Compare · 3 of 89 machines · on the Plan "My stack (Plan)" · target Interactive · recomputed in 3887 ms

None of these 3 machines is the one to buy — None of these 3 machines runs every model in "My stack (Plan)", so recommending one of them would mean recommending something that does not do the job. The closest is Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, held back by 3 things.

NONE MEETS THE PLANNO CONFIDENCE LEVEL

89 complete machines can be built from 171 evaluable catalog parts, and 0 of them meet the Plan at ≥ 15 tok/s. The machines come from the same search the Plan and Models use.

Since yesterdayNo observed price moved on any of these 3 machines.compared with the observations on record at Oct 18, 7:59 PM ET (24 h ago)

What moved

Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: $38,872, the same as 24 h ago.
Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: $44,568, the same as 24 h ago.
Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM: $45,703, the same as 24 h ago.

Capacity, bandwidth, compute

the three axes a machine is bought on (SPEC §3)

Capacity, bandwidth and compute for 3 machines side by side
Compared on	Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM7-part build · 3 blockers	Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM7-part build · 3 blockers	Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM7-part build · 3 blockers
Usable memorywhat a model can actually occupy	1,004 GB (derived) DERIVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 31.04 + System RAM 972.8 = 1,003.84 GB GPU VRAM.usable_gb 31.04 GB — 32 GB physical System RAM.usable_gb 972.8 GB — 1,024 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).	1,066 GB (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 93.12 + System RAM 972.8 = 1,065.92 GB GPU VRAM.usable_gb 93.12 GB — 96 GB physical System RAM.usable_gb 972.8 GB — 1,024 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).	1,066 GB (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 93.12 + System RAM 972.8 = 1,065.92 GB GPU VRAM.usable_gb 93.12 GB — 96 GB physical System RAM.usable_gb 972.8 GB — 1,024 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings).
Physical memorywhat is fitted, before the OS reserve	1,056 GB (observed) OBSERVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Physical memoryMeasured or published, passed through unchanged — not computed here. physical memory — published at the source, passed through unchanged; nothing is computed here 1,056 GB, as recorded from the GPU VRAM + System RAM spec of Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM physical_memory 1,056 GB — the GPU VRAM + System RAM spec of Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM What is fitted. Usable memory is lower — the OS and the driver reserve some, and the reservation is a Settings input. Observed, not derived: this app records the figure and its source, and does not re-measure it.	1,120 GB (observed) OBSERVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Physical memoryMeasured or published, passed through unchanged — not computed here. physical memory — published at the source, passed through unchanged; nothing is computed here 1,120 GB, as recorded from the GPU VRAM + System RAM spec of Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM physical_memory 1,120 GB — the GPU VRAM + System RAM spec of Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM What is fitted. Usable memory is lower — the OS and the driver reserve some, and the reservation is a Settings input. Observed, not derived: this app records the figure and its source, and does not re-measure it.	1,120 GB (observed) OBSERVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Physical memoryMeasured or published, passed through unchanged — not computed here. physical memory — published at the source, passed through unchanged; nothing is computed here 1,120 GB, as recorded from the GPU VRAM + System RAM spec of Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM physical_memory 1,120 GB — the GPU VRAM + System RAM spec of Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM What is fitted. Usable memory is lower — the OS and the driver reserve some, and the reservation is a Settings input. Observed, not derived: this app records the figure and its source, and does not re-measure it.
