = $6,757.08 per tok/s

price_floor
40354.7 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
5.97 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
15Hybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMMDIY workstation · 7 parts · break down its parts5 BLOCKERS
$36,836 (derived)

DERIVEDHybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $32,414 + $1,919 + $75 + $192 + $704 = $36,836

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM
3241392 ¢ — Newegg · new
Sapphire Radeon AI PRO R9700 32GB (32358-01-20G)
191924 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: idle_watts, Sapphire Radeon AI PRO R9700 32GB (32358-01-20G): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$48.43 (derived)

DERIVEDHybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $36,836 / 760.64 GB = $48.43 /GB

price_floor
36836.08 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
760.64 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$57.56 (derived)

DERIVEDHybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $36,836 / 640 GB/s = $57.56 per GB/s

price_floor
36836.08 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
640 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$6,301 (estimate)

ESTIMATEHybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $36,836 / 5.85 tok/s = $6,301.2 per tok/s

price_floor
36836.08 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
5.85 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
16Hybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMMDIY workstation · 7 parts · break down its parts5 BLOCKERS
$36,741 (derived)

DERIVEDHybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $32,414 + $1,824 + $75 + $192 + $704 = $36,741

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM
3241392 ¢ — Newegg · new
GIGABYTE AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD)
182424 ¢ — SabrePC · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: idle_watts, GIGABYTE AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$48.30 (derived)

DERIVEDHybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $36,741 / 760.64 GB = $48.3 /GB

price_floor
36741.08 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
760.64 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$57.41 (derived)

DERIVEDHybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $36,741 / 640 GB/s = $57.41 per GB/s

price_floor
36741.08 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
640 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$6,285 (estimate)

ESTIMATEHybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $36,741 / 5.85 tok/s = $6,284.95 per tok/s

price_floor
36741.08 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
5.85 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
17Hybrid · GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMMDIY workstation · 7 parts · break down its parts4 BLOCKERS
$36,729 (derived)

DERIVEDHybrid · GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $32,414 + $1,813 + $75 + $192 + $704 = $36,729

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM
3241392 ¢ — Newegg · new
GIGABYTE GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD)
181261 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: idle_watts, GIGABYTE GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$49.29 (derived)

DERIVEDHybrid · GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $36,729 / 745.12 GB = $49.29 /GB

price_floor
36729.45 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
745.12 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$38.26 (derived)

DERIVEDHybrid · GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $36,729 / 960 GB/s = $38.26 per GB/s

price_floor
36729.45 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
960 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$6,341 (estimate)

ESTIMATEHybrid · GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $36,729 / 5.79 tok/s = $6,341.12 per tok/s

price_floor
36729.45 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
5.79 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
18Hybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMMDIY workstation · 7 parts · break down its parts5 BLOCKERS
$36,729 (derived)

DERIVEDHybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $32,414 + $1,813 + $75 + $192 + $704 = $36,729

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM
3241392 ¢ — Newegg · new
ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT)
181261 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: idle_watts, ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$48.29 (derived)

DERIVEDHybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $36,729 / 760.64 GB = $48.29 /GB

price_floor
36729.45 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
760.64 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$57.39 (derived)

DERIVEDHybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $36,729 / 640 GB/s = $57.39 per GB/s

price_floor
36729.45 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
640 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$6,283 (estimate)

ESTIMATEHybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $36,729 / 5.85 tok/s = $6,282.96 per tok/s

price_floor
36729.45 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
5.85 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
19Hybrid · GeForce RTX 5080 16G SHADOW 3X OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMMDIY workstation · 7 parts · break down its parts4 BLOCKERS
$36,719 (derived)

DERIVEDHybrid · GeForce RTX 5080 16G SHADOW 3X OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $32,414 + $1,802 + $75 + $192 + $704 = $36,719

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM
3241392 ¢ — Newegg · new
MSI GeForce RTX 5080 16G SHADOW 3X OC
180195 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: idle_watts, MSI GeForce RTX 5080 16G SHADOW 3X OC: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$49.28 (derived)

DERIVEDHybrid · GeForce RTX 5080 16G SHADOW 3X OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $36,719 / 745.12 GB = $49.28 /GB

price_floor
36718.79 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
745.12 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$38.25 (derived)

DERIVEDHybrid · GeForce RTX 5080 16G SHADOW 3X OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $36,719 / 960 GB/s = $38.25 per GB/s

price_floor
36718.79 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
960 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$6,339 (estimate)

ESTIMATEHybrid · GeForce RTX 5080 16G SHADOW 3X OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $36,719 / 5.79 tok/s = $6,339.28 per tok/s

price_floor
36718.79 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
5.79 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
20CPU · EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMMDIY workstation · 6 parts · break down its parts5 BLOCKERS
$34,917 (derived)

DERIVEDCPU · EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $32,414 + $75 + $192 + $704 = $34,917

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM
3241392 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 6 of 6 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$47.86 (derived)

DERIVEDCPU · EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $34,917 / 729.6 GB = $47.86 /GB

price_floor
34916.84 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
729.6 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 6 of 6 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$113.66 (derived)

DERIVEDCPU · EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $34,917 / 307.2 GB/s = $113.66 per GB/s

price_floor
34916.84 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
307.2 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 6 of 6 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$6,149 (estimate)

ESTIMATECPU · EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $34,917 / 5.68 tok/s = $6,148.99 per tok/s

price_floor
34916.84 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
5.68 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 6 of 6 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from6 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
212× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts5 BLOCKERS
$29,773 (derived)

DERIVED2× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $1,828 + $25,442 + $75 + $192 + $704 = $29,773

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM
182776 ¢ — SabrePC · new
NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB
2544202 ¢ — SabrePC · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: idle_watts, NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$69.33 (derived)

DERIVED2× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $29,773 / 429.44 GB = $69.33 /GB

price_floor
29772.7 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
429.44 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$16.61 (derived)

DERIVED2× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $29,773 / 1,792 GB/s = $16.61 per GB/s

price_floor
29772.7 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
1792 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: 2× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + AMD EPYC 9015 has 429.44 GB usable — it will not load, so there is no decode speed.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
22Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts4 BLOCKERS
$24,229 (derived)

DERIVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $2,534 + $19,192 + $75 + $192 + $704 = $24,229

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM
253400 ¢ — SabrePC · new
NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB
1919249 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: idle_watts, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$41.81 (derived)

DERIVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $24,229 / 579.52 GB = $41.81 /GB

price_floor
24229.41 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
579.52 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$13.52 (derived)

DERIVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $24,229 / 1,792 GB/s = $13.52 per GB/s

price_floor
24229.41 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
1792 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$3,637 (estimate)

ESTIMATEHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $24,229 / 6.66 tok/s = $3,637.29 per tok/s

price_floor
24229.41 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
6.66 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
231× RTX PRO 6000 Blackwell Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts5 BLOCKERS
$23,523 (derived)

DERIVED1× RTX PRO 6000 Blackwell Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $1,828 + $19,192 + $75 + $192 + $704 = $23,523

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM
182776 ¢ — SabrePC · new
NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB
1919249 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: idle_watts, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$69.94 (derived)

DERIVED1× RTX PRO 6000 Blackwell Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $23,523 / 336.32 GB = $69.94 /GB

price_floor
23523.17 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
336.32 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$13.13 (derived)

DERIVED1× RTX PRO 6000 Blackwell Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $23,523 / 1,792 GB/s = $13.13 per GB/s

price_floor
23523.17 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
1792 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB + AMD EPYC 9015 has 336.32 GB usable — it will not load, so there is no decode speed.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
24Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts4 BLOCKERS
$17,758 (derived)

DERIVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $2,534 + $12,721 + $75 + $192 + $704 = $17,758

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM
253400 ¢ — SabrePC · new
NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB
1272101 ¢ — SabrePC · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: idle_watts, NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$30.64 (derived)

DERIVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $17,758 / 579.52 GB = $30.64 /GB

price_floor
17757.93 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
579.52 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$9.91 (derived)

DERIVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $17,758 / 1,792 GB/s = $9.91 per GB/s

price_floor
17757.93 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
1792 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$2,666 (estimate)

ESTIMATEHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $17,758 / 6.66 tok/s = $2,665.8 per tok/s

price_floor
17757.93 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
6.66 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
251× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts5 BLOCKERS
$17,052 (derived)

DERIVED1× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $1,828 + $12,721 + $75 + $192 + $704 = $17,052

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM
182776 ¢ — SabrePC · new
NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB
1272101 ¢ — SabrePC · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: idle_watts, NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$50.70 (derived)

DERIVED1× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $17,052 / 336.32 GB = $50.7 /GB

price_floor
17051.69 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
336.32 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$9.52 (derived)

DERIVED1× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $17,052 / 1,792 GB/s = $9.52 per GB/s

price_floor
17051.69 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
1792 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + AMD EPYC 9015 has 336.32 GB usable — it will not load, so there is no decode speed.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
262× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts5 BLOCKERS
$15,206 (derived)

DERIVED2× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $1,828 + $10,876 + $75 + $192 + $704 = $15,206

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM
182776 ¢ — SabrePC · new
MSI GeForce RTX 5090 32G GAMING TRIO OC
1087572 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: idle_watts, MSI GeForce RTX 5090 32G GAMING TRIO OC: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$49.81 (derived)

DERIVED2× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $15,206 / 305.28 GB = $49.81 /GB

price_floor
15206.4 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
305.28 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$8.49 (derived)

DERIVED2× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $15,206 / 1,792 GB/s = $8.49 per GB/s

price_floor
15206.4 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
1792 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: 2× MSI GeForce RTX 5090 32G GAMING TRIO OC + AMD EPYC 9015 has 305.28 GB usable — it will not load, so there is no decode speed.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
27Apple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 256GBsealed appliance · 1 part · break down its parts5 BLOCKERS
$11,514 (derived)

DERIVEDApple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 256GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$11,514 = $11,514

Apple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 256GB
1151443 ¢ — Apple · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
$12,709 (estimate)

ESTIMATEApple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 256GB: purchase plus five years of powerArithmetic that also depends on a tunable assumption, listed below.

purchase_plus_5yr_power = purchase + five_year_power

purchase_plus_5yr_power = $11,514.43 + $1,194.47 = $12,708.90

purchase_cents
1151443 ¢ — rig_price_floor = Σ part price_floor × quantity
five_year_power_cents
119447 ¢ — five_year_power = annual_kwh × ¢_per_kwh × 5
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Watts are at the wall (PSU conversion losses included by the caller).
  • Load and idle draws are averages over their hours, not peaks.
  • Duty cycle (Settings → Power): 6 h/day at load, 18 h/day idle.
  • 365 days per year (leap days ignored, −0.07%).
  • Flat rate for the whole period: no escalation, no time-of-use or tiered pricing, no demand charges.
  • Rounded to the nearest cent, half-up, once from the unrounded annual cost.
$59.97 (derived)

DERIVEDApple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 256GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $11,514 / 192 GB = $59.97 /GB

price_floor
11514.43 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
192 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$9.60 (derived)

DERIVEDApple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 256GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $11,514 / 1,200 GB/s = $9.6 per GB/s

price_floor
11514.43 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
1200 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but Apple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 256GB has 192 GB usable — it will not load, so there is no decode speed.
Priced fromApplecondition newaccess tier 3, route apple_store
28Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts4 BLOCKERS
$10,475 (derived)

DERIVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $2,534 + $5,438 + $75 + $192 + $704 = $10,475

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM
253400 ¢ — SabrePC · new
MSI GeForce RTX 5090 32G GAMING TRIO OC
543786 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: idle_watts, MSI GeForce RTX 5090 32G GAMING TRIO OC: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$20.24 (derived)

DERIVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $10,475 / 517.44 GB = $20.24 /GB

price_floor
10474.78 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
517.44 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$5.85 (derived)

DERIVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $10,475 / 1,792 GB/s = $5.85 per GB/s

price_floor
10474.78 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
1792 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$1,754 (estimate)

ESTIMATEHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $10,475 / 5.97 tok/s = $1,753.92 per tok/s

price_floor
10474.78 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
5.97 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
29Apple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GBsealed appliance · 1 part · broken down below5 BLOCKERS
$10,128 (derived)