The stack needsat swap on demand — the same on every machine	966 GB (estimate) ESTIMATEHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — The stack needsArithmetic that also depends on a tunable assumption, listed below. max(required_gb) over the stack largest = DeepSeek-V4-Pro-0813 at 966.21 GB GLM-5.3 507.61664759196 GB DeepSeek-V4-Pro-0813 966.2106890624001 GB Hunyuan3D 2.1 (shape + PBR paint) 29 GB Qwen3-ASR 1.7B 5.953946505559999 GB Qwen3-VL 32B Instruct 28.763790271999998 GB Models are loaded one at a time and swapped on demand; load time not included.	966 GB (estimate) ESTIMATEHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — The stack needsArithmetic that also depends on a tunable assumption, listed below. max(required_gb) over the stack largest = DeepSeek-V4-Pro-0813 at 966.21 GB GLM-5.3 507.61664759196 GB DeepSeek-V4-Pro-0813 966.2106890624001 GB Hunyuan3D 2.1 (shape + PBR paint) 29 GB Qwen3-ASR 1.7B 5.953946505559999 GB Qwen3-VL 32B Instruct 28.763790271999998 GB Models are loaded one at a time and swapped on demand; load time not included.	966 GB (estimate) ESTIMATEHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — The stack needsArithmetic that also depends on a tunable assumption, listed below. max(required_gb) over the stack largest = DeepSeek-V4-Pro-0813 at 966.21 GB GLM-5.3 507.61664759196 GB DeepSeek-V4-Pro-0813 966.2106890624001 GB Hunyuan3D 2.1 (shape + PBR paint) 29 GB Qwen3-ASR 1.7B 5.953946505559999 GB Qwen3-VL 32B Instruct 28.763790271999998 GB Models are loaded one at a time and swapped on demand; load time not included.
Memory used by the stackover 92% is Tight; over 100% it will not load	96.3% (estimate) ESTIMATEHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Memory used by the stackArithmetic that also depends on a tunable assumption, listed below. utilisation = memory the stack needs / usable memory = 966.21 GB / 1,003.84 GB = 96.3% needed_gb 966.2106890624001 GB — the largest single model usable_gb 1003.8399999999999 GB — usable memory of this machine Measured at the saved Plan's residency: swap on demand. Over 100% the stack does not fit. Over 92% it is Tight (SPEC §8.6).	90.6% (estimate) ESTIMATEHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Memory used by the stackArithmetic that also depends on a tunable assumption, listed below. utilisation = memory the stack needs / usable memory = 966.21 GB / 1,065.92 GB = 90.6% needed_gb 966.2106890624001 GB — the largest single model usable_gb 1065.92 GB — usable memory of this machine Measured at the saved Plan's residency: swap on demand. Over 100% the stack does not fit. Over 92% it is Tight (SPEC §8.6).	90.6% (estimate) ESTIMATEHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Memory used by the stackArithmetic that also depends on a tunable assumption, listed below. utilisation = memory the stack needs / usable memory = 966.21 GB / 1,065.92 GB = 90.6% needed_gb 966.2106890624001 GB — the largest single model usable_gb 1065.92 GB — usable memory of this machine Measured at the saved Plan's residency: swap on demand. Over 100% the stack does not fit. Over 92% it is Tight (SPEC §8.6).
Peak bandwidthdecides how fast tokens come out	1,792 GB/s (observed) OBSERVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Peak bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 1,792 GB/s, as recorded from the GPU VRAM spec of Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM peak_memory_bandwidth 1,792 GB/s — the GPU VRAM spec of Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM decode_efficiency 0.86 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md memory_architecture gpu_vram — component spec Theoretical peak. Decode reaches a fraction of it; the efficiency factor above is applied inside each decode estimate. The fastest pool. Weights are placed fastest-pool-first, so a second pool changes speed only once the first is full. Observed, not derived: this app records the figure and its source, and does not re-measure it.	1,792 GB/s (observed) OBSERVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Peak bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 1,792 GB/s, as recorded from the GPU VRAM spec of Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM peak_memory_bandwidth 1,792 GB/s — the GPU VRAM spec of Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM decode_efficiency 0.86 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md memory_architecture gpu_vram — component spec Theoretical peak. Decode reaches a fraction of it; the efficiency factor above is applied inside each decode estimate. The fastest pool. Weights are placed fastest-pool-first, so a second pool changes speed only once the first is full. Observed, not derived: this app records the figure and its source, and does not re-measure it.	1,792 GB/s (observed) OBSERVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Peak bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 1,792 GB/s, as recorded from the GPU VRAM spec of Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM peak_memory_bandwidth 1,792 GB/s — the GPU VRAM spec of Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM decode_efficiency 0.86 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md memory_architecture gpu_vram — component spec Theoretical peak. Decode reaches a fraction of it; the efficiency factor above is applied inside each decode estimate. The fastest pool. Weights are placed fastest-pool-first, so a second pool changes speed only once the first is full. Observed, not derived: this app records the figure and its source, and does not re-measure it.
Slowest model, decodea rig is as interactive as the model you wait on	5.6 tok/s (estimate) ESTIMATEHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Slowest model, decodeArithmetic that also depends on a tunable assumption, listed below. slowest = min(decode tok/s) over the models compared slowest of 3 models is DeepSeek-V4-Pro-0813 at 5.6 tok/s GLM-5.3 6 tok/s — this machine's own decode estimate DeepSeek-V4-Pro-0813 5.6 tok/s — this machine's own decode estimate Qwen3-VL 32B Instruct 52.7 tok/s — this machine's own decode estimate A rig is only as interactive as its slowest model: the one you wait on decides whether the machine is usable. 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).	5.9 tok/s (estimate) ESTIMATEHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Slowest model, decodeArithmetic that also depends on a tunable assumption, listed below. slowest = min(decode tok/s) over the models compared slowest of 3 models is DeepSeek-V4-Pro-0813 at 5.9 tok/s GLM-5.3 6.7 tok/s — this machine's own decode estimate DeepSeek-V4-Pro-0813 5.9 tok/s — this machine's own decode estimate Qwen3-VL 32B Instruct 52.7 tok/s — this machine's own decode estimate A rig is only as interactive as its slowest model: the one you wait on decides whether the machine is usable. 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).	5.9 tok/s (estimate) ESTIMATEHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Slowest model, decodeArithmetic that also depends on a tunable assumption, listed below. slowest = min(decode tok/s) over the models compared slowest of 3 models is DeepSeek-V4-Pro-0813 at 5.9 tok/s GLM-5.3 6.7 tok/s — this machine's own decode estimate DeepSeek-V4-Pro-0813 5.9 tok/s — this machine's own decode estimate Qwen3-VL 32B Instruct 52.7 tok/s — this machine's own decode estimate A rig is only as interactive as its slowest model: the one you wait on decides whether the machine is usable. 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).
Dense FP16 computedecides how fast a long prompt is ingested	209.5 TFLOPS (observed) OBSERVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Dense FP16 computeMeasured or published, passed through unchanged — not computed here. dense FP16 compute — published at the source, passed through unchanged; nothing is computed here 209.5 TFLOPS, as recorded from the published specification of Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM dense_FP16_compute 209.5 TFLOPS — the published specification of Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM Peak dense FP16/BF16 throughput. Prefill reaches a fraction of it (the MFU factor in Settings). Observed, not derived: this app records the figure and its source, and does not re-measure it.	Unavailable — No dense FP16 throughput is published for Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, so prompt ingestion can only be estimated from a measured prefill rate where one exists. (missing: compute_tflops_fp16)	Unavailable — No dense FP16 throughput is published for Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, so prompt ingestion can only be estimated from a measured prefill rate where one exists. (missing: compute_tflops_fp16)
GPU APIswhat a model's runtime needs to exist on the machine	cuda	cuda	cuda
Against the Plandoes this machine run every model?	3 blockers GLM-5.3: 6 tok/s, below ≥ 15 tok/s. DeepSeek-V4-Pro-0813: 5.6 tok/s, below ≥ 15 tok/s. Qwen3-ASR 1.7B: 6bit is an MLX file, which loads on Apple silicon only; this machine has cuda.	3 blockers GLM-5.3: 6.7 tok/s, below ≥ 15 tok/s. DeepSeek-V4-Pro-0813: 5.9 tok/s, below ≥ 15 tok/s. Qwen3-ASR 1.7B: 6bit is an MLX file, which loads on Apple silicon only; this machine has cuda.	3 blockers GLM-5.3: 6.7 tok/s, below ≥ 15 tok/s. DeepSeek-V4-Pro-0813: 5.9 tok/s, below ≥ 15 tok/s. Qwen3-ASR 1.7B: 6bit is an MLX file, which loads on Apple silicon only; this machine has cuda.