DERIVEDApple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$10,128 = $10,128

Apple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB
1012831 ¢ — Apple · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Idle draw is not published for 1 of 1 part in this rig, and the duty cycle in Settings spends most of the day idle — counting an unknown idle draw as zero would understate the bill. (missing: Apple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB: idle_watts)
$52.75 (derived)

DERIVEDApple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $10,128 / 192 GB = $52.75 /GB

price_floor
10128.31 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
192 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$8.44 (derived)

DERIVEDApple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $10,128 / 1,200 GB/s = $8.44 per GB/s

price_floor
10128.31 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
1200 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but Apple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB has 192 GB usable — it will not load, so there is no decode speed.
Priced fromApplecondition newaccess tier 3, route apple_store
303× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts6 BLOCKERS
$10,088 (derived)

DERIVED3× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $1,828 + $5,758 + $75 + $192 + $704 = $10,088

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM
182776 ¢ — SabrePC · new
Sapphire Radeon AI PRO R9700 32GB (32358-01-20G)
575772 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: idle_watts, Sapphire Radeon AI PRO R9700 32GB (32358-01-20G): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$30.00 (derived)

DERIVED3× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $10,088 / 336.32 GB = $30 /GB

price_floor
10088.4 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
336.32 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$15.76 (derived)

DERIVED3× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $10,088 / 640 GB/s = $15.76 per GB/s

price_floor
10088.4 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
640 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: 3× Sapphire Radeon AI PRO R9700 32GB (32358-01-20G) + AMD EPYC 9015 has 336.32 GB usable — it will not load, so there is no decode speed.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
313× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts6 BLOCKERS
$9,803 (derived)

DERIVED3× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $1,828 + $5,473 + $75 + $192 + $704 = $9,803

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM
182776 ¢ — SabrePC · new
GIGABYTE AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD)
547272 ¢ — SabrePC · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: idle_watts, GIGABYTE AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$29.15 (derived)

DERIVED3× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $9,803 / 336.32 GB = $29.15 /GB

price_floor
9803.4 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
336.32 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$15.32 (derived)

DERIVED3× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $9,803 / 640 GB/s = $15.32 per GB/s

price_floor
9803.4 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
640 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: 3× GIGABYTE AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + AMD EPYC 9015 has 336.32 GB usable — it will not load, so there is no decode speed.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
321× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts5 BLOCKERS
$9,769 (derived)

DERIVED1× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $1,828 + $5,438 + $75 + $192 + $704 = $9,769

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM
182776 ¢ — SabrePC · new
MSI GeForce RTX 5090 32G GAMING TRIO OC
543786 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: idle_watts, MSI GeForce RTX 5090 32G GAMING TRIO OC: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$35.62 (derived)

DERIVED1× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $9,769 / 274.24 GB = $35.62 /GB

price_floor
9768.54 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
274.24 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$5.45 (derived)

DERIVED1× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $9,769 / 1,792 GB/s = $5.45 per GB/s

price_floor
9768.54 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
1792 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: MSI GeForce RTX 5090 32G GAMING TRIO OC + AMD EPYC 9015 has 274.24 GB usable — it will not load, so there is no decode speed.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
333× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts5 BLOCKERS
$9,769 (derived)

DERIVED3× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $1,828 + $5,438 + $75 + $192 + $704 = $9,769

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM
182776 ¢ — SabrePC · new
GIGABYTE GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD)
543783 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: idle_watts, GIGABYTE GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$33.71 (derived)

DERIVED3× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $9,769 / 289.76 GB = $33.71 /GB

price_floor
9768.51 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
289.76 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$10.18 (derived)

DERIVED3× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $9,769 / 960 GB/s = $10.18 per GB/s

price_floor
9768.51 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
960 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: 3× GIGABYTE GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) + AMD EPYC 9015 has 289.76 GB usable — it will not load, so there is no decode speed.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
343× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts6 BLOCKERS
$9,769 (derived)

DERIVED3× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $1,828 + $5,438 + $75 + $192 + $704 = $9,769

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM
182776 ¢ — SabrePC · new
ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT)
543783 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: idle_watts, ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$29.05 (derived)

DERIVED3× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $9,769 / 336.32 GB = $29.05 /GB

price_floor
9768.51 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
336.32 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$15.26 (derived)

DERIVED3× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $9,769 / 640 GB/s = $15.26 per GB/s

price_floor
9768.51 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
640 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: 3× ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT) + AMD EPYC 9015 has 336.32 GB usable — it will not load, so there is no decode speed.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
353× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts5 BLOCKERS
$9,737 (derived)

DERIVED3× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $1,828 + $5,406 + $75 + $192 + $704 = $9,737

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM
182776 ¢ — SabrePC · new
MSI GeForce RTX 5080 16G SHADOW 3X OC
540585 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: idle_watts, MSI GeForce RTX 5080 16G SHADOW 3X OC: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$33.60 (derived)

DERIVED3× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $9,737 / 289.76 GB = $33.6 /GB

price_floor
9736.53 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
289.76 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$10.14 (derived)

DERIVED3× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $9,737 / 960 GB/s = $10.14 per GB/s

price_floor
9736.53 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
960 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: 3× MSI GeForce RTX 5080 16G SHADOW 3X OC + AMD EPYC 9015 has 289.76 GB usable — it will not load, so there is no decode speed.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
36GIGABYTE AI TOP ATOM (NVIDIA GB10), 128GBsealed appliance · 1 part · break down its parts5 BLOCKERS
$8,530 (derived)

DERIVEDGIGABYTE AI TOP ATOM (NVIDIA GB10), 128GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$8,530 = $8,530

GIGABYTE AI TOP ATOM (NVIDIA GB10), 128GB
852999 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
$9,498 (estimate)

ESTIMATEGIGABYTE AI TOP ATOM (NVIDIA GB10), 128GB: purchase plus five years of powerArithmetic that also depends on a tunable assumption, listed below.

purchase_plus_5yr_power = purchase + five_year_power

purchase_plus_5yr_power = $8,529.99 + $967.93 = $9,497.92

purchase_cents
852999 ¢ — rig_price_floor = Σ part price_floor × quantity
five_year_power_cents
96793 ¢ — five_year_power = annual_kwh × ¢_per_kwh × 5
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Watts are at the wall (PSU conversion losses included by the caller).
  • Load and idle draws are averages over their hours, not peaks.
  • Duty cycle (Settings → Power): 6 h/day at load, 18 h/day idle.
  • 365 days per year (leap days ignored, −0.07%).
  • Flat rate for the whole period: no escalation, no time-of-use or tiered pricing, no demand charges.
  • Rounded to the nearest cent, half-up, once from the unrounded annual cost.
$74.05 (derived)

DERIVEDGIGABYTE AI TOP ATOM (NVIDIA GB10), 128GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $8,530 / 115.2 GB = $74.05 /GB

price_floor
8529.99 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
115.2 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$31.25 (derived)

DERIVEDGIGABYTE AI TOP ATOM (NVIDIA GB10), 128GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $8,530 / 273 GB/s = $31.25 per GB/s

price_floor
8529.99 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
273 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but GIGABYTE AI TOP ATOM (NVIDIA GB10), 128GB has 115.2 GB usable — it will not load, so there is no decode speed.
Priced fromNeweggcondition newaccess tier 3, route newegg_realtime
37NVIDIA DGX Spark (GB10 Grace Blackwell), 128GBsealed appliance · 1 part · break down its parts5 BLOCKERS
$8,423 (derived)

DERIVEDNVIDIA DGX Spark (GB10 Grace Blackwell), 128GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$8,423 = $8,423

NVIDIA DGX Spark (GB10 Grace Blackwell), 128GB
842336 ¢ — Central Computers · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
$9,391 (estimate)

ESTIMATENVIDIA DGX Spark (GB10 Grace Blackwell), 128GB: purchase plus five years of powerArithmetic that also depends on a tunable assumption, listed below.

purchase_plus_5yr_power = purchase + five_year_power

purchase_plus_5yr_power = $8,423.36 + $967.93 = $9,391.29

purchase_cents
842336 ¢ — rig_price_floor = Σ part price_floor × quantity
five_year_power_cents
96793 ¢ — five_year_power = annual_kwh × ¢_per_kwh × 5
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Watts are at the wall (PSU conversion losses included by the caller).
  • Load and idle draws are averages over their hours, not peaks.
  • Duty cycle (Settings → Power): 6 h/day at load, 18 h/day idle.
  • 365 days per year (leap days ignored, −0.07%).
  • Flat rate for the whole period: no escalation, no time-of-use or tiered pricing, no demand charges.
  • Rounded to the nearest cent, half-up, once from the unrounded annual cost.
$73.12 (derived)

DERIVEDNVIDIA DGX Spark (GB10 Grace Blackwell), 128GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $8,423 / 115.2 GB = $73.12 /GB

price_floor
8423.36 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
115.2 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$30.85 (derived)

DERIVEDNVIDIA DGX Spark (GB10 Grace Blackwell), 128GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $8,423 / 273 GB/s = $30.85 per GB/s

price_floor
8423.36 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
273 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but NVIDIA DGX Spark (GB10 Grace Blackwell), 128GB has 115.2 GB usable — it will not load, so there is no decode speed.
Priced fromCentral Computerscondition newaccess tier 1, route jsonld_product
382× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts6 BLOCKERS
$8,169 (derived)

DERIVED2× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $1,828 + $3,838 + $75 + $192 + $704 = $8,169

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM
182776 ¢ — SabrePC · new
Sapphire Radeon AI PRO R9700 32GB (32358-01-20G)
383848 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: idle_watts, Sapphire Radeon AI PRO R9700 32GB (32358-01-20G): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$26.76 (derived)

DERIVED2× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $8,169 / 305.28 GB = $26.76 /GB

price_floor
8169.16 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
305.28 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$12.76 (derived)

DERIVED2× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $8,169 / 640 GB/s = $12.76 per GB/s

price_floor
8169.16 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
640 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: 2× Sapphire Radeon AI PRO R9700 32GB (32358-01-20G) + AMD EPYC 9015 has 305.28 GB usable — it will not load, so there is no decode speed.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
39HP Z2 Mini G1a (Ryzen AI Max+ PRO 395), 128GBsealed appliance · 1 part · break down its parts6 BLOCKERS
$8,060 (derived)

DERIVEDHP Z2 Mini G1a (Ryzen AI Max+ PRO 395), 128GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$8,060 = $8,060

HP Z2 Mini G1a (Ryzen AI Max+ PRO 395), 128GB
805978 ¢ — CDW · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Idle draw is not published for 1 of 1 part in this rig, and the duty cycle in Settings spends most of the day idle — counting an unknown idle draw as zero would understate the bill. (missing: HP Z2 Mini G1a (Ryzen AI Max+ PRO 395), 128GB: idle_watts)
$69.96 (derived)

DERIVEDHP Z2 Mini G1a (Ryzen AI Max+ PRO 395), 128GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $8,060 / 115.2 GB = $69.96 /GB

price_floor
8059.78 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
115.2 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$31.48 (derived)

DERIVEDHP Z2 Mini G1a (Ryzen AI Max+ PRO 395), 128GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $8,060 / 256 GB/s = $31.48 per GB/s

price_floor
8059.78 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
256 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but HP Z2 Mini G1a (Ryzen AI Max+ PRO 395), 128GB has 115.2 GB usable — it will not load, so there is no decode speed.
Priced fromCDWcondition newaccess tier 2, route cdw_html
402× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts6 BLOCKERS
$7,979 (derived)

DERIVED2× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $1,828 + $3,648 + $75 + $192 + $704 = $7,979

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM
182776 ¢ — SabrePC · new
GIGABYTE AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD)
364848 ¢ — SabrePC · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: idle_watts, GIGABYTE AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$26.14 (derived)

DERIVED2× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $7,979 / 305.28 GB = $26.14 /GB

price_floor
7979.16 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
305.28 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$12.47 (derived)

DERIVED2× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $7,979 / 640 GB/s = $12.47 per GB/s

price_floor
7979.16 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
640 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: 2× GIGABYTE AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + AMD EPYC 9015 has 305.28 GB usable — it will not load, so there is no decode speed.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
412× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts5 BLOCKERS
$7,956 (derived)

DERIVED2× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $1,828 + $3,625 + $75 + $192 + $704 = $7,956

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM
182776 ¢ — SabrePC · new
GIGABYTE GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD)
362522 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: idle_watts, GIGABYTE GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$29.01 (derived)

DERIVED2× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $7,956 / 274.24 GB = $29.01 /GB

price_floor
7955.9 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
274.24 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$8.29 (derived)

DERIVED2× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $7,956 / 960 GB/s = $8.29 per GB/s

price_floor
7955.9 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
960 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: 2× GIGABYTE GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) + AMD EPYC 9015 has 274.24 GB usable — it will not load, so there is no decode speed.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
422× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts6 BLOCKERS
$7,956 (derived)

DERIVED2× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $1,828 + $3,625 + $75 + $192 + $704 = $7,956

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM
182776 ¢ — SabrePC · new
ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT)
362522 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: idle_watts, ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$26.06 (derived)

DERIVED2× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $7,956 / 305.28 GB = $26.06 /GB

price_floor
7955.9 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
305.28 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$12.43 (derived)

DERIVED2× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $7,956 / 640 GB/s = $12.43 per GB/s

price_floor
7955.9 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
640 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: 2× ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT) + AMD EPYC 9015 has 305.28 GB usable — it will not load, so there is no decode speed.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
432× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts5 BLOCKERS
$7,935 (derived)

DERIVED2× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $1,828 + $3,604 + $75 + $192 + $704 = $7,935

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM
182776 ¢ — SabrePC · new
MSI GeForce RTX 5080 16G SHADOW 3X OC
360390 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: idle_watts, MSI GeForce RTX 5080 16G SHADOW 3X OC: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$28.93 (derived)

DERIVED2× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $7,935 / 274.24 GB = $28.93 /GB

price_floor
7934.58 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
274.24 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$8.27 (derived)

DERIVED2× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $7,935 / 960 GB/s = $8.27 per GB/s

price_floor
7934.58 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
960 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: 2× MSI GeForce RTX 5080 16G SHADOW 3X OC + AMD EPYC 9015 has 274.24 GB usable — it will not load, so there is no decode speed.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
44Apple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 96GBsealed appliance · 1 part · break down its parts5 BLOCKERS
$7,249 (derived)

DERIVEDApple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 96GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$7,249 = $7,249

Apple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 96GB
724943 ¢ — Apple · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
$8,444 (estimate)

ESTIMATEApple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 96GB: purchase plus five years of powerArithmetic that also depends on a tunable assumption, listed below.

purchase_plus_5yr_power = purchase + five_year_power

purchase_plus_5yr_power = $7,249.43 + $1,194.47 = $8,443.90

purchase_cents
724943 ¢ — rig_price_floor = Σ part price_floor × quantity
five_year_power_cents
119447 ¢ — five_year_power = annual_kwh × ¢_per_kwh × 5
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Watts are at the wall (PSU conversion losses included by the caller).
  • Load and idle draws are averages over their hours, not peaks.
  • Duty cycle (Settings → Power): 6 h/day at load, 18 h/day idle.
  • 365 days per year (leap days ignored, −0.07%).
  • Flat rate for the whole period: no escalation, no time-of-use or tiered pricing, no demand charges.
  • Rounded to the nearest cent, half-up, once from the unrounded annual cost.
$100.69 (derived)

DERIVEDApple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 96GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $7,249 / 72 GB = $100.69 /GB

price_floor
7249.43 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
72 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$6.04 (derived)

DERIVEDApple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 96GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $7,249 / 1,200 GB/s = $6.04 per GB/s

price_floor
7249.43 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
1200 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but Apple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 96GB has 72 GB usable — it will not load, so there is no decode speed.
Priced fromApplecondition newaccess tier 3, route apple_store
45Hybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts5 BLOCKERS
$6,956 (derived)

DERIVEDHybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $2,534 + $1,919 + $75 + $192 + $704 = $6,956

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM
253400 ¢ — SabrePC · new
Sapphire Radeon AI PRO R9700 32GB (32358-01-20G)
191924 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: idle_watts, Sapphire Radeon AI PRO R9700 32GB (32358-01-20G): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$13.44 (derived)

DERIVEDHybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $6,956 / 517.44 GB = $13.44 /GB

price_floor
6956.16 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
517.44 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$10.87 (derived)

DERIVEDHybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $6,956 / 640 GB/s = $10.87 per GB/s

price_floor
6956.16 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
640 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$1,190 (estimate)

ESTIMATEHybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $6,956 / 5.85 tok/s = $1,189.92 per tok/s

price_floor
6956.16 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
5.85 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
46Acer Veriton VGN100-UD11 (NVIDIA GB10), 128GBsealed appliance · 1 part · break down its parts5 BLOCKERS
$6,951 (derived)

DERIVEDAcer Veriton VGN100-UD11 (NVIDIA GB10), 128GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$6,951 = $6,951

Acer Veriton VGN100-UD11 (NVIDIA GB10), 128GB
695117 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
$7,919 (estimate)

ESTIMATEAcer Veriton VGN100-UD11 (NVIDIA GB10), 128GB: purchase plus five years of powerArithmetic that also depends on a tunable assumption, listed below.

purchase_plus_5yr_power = purchase + five_year_power

purchase_plus_5yr_power = $6,951.17 + $967.93 = $7,919.10

purchase_cents
695117 ¢ — rig_price_floor = Σ part price_floor × quantity
five_year_power_cents
96793 ¢ — five_year_power = annual_kwh × ¢_per_kwh × 5
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Watts are at the wall (PSU conversion losses included by the caller).
  • Load and idle draws are averages over their hours, not peaks.
  • Duty cycle (Settings → Power): 6 h/day at load, 18 h/day idle.
  • 365 days per year (leap days ignored, −0.07%).
  • Flat rate for the whole period: no escalation, no time-of-use or tiered pricing, no demand charges.
  • Rounded to the nearest cent, half-up, once from the unrounded annual cost.
$60.34 (derived)

DERIVEDAcer Veriton VGN100-UD11 (NVIDIA GB10), 128GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $6,951 / 115.2 GB = $60.34 /GB

price_floor
6951.17 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
115.2 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$25.46 (derived)

DERIVEDAcer Veriton VGN100-UD11 (NVIDIA GB10), 128GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $6,951 / 273 GB/s = $25.46 per GB/s

price_floor
6951.17 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
273 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but Acer Veriton VGN100-UD11 (NVIDIA GB10), 128GB has 115.2 GB usable — it will not load, so there is no decode speed.
Priced fromNeweggcondition newaccess tier 3, route newegg_realtime
47Hybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts5 BLOCKERS
$6,861 (derived)

DERIVEDHybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $2,534 + $1,824 + $75 + $192 + $704 = $6,861

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM
253400 ¢ — SabrePC · new
GIGABYTE AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD)
182424 ¢ — SabrePC · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: idle_watts, GIGABYTE AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$13.26 (derived)

DERIVEDHybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $6,861 / 517.44 GB = $13.26 /GB

price_floor
6861.16 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
517.44 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$10.72 (derived)

DERIVEDHybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $6,861 / 640 GB/s = $10.72 per GB/s

price_floor
6861.16 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
640 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$1,174 (estimate)

ESTIMATEHybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $6,861 / 5.85 tok/s = $1,173.67 per tok/s

price_floor
6861.16 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
5.85 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
48Hybrid · GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts5 BLOCKERS
$6,850 (derived)

DERIVEDHybrid · GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $2,534 + $1,813 + $75 + $192 + $704 = $6,850

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM
253400 ¢ — SabrePC · new
GIGABYTE GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD)
181261 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: idle_watts, GIGABYTE GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$13.65 (derived)

DERIVEDHybrid · GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $6,850 / 501.92 GB = $13.65 /GB

price_floor
6849.53 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
501.92 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$7.13 (derived)

DERIVEDHybrid · GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $6,850 / 960 GB/s = $7.13 per GB/s

price_floor
6849.53 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
960 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: GIGABYTE GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) + AMD EPYC 9015 has 501.92 GB usable — it will not load, so there is no decode speed.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
49Hybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts5 BLOCKERS
$6,850 (derived)

DERIVEDHybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $2,534 + $1,813 + $75 + $192 + $704 = $6,850

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM
253400 ¢ — SabrePC · new
ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT)
181261 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: idle_watts, ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$13.24 (derived)

DERIVEDHybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $6,850 / 517.44 GB = $13.24 /GB

price_floor
6849.53 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
517.44 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$10.70 (derived)

DERIVEDHybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $6,850 / 640 GB/s = $10.7 per GB/s

price_floor
6849.53 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
640 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
$1,172 (estimate)

ESTIMATEHybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per tok/sArithmetic that also depends on a tunable assumption, listed below.

cost_per_tok_s = price_floor / decode_tok_s

= $6,850 / 5.85 tok/s = $1,171.68 per tok/s

price_floor
6849.53 USD — rig_price_floor = Σ part price_floor × quantity
decode_tok_s
5.85 tok/s — GLM-5.3 decoding on this rig at the Plan's context
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • MoE: only active parameters are read per token (SPEC §3).
  • Effective bits include embeddings and higher-precision tensors in mixed quants.
  • Assumes active weights are quantized like the file average.
  • KV cache is re-read on every token (DECISIONS D-008).
  • Pools are read sequentially per token, so per-pool times add (harmonic combination).
  • Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device.
  • Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms).
  • Batch size 1 (single user, interactive).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per tok/s is at least this much.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
50Hybrid · GeForce RTX 5080 16G SHADOW 3X OC + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts5 BLOCKERS
$6,839 (derived)

DERIVEDHybrid · GeForce RTX 5080 16G SHADOW 3X OC + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $2,534 + $1,802 + $75 + $192 + $704 = $6,839

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM
253400 ¢ — SabrePC · new
MSI GeForce RTX 5080 16G SHADOW 3X OC
180195 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: idle_watts, MSI GeForce RTX 5080 16G SHADOW 3X OC: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$13.63 (derived)

DERIVEDHybrid · GeForce RTX 5080 16G SHADOW 3X OC + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $6,839 / 501.92 GB = $13.63 /GB

price_floor
6838.87 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
501.92 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$7.12 (derived)

DERIVEDHybrid · GeForce RTX 5080 16G SHADOW 3X OC + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $6,839 / 960 GB/s = $7.12 per GB/s

price_floor
6838.87 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
960 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: MSI GeForce RTX 5080 16G SHADOW 3X OC + AMD EPYC 9015 has 501.92 GB usable — it will not load, so there is no decode speed.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
51ASUS Ascent GX10 (NVIDIA GB10 Grace Blackwell), 128GBsealed appliance · 1 part · break down its parts5 BLOCKERS
$6,396 (derived)

DERIVEDASUS Ascent GX10 (NVIDIA GB10 Grace Blackwell), 128GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$6,396 = $6,396

ASUS Ascent GX10 (NVIDIA GB10 Grace Blackwell), 128GB
639643 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Idle draw is not published for 1 of 1 part in this rig, and the duty cycle in Settings spends most of the day idle — counting an unknown idle draw as zero would understate the bill. (missing: ASUS Ascent GX10 (NVIDIA GB10 Grace Blackwell), 128GB: idle_watts)
$55.52 (derived)

DERIVEDASUS Ascent GX10 (NVIDIA GB10 Grace Blackwell), 128GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $6,396 / 115.2 GB = $55.52 /GB

price_floor
6396.43 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
115.2 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$23.43 (derived)

DERIVEDASUS Ascent GX10 (NVIDIA GB10 Grace Blackwell), 128GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $6,396 / 273 GB/s = $23.43 per GB/s

price_floor
6396.43 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
273 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but ASUS Ascent GX10 (NVIDIA GB10 Grace Blackwell), 128GB has 115.2 GB usable — it will not load, so there is no decode speed.
Priced fromNeweggcondition newaccess tier 3, route newegg_realtime
521× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts6 BLOCKERS
$6,250 (derived)

DERIVED1× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $1,828 + $1,919 + $75 + $192 + $704 = $6,250

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM
182776 ¢ — SabrePC · new
Sapphire Radeon AI PRO R9700 32GB (32358-01-20G)
191924 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: idle_watts, Sapphire Radeon AI PRO R9700 32GB (32358-01-20G): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$22.79 (derived)