Capacity decides WHETHER a model loads, bandwidth decides HOW FAST it generates, compute decides how fast a prompt is ingested (SPEC §3) — three independent axes, and a machine can win one and lose another badly. “The stack needs” is a property of the models and the context, not of the hardware, so it is the same figure in every column.

Every model, on every machine

5 models · from the saved Plan · compare on every sizeable model instead

Every model in this comparison on each of the 3 machines, with its band, decode speed, memory and time to first token
Compared on	Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM7-part build · 3 blockers	Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM7-part build · 3 blockers	Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM7-part build · 3 blockers
GLM-5.3coding · 32,768 ctx · UD-Q4_K_XL	Usable 6.0 tok/s (estimate) ESTIMATEGLM-5.3 decode speed on Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 1.729 GB ÷ (1,792 GB/s × 0.48) + System RAM: 26.547 GB ÷ (307.2 GB/s × 0.7) + 41.98 ms fixed per-token overhead (share-weighted) → 5.97 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.0611 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.9389 — placement per_token_overhead_ms 41.978779754690706 ms — derived: 0.061 × 1.58 ms (GPU VRAM) + 0.939 × 44.61 ms (System RAM) = 41.98 ms bytes_per_token_gb 28.2764 GB 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). needs 508 GB (estimate) ESTIMATEGLM-5.3 memory requirementArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 467.29 + 2.94 + 37.38 = 507.62 GB weights_gb 467.289116837 GB — real file size kv_cache_gb 2.944401408 GB — kv formula overhead_gb 37.38312934696 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 78 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights. · 51% (estimate) ESTIMATEGLM-5.3 share of Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM's memoryArithmetic that also depends on a tunable assumption, listed below. utilization = required_gb / available_gb = 507.62 / 1,003.84 = 50.6% required_gb 507.61664759196 GB available_gb 1003.8399999999999 GB Tight when ≥ 92% of usable memory (SPEC §8.6). of memory · TTFT 36.0 s (estimate) ESTIMATEGLM-5.3 time to first token on Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 909.76 = 36 s prompt_tokens 32768 tokens prefill_tok_s 909.76 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	Usable 6.7 tok/s (estimate) ESTIMATEGLM-5.3 decode speed on Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 5.187 GB ÷ (1,792 GB/s × 0.48) + System RAM: 23.089 GB ÷ (307.2 GB/s × 0.7) + 36.72 ms fixed per-token overhead (share-weighted) → 6.66 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.1834 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.8166 — placement per_token_overhead_ms 36.71633926407211 ms — derived: 0.183 × 1.58 ms (GPU VRAM) + 0.817 × 44.61 ms (System RAM) = 36.72 ms bytes_per_token_gb 28.2764 GB 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). needs 508 GB (estimate) ESTIMATEGLM-5.3 memory requirementArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 467.29 + 2.94 + 37.38 = 507.62 GB weights_gb 467.289116837 GB — real file size kv_cache_gb 2.944401408 GB — kv formula overhead_gb 37.38312934696 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 78 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights. · 48% (estimate) ESTIMATEGLM-5.3 share of Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM's memoryArithmetic that also depends on a tunable assumption, listed below. utilization = required_gb / available_gb = 507.62 / 1,065.92 = 47.6% required_gb 507.61664759196 GB available_gb 1065.92 GB Tight when ≥ 92% of usable memory (SPEC §8.6). of memory · TTFT Unavailable — No prefill measurement or compute rating for DIY: NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB + AMD EPYC 9015. (missing: measured prefill_tokens_sec, compute_tflops_fp16)	Usable 6.7 tok/s (estimate) ESTIMATEGLM-5.3 decode speed on Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 5.187 GB ÷ (1,792 GB/s × 0.48) + System RAM: 23.089 GB ÷ (307.2 GB/s × 0.7) + 36.72 ms fixed per-token overhead (share-weighted) → 6.66 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.1834 