DERIVED1× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $6,250 / 274.24 GB = $22.79 /GB

price_floor
6249.92 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
274.24 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$9.77 (derived)

DERIVED1× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $6,250 / 640 GB/s = $9.77 per GB/s

price_floor
6249.92 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
640 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: Sapphire Radeon AI PRO R9700 32GB (32358-01-20G) + AMD EPYC 9015 has 274.24 GB usable — it will not load, so there is no decode speed.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
53HP ZGX Nano G1n AI Station (NVIDIA GB10), 128GBsealed appliance · 1 part · break down its parts5 BLOCKERS
$6,237 (derived)

DERIVEDHP ZGX Nano G1n AI Station (NVIDIA GB10), 128GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$6,237 = $6,237

HP ZGX Nano G1n AI Station (NVIDIA GB10), 128GB
623650 ¢ — HP Store · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
$7,204 (estimate)

ESTIMATEHP ZGX Nano G1n AI Station (NVIDIA GB10), 128GB: purchase plus five years of powerArithmetic that also depends on a tunable assumption, listed below.

purchase_plus_5yr_power = purchase + five_year_power

purchase_plus_5yr_power = $6,236.50 + $967.93 = $7,204.43

purchase_cents
623650 ¢ — rig_price_floor = Σ part price_floor × quantity
five_year_power_cents
96793 ¢ — five_year_power = annual_kwh × ¢_per_kwh × 5
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Watts are at the wall (PSU conversion losses included by the caller).
  • Load and idle draws are averages over their hours, not peaks.
  • Duty cycle (Settings → Power): 6 h/day at load, 18 h/day idle.
  • 365 days per year (leap days ignored, −0.07%).
  • Flat rate for the whole period: no escalation, no time-of-use or tiered pricing, no demand charges.
  • Rounded to the nearest cent, half-up, once from the unrounded annual cost.
$54.14 (derived)

DERIVEDHP ZGX Nano G1n AI Station (NVIDIA GB10), 128GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $6,237 / 115.2 GB = $54.14 /GB

price_floor
6236.5 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
115.2 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$22.84 (derived)

DERIVEDHP ZGX Nano G1n AI Station (NVIDIA GB10), 128GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $6,237 / 273 GB/s = $22.84 per GB/s

price_floor
6236.5 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
273 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but HP ZGX Nano G1n AI Station (NVIDIA GB10), 128GB has 115.2 GB usable — it will not load, so there is no decode speed.
Priced fromHP Storecondition newaccess tier 1, route jsonld_product
541× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts6 BLOCKERS
$6,155 (derived)

DERIVED1× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $1,828 + $1,824 + $75 + $192 + $704 = $6,155

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM
182776 ¢ — SabrePC · new
GIGABYTE AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD)
182424 ¢ — SabrePC · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: idle_watts, GIGABYTE AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$22.44 (derived)

DERIVED1× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $6,155 / 274.24 GB = $22.44 /GB

price_floor
6154.92 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
274.24 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$9.62 (derived)

DERIVED1× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $6,155 / 640 GB/s = $9.62 per GB/s

price_floor
6154.92 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
640 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: GIGABYTE AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + AMD EPYC 9015 has 274.24 GB usable — it will not load, so there is no decode speed.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
551× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts5 BLOCKERS
$6,143 (derived)

DERIVED1× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $1,828 + $1,813 + $75 + $192 + $704 = $6,143

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM
182776 ¢ — SabrePC · new
GIGABYTE GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD)
181261 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: idle_watts, GIGABYTE GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$23.74 (derived)

DERIVED1× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $6,143 / 258.72 GB = $23.74 /GB

price_floor
6143.29 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
258.72 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$6.40 (derived)

DERIVED1× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $6,143 / 960 GB/s = $6.4 per GB/s

price_floor
6143.29 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
960 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: GIGABYTE GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) + AMD EPYC 9015 has 258.72 GB usable — it will not load, so there is no decode speed.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
561× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts6 BLOCKERS
$6,143 (derived)

DERIVED1× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $1,828 + $1,813 + $75 + $192 + $704 = $6,143

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM
182776 ¢ — SabrePC · new
ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT)
181261 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: idle_watts, ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT): idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$22.40 (derived)

DERIVED1× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $6,143 / 274.24 GB = $22.4 /GB

price_floor
6143.29 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
274.24 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$9.60 (derived)

DERIVED1× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $6,143 / 640 GB/s = $9.6 per GB/s

price_floor
6143.29 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
640 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT) + AMD EPYC 9015 has 274.24 GB usable — it will not load, so there is no decode speed.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
571× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMDIY workstation · 7 parts · break down its parts5 BLOCKERS
$6,133 (derived)

DERIVED1× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $1,828 + $1,802 + $75 + $192 + $704 = $6,133

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM
182776 ¢ — SabrePC · new
MSI GeForce RTX 5080 16G SHADOW 3X OC
180195 ¢ — Newegg · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: idle_watts, MSI GeForce RTX 5080 16G SHADOW 3X OC: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$23.70 (derived)

DERIVED1× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $6,133 / 258.72 GB = $23.7 /GB

price_floor
6132.63 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
258.72 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$6.39 (derived)

DERIVED1× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $6,133 / 960 GB/s = $6.39 per GB/s

price_floor
6132.63 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
960 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 7 of 7 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: MSI GeForce RTX 5080 16G SHADOW 3X OC + AMD EPYC 9015 has 258.72 GB usable — it will not load, so there is no decode speed.
Priced from7 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
58Apple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 96GBsealed appliance · 1 part · break down its parts5 BLOCKERS
$5,863 (derived)

DERIVEDApple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 96GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$5,863 = $5,863

Apple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 96GB
586331 ¢ — Apple · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Idle draw is not published for 1 of 1 part in this rig, and the duty cycle in Settings spends most of the day idle — counting an unknown idle draw as zero would understate the bill. (missing: Apple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 96GB: idle_watts)
$81.43 (derived)

DERIVEDApple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 96GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $5,863 / 72 GB = $81.43 /GB

price_floor
5863.31 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
72 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$4.89 (derived)

DERIVEDApple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 96GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $5,863 / 1,200 GB/s = $4.89 per GB/s

price_floor
5863.31 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
1200 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but Apple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 96GB has 72 GB usable — it will not load, so there is no decode speed.
Priced fromApplecondition newaccess tier 3, route apple_store
59Apple Mac Studio (2026), M5 Max 18-core CPU / 40-core GPU, 128GBsealed appliance · 1 part · break down its parts5 BLOCKERS
$5,757 (derived)

DERIVEDApple Mac Studio (2026), M5 Max 18-core CPU / 40-core GPU, 128GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$5,757 = $5,757

Apple Mac Studio (2026), M5 Max 18-core CPU / 40-core GPU, 128GB
575668 ¢ — Apple · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
$7,222 (estimate)

ESTIMATEApple Mac Studio (2026), M5 Max 18-core CPU / 40-core GPU, 128GB: purchase plus five years of powerArithmetic that also depends on a tunable assumption, listed below.

purchase_plus_5yr_power = purchase + five_year_power

purchase_plus_5yr_power = $5,756.68 + $1,465.14 = $7,221.82

purchase_cents
575668 ¢ — rig_price_floor = Σ part price_floor × quantity
five_year_power_cents
146514 ¢ — five_year_power = annual_kwh × ¢_per_kwh × 5
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Watts are at the wall (PSU conversion losses included by the caller).
  • Load and idle draws are averages over their hours, not peaks.
  • Duty cycle (Settings → Power): 6 h/day at load, 18 h/day idle.
  • 365 days per year (leap days ignored, −0.07%).
  • Flat rate for the whole period: no escalation, no time-of-use or tiered pricing, no demand charges.
  • Rounded to the nearest cent, half-up, once from the unrounded annual cost.
$59.97 (derived)

DERIVEDApple Mac Studio (2026), M5 Max 18-core CPU / 40-core GPU, 128GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $5,757 / 96 GB = $59.97 /GB

price_floor
5756.68 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
96 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$9.38 (derived)

DERIVEDApple Mac Studio (2026), M5 Max 18-core CPU / 40-core GPU, 128GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $5,757 / 614 GB/s = $9.38 per GB/s

price_floor
5756.68 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
614 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but Apple Mac Studio (2026), M5 Max 18-core CPU / 40-core GPU, 128GB has 96 GB usable — it will not load, so there is no decode speed.
Priced fromApplecondition newaccess tier 3, route apple_store
60MSI EdgeXpert (NVIDIA GB10), 128GBsealed appliance · 1 part · break down its parts5 BLOCKERS
$5,378 (derived)

DERIVEDMSI EdgeXpert (NVIDIA GB10), 128GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$5,378 = $5,378

MSI EdgeXpert (NVIDIA GB10), 128GB
537789 ¢ — SabrePC · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
$6,346 (estimate)

ESTIMATEMSI EdgeXpert (NVIDIA GB10), 128GB: purchase plus five years of powerArithmetic that also depends on a tunable assumption, listed below.

purchase_plus_5yr_power = purchase + five_year_power

purchase_plus_5yr_power = $5,377.89 + $967.93 = $6,345.82

purchase_cents
537789 ¢ — rig_price_floor = Σ part price_floor × quantity
five_year_power_cents
96793 ¢ — five_year_power = annual_kwh × ¢_per_kwh × 5
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Watts are at the wall (PSU conversion losses included by the caller).
  • Load and idle draws are averages over their hours, not peaks.
  • Duty cycle (Settings → Power): 6 h/day at load, 18 h/day idle.
  • 365 days per year (leap days ignored, −0.07%).
  • Flat rate for the whole period: no escalation, no time-of-use or tiered pricing, no demand charges.
  • Rounded to the nearest cent, half-up, once from the unrounded annual cost.
$46.68 (derived)

DERIVEDMSI EdgeXpert (NVIDIA GB10), 128GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $5,378 / 115.2 GB = $46.68 /GB

price_floor
5377.89 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
115.2 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$19.70 (derived)

DERIVEDMSI EdgeXpert (NVIDIA GB10), 128GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $5,378 / 273 GB/s = $19.7 per GB/s

price_floor
5377.89 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
273 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but MSI EdgeXpert (NVIDIA GB10), 128GB has 115.2 GB usable — it will not load, so there is no decode speed.
Priced fromSabrePCcondition newaccess tier 2, route sabrepc_ngstate
61Lenovo ThinkStation PGX (NVIDIA GB10), 128GBsealed appliance · 1 part · break down its parts5 BLOCKERS
$5,118 (derived)

DERIVEDLenovo ThinkStation PGX (NVIDIA GB10), 128GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$5,118 = $5,118

Lenovo ThinkStation PGX (NVIDIA GB10), 128GB
511799 ¢ — Central Computers · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
$6,086 (estimate)

ESTIMATELenovo ThinkStation PGX (NVIDIA GB10), 128GB: purchase plus five years of powerArithmetic that also depends on a tunable assumption, listed below.

purchase_plus_5yr_power = purchase + five_year_power

purchase_plus_5yr_power = $5,117.99 + $967.93 = $6,085.92

purchase_cents
511799 ¢ — rig_price_floor = Σ part price_floor × quantity
five_year_power_cents
96793 ¢ — five_year_power = annual_kwh × ¢_per_kwh × 5
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Watts are at the wall (PSU conversion losses included by the caller).
  • Load and idle draws are averages over their hours, not peaks.
  • Duty cycle (Settings → Power): 6 h/day at load, 18 h/day idle.
  • 365 days per year (leap days ignored, −0.07%).
  • Flat rate for the whole period: no escalation, no time-of-use or tiered pricing, no demand charges.
  • Rounded to the nearest cent, half-up, once from the unrounded annual cost.
$44.43 (derived)

DERIVEDLenovo ThinkStation PGX (NVIDIA GB10), 128GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $5,118 / 115.2 GB = $44.43 /GB

price_floor
5117.99 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
115.2 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$18.75 (derived)