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.8166 — placement per_token_overhead_ms 36.71633926407211 ms — derived: 0.183 × 1.58 ms (GPU VRAM) + 0.817 × 44.61 ms (System RAM) = 36.72 ms bytes_per_token_gb 28.2764 GB 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). needs 508 GB (estimate) ESTIMATEGLM-5.3 memory requirementArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 467.29 + 2.94 + 37.38 = 507.62 GB weights_gb 467.289116837 GB — real file size kv_cache_gb 2.944401408 GB — kv formula overhead_gb 37.38312934696 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 78 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights. · 48% (estimate) ESTIMATEGLM-5.3 share of Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM's memoryArithmetic that also depends on a tunable assumption, listed below. utilization = required_gb / available_gb = 507.62 / 1,065.92 = 47.6% required_gb 507.61664759196 GB available_gb 1065.92 GB Tight when ≥ 92% of usable memory (SPEC §8.6). of memory · TTFT Unavailable — No prefill measurement or compute rating for DIY: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + AMD EPYC 9015. (missing: measured prefill_tokens_sec, compute_tflops_fp16)
DeepSeek-V4-Pro-0813reasoning · 32,768 ctx · native	Tight · Usable 5.6 tok/s (estimate) ESTIMATEDeepSeek-V4-Pro-0813 decode speed on Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 0.956 GB ÷ (1,792 GB/s × 0.48) + System RAM: 28.816 GB ÷ (307.2 GB/s × 0.7) + 43.23 ms fixed per-token overhead (share-weighted) → 5.61 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.0321 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.9679 — placement per_token_overhead_ms 43.2276398014225 ms — derived: 0.032 × 1.58 ms (GPU VRAM) + 0.968 × 44.61 ms (System RAM) = 43.23 ms bytes_per_token_gb 29.7727 GB 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). needs 966 GB (estimate) ESTIMATEDeepSeek-V4-Pro-0813 memory requirementArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 892.74 + 2.05 + 71.42 = 966.21 GB weights_gb 892.74432288 GB — real file size kv_cache_gb 2.046820352 GB — kv formula overhead_gb 71.4195458304 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 61 layers keep a full KV cache. KV width inferred from a single latent 'wkv' projection (unrecognized attention family); any KV compression the runtime applies is not modeled, so the cache is likely overstated. Runtime buffers, activations, and allocator slack approximated as a fraction of weights. · 96% (estimate) ESTIMATEDeepSeek-V4-Pro-0813 share of Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM's memoryArithmetic that also depends on a tunable assumption, listed below. utilization = required_gb / available_gb = 966.21 / 1,003.84 = 96.3% required_gb 966.2106890624001 GB available_gb 1003.8399999999999 GB Tight when ≥ 92% of usable memory (SPEC §8.6). of memory · TTFT 43.7 s (estimate) ESTIMATEDeepSeek-V4-Pro-0813 time to first token on Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 750.47 = 43.7 s prompt_tokens 32768 tokens prefill_tok_s 750.47 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	Usable 5.9 tok/s (estimate) ESTIMATEDeepSeek-V4-Pro-0813 decode speed on Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 2.869 GB ÷ (1,792 GB/s × 0.48) + System RAM: 26.903 GB ÷ (307.2 GB/s × 0.7) + 40.46 ms fixed per-token overhead (share-weighted) → 5.92 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.0964 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.9036 — placement per_token_overhead_ms 40.46291940426751 ms — derived: 0.096 × 1.58 ms (GPU VRAM) + 0.904 × 44.61 ms (System RAM) = 40.46 ms bytes_per_token_gb 29.7727 GB 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). needs 966 GB (estimate) ESTIMATEDeepSeek-V4-Pro-0813 memory requirementArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 892.74 + 2.05 + 71.42 = 966.21 GB weights_gb 892.74432288 GB — real file size kv_cache_gb 2.046820352 GB — kv formula overhead_gb 71.4195458304 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 61 layers keep a full KV cache. KV width inferred from a single latent 'wkv' projection (unrecognized attention family); any KV compression the runtime applies is not modeled, so the cache is likely overstated. Runtime buffers, activations, and allocator slack approximated as a fraction of weights. · 91% (estimate) ESTIMATEDeepSeek-V4-Pro-0813 share of Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM's memoryArithmetic that also depends on a tunable assumption, listed below. utilization = required_gb / available_gb = 966.21 / 1,065.92 = 90.6% required_gb 966.2106890624001 GB available_gb 1065.92 GB Tight when ≥ 92% of usable memory (SPEC §8.6). of memory · TTFT Unavailable — No prefill measurement or compute rating for DIY: NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB + AMD EPYC 9015. (missing: measured prefill_tokens_sec, compute_tflops_fp16)	Usable 5.9 tok/s (estimate) ESTIMATEDeepSeek-V4-Pro-0813 decode speed on Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 2.869 GB ÷ (1,792 GB/s × 0.48) + System RAM: 26.903 GB ÷ (307.2 GB/s × 0.7) + 40.46 ms fixed per-token overhead (share-weighted) → 5.92 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.0964 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.9036 — placement per_token_overhead_ms 40.46291940426751 ms — derived: 0.096 × 1.58 ms (GPU VRAM) + 0.904 × 44.61 ms (System RAM) = 40.46 ms bytes_per_token_gb 29.7727 GB 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). needs 966 GB (estimate) ESTIMATEDeepSeek-V4-Pro-0813 memory requirementArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 892.74 + 2.05 + 71.42 = 966.21 GB weights_gb 892.74432288 GB — real file size kv_cache_gb 2.046820352 GB — kv formula overhead_gb 71.4195458304 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 61 layers keep a full KV cache. KV width inferred from a single latent 'wkv' projection (unrecognized attention family); any KV compression the runtime applies is not modeled, so the cache is likely overstated. Runtime buffers, activations, and allocator slack approximated as a fraction of weights. · 91% (estimate) ESTIMATEDeepSeek-V4-Pro-0813 share of Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM's memoryArithmetic that also depends on a tunable assumption, listed below. utilization = required_gb / available_gb = 966.21 / 1,065.92 = 90.6% required_gb 966.2106890624001 GB available_gb 1065.92 GB Tight when ≥ 92% of usable memory (SPEC §8.6). of memory · TTFT Unavailable — No prefill measurement or compute rating for DIY: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + AMD EPYC 9015. (missing: measured prefill_tokens_sec, compute_tflops_fp16)
Hunyuan3D 2.1 (shape + PBR paint)3d · 32,768 ctx	Runsneeds 29.0 GB · Fits: 29 GB of 1,003.84 GB.	Runsneeds 29.0 GB · Fits: 29 GB of 1,065.92 GB.	Runsneeds 29.0 GB · Fits: 29 GB of 1,065.92 GB.
Qwen3-ASR 1.7Bstt · 32,768 ctx · 6bit	Unavailable — 6bit is an MLX file, which loads on Apple silicon only; this machine has cuda. (missing: a file format this machine can load)	Unavailable — 6bit is an MLX file, which loads on Apple silicon only; this machine has cuda. (missing: a file format this machine can load)	Unavailable — 6bit is an MLX file, which loads on Apple silicon only; this machine has cuda. (missing: a file format this machine can load)
Qwen3-VL 32B Instructvision · 32,768 ctx · IQ4_NL	Fast 52.7 tok/s (estimate) ESTIMATEQwen3-VL 32B Instruct decode speed on Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (bytes_per_token / (bandwidth × efficiency) + overhead) GPU VRAM: 26.826 GB ÷ (1,792 GB/s × 0.86) + 1.58 ms fixed per-token overhead → 52.67 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency 0.86 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 1 — placement per_token_overhead_ms 1.58 ms — Settings efficiency.gpu_vram.overheadMs — calibrated fit (D-013): Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md bytes_per_token_gb 26.8259 GB Dense: every parameter is read per token. 