DERIVEDLenovo ThinkStation PGX (NVIDIA GB10), 128GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $5,118 / 273 GB/s = $18.75 per GB/s

price_floor
5117.99 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
273 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but Lenovo ThinkStation PGX (NVIDIA GB10), 128GB has 115.2 GB usable — it will not load, so there is no decode speed.
Priced fromCentral Computerscondition newaccess tier 1, route jsonld_product
62CPU · EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMMDIY workstation · 6 parts · break down its parts6 BLOCKERS
$5,037 (derived)

DERIVEDCPU · EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $2,534 + $75 + $192 + $704 = $5,037

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM
253400 ¢ — SabrePC · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 6 of 6 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$10.36 (derived)

DERIVEDCPU · EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $5,037 / 486.4 GB = $10.36 /GB

price_floor
5036.92 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
486.4 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 6 of 6 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$16.40 (derived)

DERIVEDCPU · EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $5,037 / 307.2 GB/s = $16.4 per GB/s

price_floor
5036.92 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
307.2 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 6 of 6 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: AMD EPYC 9015 (CPU only) has 486.4 GB usable — it will not load, so there is no decode speed.
Priced from6 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
63CPU · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMDIY workstation · 6 parts · break down its parts6 BLOCKERS
$4,331 (derived)

DERIVEDCPU · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$638 + $895 + $1,828 + $75 + $192 + $704 = $4,331

AMD EPYC 9015
63809 ¢ — SabrePC · new
ASRock Rack GENOAD8QM3-2T/BCM
89458 ¢ — Newegg · new
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM
182776 ¢ — SabrePC · new
ARCTIC Freezer 4U-SP5
7463 ¢ — ARCTIC · new
Phanteks Enthoo Pro 2 Server Edition
19191 ¢ — Newegg · new
Seasonic PRIME PX-2200
70371 ¢ — Newegg · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 6 of 6 priced offers in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition. (missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts, Seasonic PRIME PX-2200: idle_watts)
$17.81 (derived)

DERIVEDCPU · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $4,331 / 243.2 GB = $17.81 /GB

price_floor
4330.68 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
243.2 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 6 of 6 priced offers in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$14.10 (derived)

DERIVEDCPU · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $4,331 / 307.2 GB/s = $14.1 per GB/s

price_floor
4330.68 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
307.2 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 6 of 6 priced offers in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but DIY: AMD EPYC 9015 (CPU only) has 243.2 GB usable — it will not load, so there is no decode speed.
Priced from6 offers across 3 retailersoldestopen the rig below for each part’s retailer, condition and route
64Minisforum MS-S1 MAX (Ryzen AI Max+ 395), 128GBsealed appliance · 1 part · break down its parts6 BLOCKERS
$4,136 (derived)

DERIVEDMinisforum MS-S1 MAX (Ryzen AI Max+ 395), 128GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$4,136 = $4,136

Minisforum MS-S1 MAX (Ryzen AI Max+ 395), 128GB
413598 ¢ — Minisforum · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for Minisforum MS-S1 MAX (Ryzen AI Max+ 395), 128GB. (missing: Minisforum MS-S1 MAX (Ryzen AI Max+ 395), 128GB: load_watts or tdp_watts, Minisforum MS-S1 MAX (Ryzen AI Max+ 395), 128GB: idle_watts)
$35.90 (derived)

DERIVEDMinisforum MS-S1 MAX (Ryzen AI Max+ 395), 128GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $4,136 / 115.2 GB = $35.9 /GB

price_floor
4135.98 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
115.2 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$16.16 (derived)

DERIVEDMinisforum MS-S1 MAX (Ryzen AI Max+ 395), 128GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $4,136 / 256 GB/s = $16.16 per GB/s

price_floor
4135.98 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
256 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but Minisforum MS-S1 MAX (Ryzen AI Max+ 395), 128GB has 115.2 GB usable — it will not load, so there is no decode speed.
Priced fromMinisforumcondition newaccess tier 3, route shopify_store
65GMKtec EVO-X3 (Ryzen AI Max+ 395), 128GBsealed appliance · 1 part · break down its parts6 BLOCKERS
$4,052 (derived)

DERIVEDGMKtec EVO-X3 (Ryzen AI Max+ 395), 128GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$4,052 = $4,052

GMKtec EVO-X3 (Ryzen AI Max+ 395), 128GB
405174 ¢ — GMKtec · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for GMKtec EVO-X3 (Ryzen AI Max+ 395), 128GB. (missing: GMKtec EVO-X3 (Ryzen AI Max+ 395), 128GB: load_watts or tdp_watts, GMKtec EVO-X3 (Ryzen AI Max+ 395), 128GB: idle_watts)
$35.17 (derived)

DERIVEDGMKtec EVO-X3 (Ryzen AI Max+ 395), 128GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $4,052 / 115.2 GB = $35.17 /GB

price_floor
4051.74 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
115.2 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$15.83 (derived)

DERIVEDGMKtec EVO-X3 (Ryzen AI Max+ 395), 128GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $4,052 / 256 GB/s = $15.83 per GB/s

price_floor
4051.74 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
256 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but GMKtec EVO-X3 (Ryzen AI Max+ 395), 128GB has 115.2 GB usable — it will not load, so there is no decode speed.
Priced fromGMKteccondition newaccess tier 3, route shopify_store
66Apple Mac Studio (2026), M5 Max 18-core CPU / 40-core GPU, 64GBsealed appliance · 1 part · break down its parts5 BLOCKERS
$4,051 (derived)

DERIVEDApple Mac Studio (2026), M5 Max 18-core CPU / 40-core GPU, 64GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$4,051 = $4,051

Apple Mac Studio (2026), M5 Max 18-core CPU / 40-core GPU, 64GB
405068 ¢ — Apple · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
$5,516 (estimate)

ESTIMATEApple Mac Studio (2026), M5 Max 18-core CPU / 40-core GPU, 64GB: purchase plus five years of powerArithmetic that also depends on a tunable assumption, listed below.

purchase_plus_5yr_power = purchase + five_year_power

purchase_plus_5yr_power = $4,050.68 + $1,465.14 = $5,515.82

purchase_cents
405068 ¢ — rig_price_floor = Σ part price_floor × quantity
five_year_power_cents
146514 ¢ — five_year_power = annual_kwh × ¢_per_kwh × 5
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Watts are at the wall (PSU conversion losses included by the caller).
  • Load and idle draws are averages over their hours, not peaks.
  • Duty cycle (Settings → Power): 6 h/day at load, 18 h/day idle.
  • 365 days per year (leap days ignored, −0.07%).
  • Flat rate for the whole period: no escalation, no time-of-use or tiered pricing, no demand charges.
  • Rounded to the nearest cent, half-up, once from the unrounded annual cost.
$84.39 (derived)

DERIVEDApple Mac Studio (2026), M5 Max 18-core CPU / 40-core GPU, 64GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $4,051 / 48 GB = $84.39 /GB

price_floor
4050.68 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
48 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$6.60 (derived)

DERIVEDApple Mac Studio (2026), M5 Max 18-core CPU / 40-core GPU, 64GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $4,051 / 614 GB/s = $6.6 per GB/s

price_floor
4050.68 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
614 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but Apple Mac Studio (2026), M5 Max 18-core CPU / 40-core GPU, 64GB has 48 GB usable — it will not load, so there is no decode speed.
Priced fromApplecondition newaccess tier 3, route apple_store
67Minisforum N5 MAX AI NAS (Ryzen AI Max+ 395), 128GBsealed appliance · 1 part · break down its parts6 BLOCKERS
$3,837 (derived)

DERIVEDMinisforum N5 MAX AI NAS (Ryzen AI Max+ 395), 128GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$3,837 = $3,837

Minisforum N5 MAX AI NAS (Ryzen AI Max+ 395), 128GB
383743 ¢ — Minisforum · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
$4,482 (estimate)

ESTIMATEMinisforum N5 MAX AI NAS (Ryzen AI Max+ 395), 128GB: purchase plus five years of powerArithmetic that also depends on a tunable assumption, listed below.

purchase_plus_5yr_power = purchase + five_year_power

purchase_plus_5yr_power = $3,837.43 + $644.31 = $4,481.74

purchase_cents
383743 ¢ — rig_price_floor = Σ part price_floor × quantity
five_year_power_cents
64431 ¢ — five_year_power = annual_kwh × ¢_per_kwh × 5
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Watts are at the wall (PSU conversion losses included by the caller).
  • Load and idle draws are averages over their hours, not peaks.
  • Duty cycle (Settings → Power): 6 h/day at load, 18 h/day idle.
  • 365 days per year (leap days ignored, −0.07%).
  • Flat rate for the whole period: no escalation, no time-of-use or tiered pricing, no demand charges.
  • Rounded to the nearest cent, half-up, once from the unrounded annual cost.
$33.31 (derived)

DERIVEDMinisforum N5 MAX AI NAS (Ryzen AI Max+ 395), 128GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $3,837 / 115.2 GB = $33.31 /GB

price_floor
3837.43 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
115.2 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$14.99 (derived)

DERIVEDMinisforum N5 MAX AI NAS (Ryzen AI Max+ 395), 128GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $3,837 / 256 GB/s = $14.99 per GB/s

price_floor
3837.43 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
256 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but Minisforum N5 MAX AI NAS (Ryzen AI Max+ 395), 128GB has 115.2 GB usable — it will not load, so there is no decode speed.
Priced fromMinisforumcondition newaccess tier 3, route shopify_store
68GMKtec EVO-X2 (Ryzen AI Max+ 395), 128GBsealed appliance · 1 part · break down its parts6 BLOCKERS
$3,732 (derived)

DERIVEDGMKtec EVO-X2 (Ryzen AI Max+ 395), 128GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$3,732 = $3,732

GMKtec EVO-X2 (Ryzen AI Max+ 395), 128GB
373186 ¢ — GMKtec · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Idle draw is not published for 1 of 1 part in this rig, and the duty cycle in Settings spends most of the day idle — counting an unknown idle draw as zero would understate the bill. (missing: GMKtec EVO-X2 (Ryzen AI Max+ 395), 128GB: idle_watts)
$32.39 (derived)

DERIVEDGMKtec EVO-X2 (Ryzen AI Max+ 395), 128GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $3,732 / 115.2 GB = $32.39 /GB

price_floor
3731.86 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
115.2 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$14.58 (derived)

DERIVEDGMKtec EVO-X2 (Ryzen AI Max+ 395), 128GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $3,732 / 256 GB/s = $14.58 per GB/s

price_floor
3731.86 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
256 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but GMKtec EVO-X2 (Ryzen AI Max+ 395), 128GB has 115.2 GB usable — it will not load, so there is no decode speed.
Priced fromGMKteccondition newaccess tier 3, route shopify_store
69Framework Framework Desktop DIY Edition (Ryzen AI Max+ 395), 128GBsealed appliance · 1 part · break down its parts6 BLOCKERS
$3,678 (derived)

DERIVEDFramework Framework Desktop DIY Edition (Ryzen AI Max+ 395), 128GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$3,678 = $3,678

Framework Framework Desktop DIY Edition (Ryzen AI Max+ 395), 128GB
367750 ¢ — Framework · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Idle draw is not published for 1 of 1 part in this rig, and the duty cycle in Settings spends most of the day idle — counting an unknown idle draw as zero would understate the bill. (missing: Framework Framework Desktop DIY Edition (Ryzen AI Max+ 395), 128GB: idle_watts)
$31.92 (derived)

DERIVEDFramework Framework Desktop DIY Edition (Ryzen AI Max+ 395), 128GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $3,678 / 115.2 GB = $31.92 /GB

price_floor
3677.5 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
115.2 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$14.37 (derived)

DERIVEDFramework Framework Desktop DIY Edition (Ryzen AI Max+ 395), 128GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $3,678 / 256 GB/s = $14.37 per GB/s

price_floor
3677.5 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
256 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but Framework Framework Desktop DIY Edition (Ryzen AI Max+ 395), 128GB has 115.2 GB usable — it will not load, so there is no decode speed.
Priced fromFrameworkcondition newaccess tier 2, route framework_html
70Apple Mac mini (2026), M5 Pro 18-core CPU / 20-core GPU, 64GBsealed appliance · 1 part · break down its parts5 BLOCKERS
$3,411 (derived)

DERIVEDApple Mac mini (2026), M5 Pro 18-core CPU / 20-core GPU, 64GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$3,411 = $3,411

Apple Mac mini (2026), M5 Pro 18-core CPU / 20-core GPU, 64GB
341093 ¢ — Apple · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
$3,911 (estimate)

ESTIMATEApple Mac mini (2026), M5 Pro 18-core CPU / 20-core GPU, 64GB: purchase plus five years of powerArithmetic that also depends on a tunable assumption, listed below.

purchase_plus_5yr_power = purchase + five_year_power

purchase_plus_5yr_power = $3,410.93 + $500.15 = $3,911.08

purchase_cents
341093 ¢ — rig_price_floor = Σ part price_floor × quantity
five_year_power_cents
50015 ¢ — five_year_power = annual_kwh × ¢_per_kwh × 5
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Watts are at the wall (PSU conversion losses included by the caller).
  • Load and idle draws are averages over their hours, not peaks.
  • Duty cycle (Settings → Power): 6 h/day at load, 18 h/day idle.
  • 365 days per year (leap days ignored, −0.07%).
  • Flat rate for the whole period: no escalation, no time-of-use or tiered pricing, no demand charges.
  • Rounded to the nearest cent, half-up, once from the unrounded annual cost.
$71.06 (derived)

DERIVEDApple Mac mini (2026), M5 Pro 18-core CPU / 20-core GPU, 64GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $3,411 / 48 GB = $71.06 /GB

price_floor
3410.93 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
48 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$11.11 (derived)

DERIVEDApple Mac mini (2026), M5 Pro 18-core CPU / 20-core GPU, 64GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $3,411 / 307 GB/s = $11.11 per GB/s

price_floor
3410.93 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
307 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but Apple Mac mini (2026), M5 Pro 18-core CPU / 20-core GPU, 64GB has 48 GB usable — it will not load, so there is no decode speed.
Priced fromApplecondition newaccess tier 3, route apple_store
71Apple Mac Studio (2026), M5 Max 18-core CPU / 40-core GPU, 48GBsealed appliance · 1 part · break down its parts5 BLOCKERS
$3,304 (derived)

DERIVEDApple Mac Studio (2026), M5 Max 18-core CPU / 40-core GPU, 48GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$3,304 = $3,304

Apple Mac Studio (2026), M5 Max 18-core CPU / 40-core GPU, 48GB
330431 ¢ — Apple · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
$4,769 (estimate)

ESTIMATEApple Mac Studio (2026), M5 Max 18-core CPU / 40-core GPU, 48GB: purchase plus five years of powerArithmetic that also depends on a tunable assumption, listed below.

purchase_plus_5yr_power = purchase + five_year_power

purchase_plus_5yr_power = $3,304.31 + $1,465.14 = $4,769.45

purchase_cents
330431 ¢ — rig_price_floor = Σ part price_floor × quantity
five_year_power_cents
146514 ¢ — five_year_power = annual_kwh × ¢_per_kwh × 5
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Watts are at the wall (PSU conversion losses included by the caller).
  • Load and idle draws are averages over their hours, not peaks.
  • Duty cycle (Settings → Power): 6 h/day at load, 18 h/day idle.
  • 365 days per year (leap days ignored, −0.07%).
  • Flat rate for the whole period: no escalation, no time-of-use or tiered pricing, no demand charges.
  • Rounded to the nearest cent, half-up, once from the unrounded annual cost.
$91.79 (derived)

DERIVEDApple Mac Studio (2026), M5 Max 18-core CPU / 40-core GPU, 48GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $3,304 / 36 GB = $91.79 /GB

price_floor
3304.31 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
36 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$5.38 (derived)

DERIVEDApple Mac Studio (2026), M5 Max 18-core CPU / 40-core GPU, 48GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $3,304 / 614 GB/s = $5.38 per GB/s

price_floor
3304.31 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
614 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but Apple Mac Studio (2026), M5 Max 18-core CPU / 40-core GPU, 48GB has 36 GB usable — it will not load, so there is no decode speed.
Priced fromApplecondition newaccess tier 3, route apple_store
72Apple Mac mini (2026), M5 Pro 15-core CPU / 16-core GPU, 64GBsealed appliance · 1 part · break down its parts5 BLOCKERS
$2,878 (derived)

DERIVEDApple Mac mini (2026), M5 Pro 15-core CPU / 16-core GPU, 64GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$2,878 = $2,878

Apple Mac mini (2026), M5 Pro 15-core CPU / 16-core GPU, 64GB
287781 ¢ — Apple · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Idle draw is not published for 1 of 1 part in this rig, and the duty cycle in Settings spends most of the day idle — counting an unknown idle draw as zero would understate the bill. (missing: Apple Mac mini (2026), M5 Pro 15-core CPU / 16-core GPU, 64GB: idle_watts)
$59.95 (derived)

DERIVEDApple Mac mini (2026), M5 Pro 15-core CPU / 16-core GPU, 64GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $2,878 / 48 GB = $59.95 /GB

price_floor
2877.81 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
48 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$9.37 (derived)

DERIVEDApple Mac mini (2026), M5 Pro 15-core CPU / 16-core GPU, 64GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $2,878 / 307 GB/s = $9.37 per GB/s

price_floor
2877.81 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
307 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but Apple Mac mini (2026), M5 Pro 15-core CPU / 16-core GPU, 64GB has 48 GB usable — it will not load, so there is no decode speed.
Priced fromApplecondition newaccess tier 3, route apple_store
73Minisforum MS-S1 MAX (Ryzen AI Max+ 395), 64GBsealed appliance · 1 part · break down its parts6 BLOCKERS
$2,771 (derived)

DERIVEDMinisforum MS-S1 MAX (Ryzen AI Max+ 395), 64GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$2,771 = $2,771

Minisforum MS-S1 MAX (Ryzen AI Max+ 395), 64GB
277118 ¢ — Minisforum · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Load draw at the wall is unavailable: Electrical feasibility cannot be computed exactly: no load or TDP watts for Minisforum MS-S1 MAX (Ryzen AI Max+ 395), 64GB. (missing: Minisforum MS-S1 MAX (Ryzen AI Max+ 395), 64GB: load_watts or tdp_watts, Minisforum MS-S1 MAX (Ryzen AI Max+ 395), 64GB: idle_watts)
$48.11 (derived)

DERIVEDMinisforum MS-S1 MAX (Ryzen AI Max+ 395), 64GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $2,771 / 57.6 GB = $48.11 /GB

price_floor
2771.18 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
57.6 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$10.82 (derived)

DERIVEDMinisforum MS-S1 MAX (Ryzen AI Max+ 395), 64GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $2,771 / 256 GB/s = $10.82 per GB/s

price_floor
2771.18 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
256 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but Minisforum MS-S1 MAX (Ryzen AI Max+ 395), 64GB has 57.6 GB usable — it will not load, so there is no decode speed.
Priced fromMinisforumcondition newaccess tier 3, route shopify_store
74Apple Mac Studio (2026), M5 Max 18-core CPU / 32-core GPU, 36GBsealed appliance · 1 part · break down its parts6 BLOCKERS
$2,665 (derived)

DERIVEDApple Mac Studio (2026), M5 Max 18-core CPU / 32-core GPU, 36GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$2,665 = $2,665

Apple Mac Studio (2026), M5 Max 18-core CPU / 32-core GPU, 36GB
266456 ¢ — Apple · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
$3,306 (estimate)

ESTIMATEApple Mac Studio (2026), M5 Max 18-core CPU / 32-core GPU, 36GB: purchase plus five years of powerArithmetic that also depends on a tunable assumption, listed below.

purchase_plus_5yr_power = purchase + five_year_power

purchase_plus_5yr_power = $2,664.56 + $641.37 = $3,305.93

purchase_cents
266456 ¢ — rig_price_floor = Σ part price_floor × quantity
five_year_power_cents
64137 ¢ — five_year_power = annual_kwh × ¢_per_kwh × 5
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Watts are at the wall (PSU conversion losses included by the caller).
  • Load and idle draws are averages over their hours, not peaks.
  • Duty cycle (Settings → Power): 6 h/day at load, 18 h/day idle.
  • 365 days per year (leap days ignored, −0.07%).
  • Flat rate for the whole period: no escalation, no time-of-use or tiered pricing, no demand charges.
  • Rounded to the nearest cent, half-up, once from the unrounded annual cost.
$98.69 (derived)

DERIVEDApple Mac Studio (2026), M5 Max 18-core CPU / 32-core GPU, 36GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $2,665 / 27 GB = $98.69 /GB

price_floor
2664.56 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
27 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$5.79 (derived)

DERIVEDApple Mac Studio (2026), M5 Max 18-core CPU / 32-core GPU, 36GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $2,665 / 460 GB/s = $5.79 per GB/s

price_floor
2664.56 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
460 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but Apple Mac Studio (2026), M5 Max 18-core CPU / 32-core GPU, 36GB has 27 GB usable — it will not load, so there is no decode speed.
Priced fromApplecondition newaccess tier 3, route apple_store
75Minisforum N5 MAX AI NAS (Ryzen AI Max+ 395), 64GBsealed appliance · 1 part · break down its parts6 BLOCKERS
$2,558 (derived)

DERIVEDMinisforum N5 MAX AI NAS (Ryzen AI Max+ 395), 64GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$2,558 = $2,558

Minisforum N5 MAX AI NAS (Ryzen AI Max+ 395), 64GB
255793 ¢ — Minisforum · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
$3,202 (estimate)

ESTIMATEMinisforum N5 MAX AI NAS (Ryzen AI Max+ 395), 64GB: purchase plus five years of powerArithmetic that also depends on a tunable assumption, listed below.

purchase_plus_5yr_power = purchase + five_year_power

purchase_plus_5yr_power = $2,557.93 + $644.31 = $3,202.24

purchase_cents
255793 ¢ — rig_price_floor = Σ part price_floor × quantity
five_year_power_cents
64431 ¢ — five_year_power = annual_kwh × ¢_per_kwh × 5
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Watts are at the wall (PSU conversion losses included by the caller).
  • Load and idle draws are averages over their hours, not peaks.
  • Duty cycle (Settings → Power): 6 h/day at load, 18 h/day idle.
  • 365 days per year (leap days ignored, −0.07%).
  • Flat rate for the whole period: no escalation, no time-of-use or tiered pricing, no demand charges.
  • Rounded to the nearest cent, half-up, once from the unrounded annual cost.
$44.41 (derived)

DERIVEDMinisforum N5 MAX AI NAS (Ryzen AI Max+ 395), 64GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $2,558 / 57.6 GB = $44.41 /GB

price_floor
2557.93 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
57.6 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$9.99 (derived)

DERIVEDMinisforum N5 MAX AI NAS (Ryzen AI Max+ 395), 64GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $2,558 / 256 GB/s = $9.99 per GB/s

price_floor
2557.93 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
256 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but Minisforum N5 MAX AI NAS (Ryzen AI Max+ 395), 64GB has 57.6 GB usable — it will not load, so there is no decode speed.
Priced fromMinisforumcondition newaccess tier 3, route shopify_store
76GMKtec EVO-X2 (Ryzen AI Max+ 395), 64GBsealed appliance · 1 part · break down its parts6 BLOCKERS
$2,346 (derived)

DERIVEDGMKtec EVO-X2 (Ryzen AI Max+ 395), 64GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$2,346 = $2,346

GMKtec EVO-X2 (Ryzen AI Max+ 395), 64GB
234574 ¢ — GMKtec · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Idle draw is not published for 1 of 1 part in this rig, and the duty cycle in Settings spends most of the day idle — counting an unknown idle draw as zero would understate the bill. (missing: GMKtec EVO-X2 (Ryzen AI Max+ 395), 64GB: idle_watts)
$40.72 (derived)

DERIVEDGMKtec EVO-X2 (Ryzen AI Max+ 395), 64GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $2,346 / 57.6 GB = $40.72 /GB

price_floor
2345.74 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
57.6 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$9.16 (derived)

DERIVEDGMKtec EVO-X2 (Ryzen AI Max+ 395), 64GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $2,346 / 256 GB/s = $9.16 per GB/s

price_floor
2345.74 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
256 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but GMKtec EVO-X2 (Ryzen AI Max+ 395), 64GB has 57.6 GB usable — it will not load, so there is no decode speed.
Priced fromGMKteccondition newaccess tier 3, route shopify_store
77Framework Framework Desktop DIY Edition (Ryzen AI Max+ 395), 64GBsealed appliance · 1 part · break down its parts6 BLOCKERS
$2,089 (derived)