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). Per-token overhead of System RAM (44.61 ms) is not charged: this model does not read from that pool. Batch size 1 (single user, interactive). needs 28.8 GB (estimate) ESTIMATEQwen3-VL 32B Instruct memory requirementArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 18.68 + 8.59 + 1.49 = 28.76 GB weights_gb 18.679496 GB — real file size kv_cache_gb 8.589934592 GB — kv formula overhead_gb 1.49435968 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 64 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights. · 3% (estimate) ESTIMATEQwen3-VL 32B Instruct share of Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM's memoryArithmetic that also depends on a tunable assumption, listed below. utilization = required_gb / available_gb = 28.76 / 1,003.84 = 2.9% required_gb 28.763790271999998 GB available_gb 1003.8399999999999 GB Tight when ≥ 92% of usable memory (SPEC §8.6). of memory · TTFT 28.6 s (estimate) ESTIMATEQwen3-VL 32B Instruct time to first token on Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,146.27 = 28.6 s prompt_tokens 32768 tokens prefill_tok_s 1146.27 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	Fast 52.7 tok/s (estimate) ESTIMATEQwen3-VL 32B Instruct decode speed on Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (bytes_per_token / (bandwidth × efficiency) + overhead) GPU VRAM: 26.826 GB ÷ (1,792 GB/s × 0.86) + 1.58 ms fixed per-token overhead → 52.67 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency 0.86 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 1 — placement per_token_overhead_ms 1.58 ms — Settings efficiency.gpu_vram.overheadMs — calibrated fit (D-013): Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md bytes_per_token_gb 26.8259 GB Dense: every parameter is read per token. 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). Per-token overhead of System RAM (44.61 ms) is not charged: this model does not read from that pool. Batch size 1 (single user, interactive). needs 28.8 GB (estimate) ESTIMATEQwen3-VL 32B Instruct memory requirementArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 18.68 + 8.59 + 1.49 = 28.76 GB weights_gb 18.679496 GB — real file size kv_cache_gb 8.589934592 GB — kv formula overhead_gb 1.49435968 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 64 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights. · 3% (estimate) ESTIMATEQwen3-VL 32B Instruct share of Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM's memoryArithmetic that also depends on a tunable assumption, listed below. utilization = required_gb / available_gb = 28.76 / 1,065.92 = 2.7% required_gb 28.763790271999998 GB available_gb 1065.92 GB Tight when ≥ 92% of usable memory (SPEC §8.6). of memory · TTFT Unavailable — No prefill measurement or compute rating for DIY: NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB + AMD EPYC 9015. (missing: measured prefill_tokens_sec, compute_tflops_fp16)	Fast 52.7 tok/s (estimate) ESTIMATEQwen3-VL 32B Instruct decode speed on Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (bytes_per_token / (bandwidth × efficiency) + overhead) GPU VRAM: 26.826 GB ÷ (1,792 GB/s × 0.86) + 1.58 ms fixed per-token overhead → 52.67 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency 0.86 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 1 — placement per_token_overhead_ms 1.58 ms — Settings efficiency.gpu_vram.overheadMs — calibrated fit (D-013): Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md bytes_per_token_gb 26.8259 GB Dense: every parameter is read per token. 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). Per-token overhead of System RAM (44.61 ms) is not charged: this model does not read from that pool. Batch size 1 (single user, interactive). needs 28.8 GB (estimate) ESTIMATEQwen3-VL 32B Instruct memory requirementArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 18.68 + 8.59 + 1.49 = 28.76 GB weights_gb 18.679496 GB — real file size kv_cache_gb 8.589934592 GB — kv formula overhead_gb 1.49435968 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 64 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights. · 3% (estimate) ESTIMATEQwen3-VL 32B Instruct share of Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM's memoryArithmetic that also depends on a tunable assumption, listed below. utilization = required_gb / available_gb = 28.76 / 1,065.92 = 2.7% required_gb 28.763790271999998 GB available_gb 1065.92 GB Tight when ≥ 92% of usable memory (SPEC §8.6). of memory · TTFT Unavailable — No prefill measurement or compute rating for DIY: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + AMD EPYC 9015. (missing: measured prefill_tokens_sec, compute_tflops_fp16)