DERIVEDFramework Framework Desktop DIY Edition (Ryzen AI Max+ 395), 64GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$2,089 = $2,089

Framework Framework Desktop DIY Edition (Ryzen AI Max+ 395), 64GB
208878 ¢ — Framework · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Idle draw is not published for 1 of 1 part in this rig, and the duty cycle in Settings spends most of the day idle — counting an unknown idle draw as zero would understate the bill. (missing: Framework Framework Desktop DIY Edition (Ryzen AI Max+ 395), 64GB: idle_watts)
$36.26 (derived)

DERIVEDFramework Framework Desktop DIY Edition (Ryzen AI Max+ 395), 64GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $2,089 / 57.6 GB = $36.26 /GB

price_floor
2088.78 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
57.6 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$8.16 (derived)

DERIVEDFramework Framework Desktop DIY Edition (Ryzen AI Max+ 395), 64GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $2,089 / 256 GB/s = $8.16 per GB/s

price_floor
2088.78 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
256 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but Framework Framework Desktop DIY Edition (Ryzen AI Max+ 395), 64GB has 57.6 GB usable — it will not load, so there is no decode speed.
Priced fromFrameworkcondition newaccess tier 2, route framework_html
78Apple Mac mini (2026), M6, 32GBsealed appliance · 1 part · break down its parts6 BLOCKERS
$1,918 (derived)

DERIVEDApple Mac mini (2026), M6, 32GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$1,918 = $1,918

Apple Mac mini (2026), M6, 32GB
191818 ¢ — Apple · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
$2,410 (estimate)

ESTIMATEApple Mac mini (2026), M6, 32GB: purchase plus five years of powerArithmetic that also depends on a tunable assumption, listed below.

purchase_plus_5yr_power = purchase + five_year_power

purchase_plus_5yr_power = $1,918.18 + $491.32 = $2,409.50

purchase_cents
191818 ¢ — rig_price_floor = Σ part price_floor × quantity
five_year_power_cents
49132 ¢ — five_year_power = annual_kwh × ¢_per_kwh × 5
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Watts are at the wall (PSU conversion losses included by the caller).
  • Load and idle draws are averages over their hours, not peaks.
  • Duty cycle (Settings → Power): 6 h/day at load, 18 h/day idle.
  • 365 days per year (leap days ignored, −0.07%).
  • Flat rate for the whole period: no escalation, no time-of-use or tiered pricing, no demand charges.
  • Rounded to the nearest cent, half-up, once from the unrounded annual cost.
$79.92 (derived)

DERIVEDApple Mac mini (2026), M6, 32GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $1,918 / 24 GB = $79.92 /GB

price_floor
1918.18 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
24 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$11.28 (derived)

DERIVEDApple Mac mini (2026), M6, 32GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $1,918 / 170 GB/s = $11.28 per GB/s

price_floor
1918.18 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
170 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but Apple Mac mini (2026), M6, 32GB has 24 GB usable — it will not load, so there is no decode speed.
Priced fromApplecondition newaccess tier 3, route apple_store
79Apple Mac mini (2026), M5 Pro 15-core CPU / 16-core GPU, 24GBsealed appliance · 1 part · break down its parts6 BLOCKERS
$1,812 (derived)

DERIVEDApple Mac mini (2026), M5 Pro 15-core CPU / 16-core GPU, 24GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$1,812 = $1,812

Apple Mac mini (2026), M5 Pro 15-core CPU / 16-core GPU, 24GB
181156 ¢ — Apple · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Idle draw is not published for 1 of 1 part in this rig, and the duty cycle in Settings spends most of the day idle — counting an unknown idle draw as zero would understate the bill. (missing: Apple Mac mini (2026), M5 Pro 15-core CPU / 16-core GPU, 24GB: idle_watts)
$100.64 (derived)

DERIVEDApple Mac mini (2026), M5 Pro 15-core CPU / 16-core GPU, 24GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $1,812 / 18 GB = $100.64 /GB

price_floor
1811.56 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
18 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$5.90 (derived)

DERIVEDApple Mac mini (2026), M5 Pro 15-core CPU / 16-core GPU, 24GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $1,812 / 307 GB/s = $5.9 per GB/s

price_floor
1811.56 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
307 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but Apple Mac mini (2026), M5 Pro 15-core CPU / 16-core GPU, 24GB has 18 GB usable — it will not load, so there is no decode speed.
Priced fromApplecondition newaccess tier 3, route apple_store
80Framework Framework Desktop DIY Edition (Ryzen AI Max 385), 32GBsealed appliance · 1 part · break down its parts6 BLOCKERS
$1,353 (derived)

DERIVEDFramework Framework Desktop DIY Edition (Ryzen AI Max 385), 32GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$1,353 = $1,353

Framework Framework Desktop DIY Edition (Ryzen AI Max 385), 32GB
135307 ¢ — Framework · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
Unavailable — Purchase + 5-year power needs the power term. Idle draw is not published for 1 of 1 part in this rig, and the duty cycle in Settings spends most of the day idle — counting an unknown idle draw as zero would understate the bill. (missing: Framework Framework Desktop DIY Edition (Ryzen AI Max 385), 32GB: idle_watts)
$46.98 (derived)

DERIVEDFramework Framework Desktop DIY Edition (Ryzen AI Max 385), 32GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $1,353 / 28.8 GB = $46.98 /GB

price_floor
1353.07 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
28.8 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$5.29 (derived)

DERIVEDFramework Framework Desktop DIY Edition (Ryzen AI Max 385), 32GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $1,353 / 256 GB/s = $5.29 per GB/s

price_floor
1353.07 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
256 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but Framework Framework Desktop DIY Edition (Ryzen AI Max 385), 32GB has 28.8 GB usable — it will not load, so there is no decode speed.
Priced fromFrameworkcondition newaccess tier 2, route framework_html
81Apple Mac mini (2026), M6, 16GBsealed appliance · 1 part · break down its parts6 BLOCKERS
$959 (derived)

DERIVEDApple Mac mini (2026), M6, 16GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$959 = $959

Apple Mac mini (2026), M6, 16GB
95856 ¢ — Apple · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
at least
$1,191 (estimate)

ESTIMATEApple Mac mini (2026), M6, 16GB: purchase plus five years of powerArithmetic that also depends on a tunable assumption, listed below.

purchase_plus_5yr_power = purchase + five_year_power

purchase_plus_5yr_power = $958.56 + $232.42 = $1,190.98

purchase_cents
95856 ¢ — rig_price_floor = Σ part price_floor × quantity
five_year_power_cents
23242 ¢ — five_year_power = annual_kwh × ¢_per_kwh × 5
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Watts are at the wall (PSU conversion losses included by the caller).
  • Load and idle draws are averages over their hours, not peaks.
  • Duty cycle (Settings → Power): 6 h/day at load, 18 h/day idle.
  • 365 days per year (leap days ignored, −0.07%).
  • Flat rate for the whole period: no escalation, no time-of-use or tiered pricing, no demand charges.
  • Rounded to the nearest cent, half-up, once from the unrounded annual cost.
$79.88 (derived)

DERIVEDApple Mac mini (2026), M6, 16GB: cost per GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $959 / 12 GB = $79.88 /GB

price_floor
958.56 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
12 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.
$6.27 (derived)

DERIVEDApple Mac mini (2026), M6, 16GB: cost per GB/sArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $959 / 153 GB/s = $6.27 per GB/s

price_floor
958.56 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
153 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but Apple Mac mini (2026), M6, 16GB has 12 GB usable — it will not load, so there is no decode speed.
Priced fromApplecondition newaccess tier 3, route apple_store

Total cost, most expensive first — SPEC §15.11 target 1 ranks complete rigs by total landed cost descending. Cost basis: Price + NJ sales tax, with shipping EXCLUDED because no listing states one. Shipping is never negative, so this is a strict lower bound on landed cost, not an estimate of it. A rig’s cost is the sum of its parts’ cheapest usable offers; the row beneath each rig names the offer or offers the figure came from, with the condition, the access tier and route that produced it, and how old the observation is. Past the 36 h freshness window set in Settings the age reads “stale” in caution ink. Where the columns need more room than the panel has — below about 1,180 px — the table scrolls sideways inside its own labelled, keyboard-reachable region rather than dropping a column or truncating a rig’s name.

Listed, not ranked

8 of 89 rigs have no total cost
Rigs that could not be ranked on total cost, each with the reason
RigWhy total cost could not be computedCostUsable GB
AMD Ryzen AI Halo Developer Platform (Ryzen AI Max+ 395), 128GBsealed appliance · 1 part · break down its partsUnavailable — A rig total needs a price floor for every part, and 1 of 1 has no usable offer. (missing: AMD Ryzen AI Halo Developer Platform (Ryzen AI Max+ 395), 128GB)Unavailable — A rig total needs a price floor for every part, and 1 of 1 has no usable offer. (missing: AMD Ryzen AI Halo Developer Platform (Ryzen AI Max+ 395), 128GB)
115 (derived)

DERIVEDAMD Ryzen AI Halo Developer Platform (Ryzen AI Max+ 395), 128GB: usable memoryArithmetic on observed inputs only. No tunable assumption enters it.

available_gb = Σ usable_gb over memory pools

unified LPDDR5x 115.2 = 115.2 GB

unified LPDDR5x.usable_gb
115.2 GB — 128 GB physical
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
Priced fromUnavailable — No monitored, non-quarantined offer on record prices any part of this rig, so there is nothing to attribute a cost to. 1 part unpriced. (missing: AMD Ryzen AI Halo Developer Platform (Ryzen AI Max+ 395), 128GB)
Apple Mac mini (2026), M5 Pro 15-core CPU / 16-core GPU, 48GBsealed appliance · 1 part · break down its partsUnavailable — A rig total needs a price floor for every part, and 1 of 1 has no usable offer. (missing: Apple Mac mini (2026), M5 Pro 15-core CPU / 16-core GPU, 48GB)Unavailable — A rig total needs a price floor for every part, and 1 of 1 has no usable offer. (missing: Apple Mac mini (2026), M5 Pro 15-core CPU / 16-core GPU, 48GB)
36.0 (derived)

DERIVEDApple Mac mini (2026), M5 Pro 15-core CPU / 16-core GPU, 48GB: usable memoryArithmetic on observed inputs only. No tunable assumption enters it.

available_gb = Σ usable_gb over memory pools

unified memory 36 = 36 GB

unified memory.usable_gb
36 GB — 48 GB physical
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
Priced fromUnavailable — No monitored, non-quarantined offer on record prices any part of this rig, so there is nothing to attribute a cost to. 1 part unpriced. (missing: Apple Mac mini (2026), M5 Pro 15-core CPU / 16-core GPU, 48GB)
Apple Mac mini (2026), M6, 24GBsealed appliance · 1 part · break down its partsUnavailable — A rig total needs a price floor for every part, and 1 of 1 has no usable offer. (missing: Apple Mac mini (2026), M6, 24GB)Unavailable — A rig total needs a price floor for every part, and 1 of 1 has no usable offer. (missing: Apple Mac mini (2026), M6, 24GB)
18.0 (derived)

DERIVEDApple Mac mini (2026), M6, 24GB: usable memoryArithmetic on observed inputs only. No tunable assumption enters it.

available_gb = Σ usable_gb over memory pools

unified memory 18 = 18 GB

unified memory.usable_gb
18 GB — 24 GB physical
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
Priced fromUnavailable — No monitored, non-quarantined offer on record prices any part of this rig, so there is nothing to attribute a cost to. 1 part unpriced. (missing: Apple Mac mini (2026), M6, 24GB)
Apple Mac Studio (2025), M3 Ultra 28-core CPU / 60-core GPU, 96GBsealed appliance · 1 part · break down its partsUnavailable — A rig total needs a price floor for every part, and 1 of 1 has no usable offer. (missing: Apple Mac Studio (2025), M3 Ultra 28-core CPU / 60-core GPU, 96GB)Unavailable — A rig total needs a price floor for every part, and 1 of 1 has no usable offer. (missing: Apple Mac Studio (2025), M3 Ultra 28-core CPU / 60-core GPU, 96GB)
72.0 (derived)

DERIVEDApple Mac Studio (2025), M3 Ultra 28-core CPU / 60-core GPU, 96GB: usable memoryArithmetic on observed inputs only. No tunable assumption enters it.

available_gb = Σ usable_gb over memory pools

unified memory 72 = 72 GB

unified memory.usable_gb
72 GB — 96 GB physical
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
Priced fromUnavailable — No monitored, non-quarantined offer on record prices any part of this rig, so there is nothing to attribute a cost to. 1 part unpriced. (missing: Apple Mac Studio (2025), M3 Ultra 28-core CPU / 60-core GPU, 96GB)
Apple Mac Studio (2025), M3 Ultra 32-core CPU / 80-core GPU, 96GBsealed appliance · 1 part · break down its partsUnavailable — A rig total needs a price floor for every part, and 1 of 1 has no usable offer. (missing: Apple Mac Studio (2025), M3 Ultra 32-core CPU / 80-core GPU, 96GB)Unavailable — A rig total needs a price floor for every part, and 1 of 1 has no usable offer. (missing: Apple Mac Studio (2025), M3 Ultra 32-core CPU / 80-core GPU, 96GB)
72.0 (derived)

DERIVEDApple Mac Studio (2025), M3 Ultra 32-core CPU / 80-core GPU, 96GB: usable memoryArithmetic on observed inputs only. No tunable assumption enters it.

available_gb = Σ usable_gb over memory pools

unified memory 72 = 72 GB

unified memory.usable_gb
72 GB — 96 GB physical
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
Priced fromUnavailable — No monitored, non-quarantined offer on record prices any part of this rig, so there is nothing to attribute a cost to. 1 part unpriced. (missing: Apple Mac Studio (2025), M3 Ultra 32-core CPU / 80-core GPU, 96GB)
Apple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 512GBsealed appliance · 1 part · break down its partsUnavailable — A rig total needs a price floor for every part, and 1 of 1 has no usable offer. (missing: Apple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 512GB)Unavailable — A rig total needs a price floor for every part, and 1 of 1 has no usable offer. (missing: Apple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 512GB)
384 (derived)

DERIVEDApple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 512GB: usable memoryArithmetic on observed inputs only. No tunable assumption enters it.

available_gb = Σ usable_gb over memory pools

unified memory 384 = 384 GB

unified memory.usable_gb
384 GB — 512 GB physical
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
Priced fromUnavailable — No monitored, non-quarantined offer on record prices any part of this rig, so there is nothing to attribute a cost to. 1 part unpriced. (missing: Apple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 512GB)
Beelink GTR9 Pro (Ryzen AI Max+ 395), 128GBsealed appliance · 1 part · break down its partsUnavailable — A rig total needs a price floor for every part, and 1 of 1 has no usable offer. (missing: Beelink GTR9 Pro (Ryzen AI Max+ 395), 128GB)Unavailable — A rig total needs a price floor for every part, and 1 of 1 has no usable offer. (missing: Beelink GTR9 Pro (Ryzen AI Max+ 395), 128GB)
115 (derived)

DERIVEDBeelink GTR9 Pro (Ryzen AI Max+ 395), 128GB: usable memoryArithmetic on observed inputs only. No tunable assumption enters it.

available_gb = Σ usable_gb over memory pools

unified LPDDR5x 115.2 = 115.2 GB

unified LPDDR5x.usable_gb
115.2 GB — 128 GB physical
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
Priced fromUnavailable — No monitored, non-quarantined offer on record prices any part of this rig, so there is nothing to attribute a cost to. 1 part unpriced. (missing: Beelink GTR9 Pro (Ryzen AI Max+ 395), 128GB)
Minisforum MS-S1 MAX-P495 (Ryzen AI Max+ PRO 495), 192GBsealed appliance · 1 part · break down its partsUnavailable — A rig total needs a price floor for every part, and 1 of 1 has no usable offer. (missing: Minisforum MS-S1 MAX-P495 (Ryzen AI Max+ PRO 495), 192GB)Unavailable — A rig total needs a price floor for every part, and 1 of 1 has no usable offer. (missing: Minisforum MS-S1 MAX-P495 (Ryzen AI Max+ PRO 495), 192GB)
173 (derived)

DERIVEDMinisforum MS-S1 MAX-P495 (Ryzen AI Max+ PRO 495), 192GB: usable memoryArithmetic on observed inputs only. No tunable assumption enters it.

available_gb = Σ usable_gb over memory pools

unified LPDDR5x 172.8 = 172.8 GB

unified LPDDR5x.usable_gb
172.8 GB — 192 GB physical
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
Priced fromUnavailable — No monitored, non-quarantined offer on record prices any part of this rig, so there is nothing to attribute a cost to. 1 part unpriced. (missing: Minisforum MS-S1 MAX-P495 (Ryzen AI Max+ PRO 495), 192GB)

These rigs are listed, not ranked. Sorting a rig whose total cost could not be computed to the end of the table would read as “worst value for money”, and the truth is “not measured”. Each row states which input is missing.

Inside Apple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB

1 part · the rig you opened · back to rank 1

Price floor

$10,128 (derived)

DERIVEDApple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB: price floorArithmetic on observed inputs only. No tunable assumption enters it.

rig_price_floor = Σ part price_floor × quantity

$10,128 = $10,128

Apple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB
1012831 ¢ — Apple · new
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.

+ 5 yr power

Unavailable — Purchase + 5-year power needs the power term. Idle draw is not published for 1 of 1 part in this rig, and the duty cycle in Settings spends most of the day idle — counting an unknown idle draw as zero would understate the bill. (missing: Apple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB: idle_watts)

Usable memory

192 GB (derived)

DERIVEDApple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB: usable memoryArithmetic on observed inputs only. No tunable assumption enters it.

available_gb = Σ usable_gb over memory pools

unified memory 192 = 192 GB

unified memory.usable_gb
192 GB — 256 GB physical
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).

Cost per GB

$52.75 (derived)

DERIVEDApple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB: cost per usable GBArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gb = price_floor / usable_memory_gb

= $10,128 / 192 GB = $52.75 /GB

price_floor
10128.31 USD — rig_price_floor = Σ part price_floor × quantity
usable_memory_gb
192 GB — usable memory after the OS and framework reservation (Settings → Estimator)
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB is at least this much.

Cost per tok/s · GLM-5.3

Unavailable — Cost per tok/s needs decode_tok_s: GLM-5.3 has no decode speed on this rig: Model needs 507.62 GB but Apple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB has 192 GB usable — it will not load, so there is no decode speed.
Plan fit
DOES NOT MEET THE PLAN
  • GLM-5.3: won't load (507.62 GB needed, 192 GB usable).
  • Largest model needs 507.62 GB; 192 GB usable.
  • DeepSeek-V4-Pro-0813: No quantization chosen or sized.
  • Qwen3-ASR 1.7B: No quantization chosen or sized.
  • and 1 more, listed in full on the Builds screen’s report for this rig.
Memory used
537 GB (estimate)

ESTIMATEApple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB: memory with every Plan model residentArithmetic that also depends on a tunable assumption, listed below.

Σ required_gb over every model in the stack

GLM-5.3 507.62 + Hunyuan3D 2.1 (shape + PBR paint) 29 = 536.62 GB

GLM-5.3
507.61664759196 GB
Hunyuan3D 2.1 (shape + PBR paint)
29 GB
  • All models loaded at once (Keep all resident).
of
192 GB (derived)

DERIVEDApple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB: usable memoryArithmetic on observed inputs only. No tunable assumption enters it.

available_gb = Σ usable_gb over memory pools

unified memory 192 = 192 GB

unified memory.usable_gb
192 GB — 256 GB physical
  • Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
(279%)
Cost per GB/s
$8.44 (derived)

DERIVEDApple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB: cost per GB/s of memory bandwidthArithmetic on observed inputs only. No tunable assumption enters it.

cost_per_gbs = price_floor / memory_bandwidth_gbs

= $10,128 / 1,200 GB/s = $8.44 per GB/s

price_floor
10128.31 USD — rig_price_floor = Σ part price_floor × quantity
memory_bandwidth_gbs
1200 GB/s — the published bandwidth of the rig's fastest memory pool
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Shipping is unstated on 1 of 1 priced offer in this rig, so no part of the total carries a delivery charge.
  • Weights are read from the fastest pool first; a rig whose weights spill into a slower pool decodes below this figure.
  • Because the numerator is a lower bound, this ratio is a lower bound too: the true cost per GB/s is at least this much.
Power / yr
Unavailable — Idle draw is not published for 1 of 1 part in this rig, and the duty cycle in Settings spends most of the day idle — counting an unknown idle draw as zero would understate the bill. (missing: Apple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB: idle_watts)

Unavailable — 3 models in this Plan contribute no memory to either residency total on this rig (missing: DeepSeek-V4-Pro-0813, Qwen3-ASR 1.7B, Qwen3-VL 32B Instruct)Both totals are therefore lower bounds, not estimates of the whole Plan — and so is every figure below that divides a cost by them.

Every part of Apple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB, ranked by its own cost with its share of the rig total, the offer each price came from, and what the step up to it buys on this Plan
# rankPartPrice floorShare of rigWhat the step up buys
1Apple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GBunified box · ×1 · 1 offer on record
$10,128 (derived)

DERIVEDApple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB: price floor for 1Arithmetic on observed inputs only. No tunable assumption enters it.

part_total = unit_cost × quantity

= $10,128.31 × 1 = $10,128.31

unit_cost_cents
1012831 ¢ — price_floor = price + round_half_up(price × tax_rate) (shipping excluded: the listing does not state one)
quantity
1 — Apple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB in this rig
  • LOWER BOUND: shipping is not stated on this offer and is therefore excluded entirely. The landed cost is this figure plus shipping, never less.
  • Sales tax rounded to the nearest cent, half-up (x.5¢ → x+1¢), in exact decimal arithmetic on the rate.
  • Listed price is taken as pre-tax.
100.0% (derived)

DERIVEDApple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB: share of the rig's costArithmetic on observed inputs only. No tunable assumption enters it.

share_of_total = part price_floor / rig price_floor

= $10,128 / $10,128 = 100%

part_cents
1012831 ¢
rig_cents
1012831 ¢
  • Both terms are lower bounds (price floor), so the share is exact only if every part's unstated shipping is proportional to its price.
OVERSPEND Stepping down to GIGABYTE AI TOP ATOM (NVIDIA GB10), 128GB saves $1,598 on the price floor and costs nothing the Plan uses: the same models load, at the same or higher total speed. This step up is overspend.measured against GIGABYTE AI TOP ATOM (NVIDIA GB10), 128GB at $8,530 for the whole rig
0.00 tok/s per $1k (derived)

DERIVEDApple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB: tok/s bought per extra $1,000Arithmetic on observed inputs only. No tunable assumption enters it.

tok_s_per_$1k = Δ total decode tok/s across the Plan / (Δ cost / $1,000)

= 0 tok/s / ($1,598 / $1,000) = 0 tok/s per $1k

delta_tok_s
0 tok/s — this rig minus GIGABYTE AI TOP ATOM (NVIDIA GB10), 128GB, summed over every Plan model that decodes
delta_cost_cents
159832 ¢ — Price floor of this rig minus price floor of the same rig with GIGABYTE AI TOP ATOM (NVIDIA GB10), 128GB in the slot
  • Every other part is held fixed; only the one slot is swapped.
  • A model that does not load contributes 0 tok/s to the total, so a swap that drops a model shows up as a large negative Δ.
  • Both costs are lower bounds (price floor), so the rate is exact only if the two options' unstated shipping charges are equal.
Priced fromApplecondition newaccess tier 3, route apple_store

Each part is compared with the most expensive cheaper part in the same category that leaves the rig free of compatibility errors and fully priced, with everything else held fixed. Load-bearing: the step up makes a Plan model load that otherwise would not. Overspend: it costs more and buys nothing this Plan uses. Proportional: it buys decode speed, priced in tok/s per $1,000. Baseline: nothing cheaper exists for the slot.

What this screen could not measure

0 gaps

Every input this screen needs is present: 171 complete parts, 137 of them priced, 8739 price observations, and an active Plan to measure cost per tok/s against.