Bands are SPEC §8.6: won’t load → tight (≥ 92% of memory) → batch only (< 5 tok/s) → usable (5–15) → interactive (15–40) → fast (> 40). The Plan asks for ≥ 15 tok/s. Every model is sized at the context shown beside its name, and TTFT is measured on a 32,768-token prompt — the number you feel at the start of an agentic coding turn.

What each costs

observed price, landed floor and landed cost are three different figures

What each of the 3 machines costs, and what each dollar buys
Compared on	Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM7-part build · 3 blockers	Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM7-part build · 3 blockers	Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM7-part build · 3 blockers
Observed pricewhat the retailers were seen asking, before shipping and tax	$38,872 (derived) DERIVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $3,928 × 8 + $5,100 + $70 + $180 + $660 = $38,872 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM 392811 ¢ — SabrePC · new · tier 2/sabrepc_ngstate MSI GeForce RTX 5090 32G GAMING TRIO OC 509999 ¢ — Newegg · new · tier 3/newegg_realtime ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation. BEST — the lowest observed price of the machines compared	$44,568 (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $3,928 × 8 + $10,795 + $70 + $180 + $660 = $44,568 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM 392811 ¢ — SabrePC · new · tier 2/sabrepc_ngstate NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB 1079547 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.	$45,703 (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $3,928 × 8 + $11,931 + $70 + $180 + $660 = $45,703 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM 392811 ¢ — SabrePC · new · tier 2/sabrepc_ngstate NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB 1193061 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
Landed floorprice + NJ sales tax, shipping unknown — a lower bound	$41,448 (derived) DERIVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — landed floorArithmetic on observed inputs only. No tunable assumption enters it. landed_floor = observed_price + sales_tax(observed_price) [shipping ≥ 0 and unknown] $38,872.28 + $2,575.29 = $41,447.57 observed_price_cents 3887228 ¢ — Σ cheapest observed offer per part sales_tax_cents 257529 ¢ — SPEC §2: NJ sales tax 6.625% offers_without_shipping 7 offers — price_observations.shipping_cents IS NULL Sales tax rounded to the nearest cent, half-up (x.5¢ → x+1¢), in exact decimal arithmetic on the rate. Shipping is not published by every retailer here. It is omitted, not assumed to be zero — which is why this is a floor. BEST — the lowest landed floor of the machines comparedlower bound — 7 of 7 offers do not publish a delivery charge (SabrePC, Newegg, SabrePC, Newegg, ARCTIC, Newegg, Newegg), so this total is a lower bound on what the machine lands for, not an estimate of it.	$47,520 (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — landed floorArithmetic on observed inputs only. No tunable assumption enters it. landed_floor = observed_price + sales_tax(observed_price) [shipping ≥ 0 and unknown] $44,567.76 + $2,952.61 = $47,520.37 observed_price_cents 4456776 ¢ — Σ cheapest observed offer per part sales_tax_cents 295261 ¢ — SPEC §2: NJ sales tax 6.625% offers_without_shipping 7 offers — price_observations.shipping_cents IS NULL Sales tax rounded to the nearest cent, half-up (x.5¢ → x+1¢), in exact decimal arithmetic on the rate. Shipping is not published by every retailer here. It is omitted, not assumed to be zero — which is why this is a floor. lower bound — 7 of 7 offers do not publish a delivery charge (SabrePC, Newegg, SabrePC, SabrePC, ARCTIC, Newegg, Newegg), so this total is a lower bound on what the machine lands for, not an estimate of it.	$48,731 (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — landed floorArithmetic on observed inputs only. No tunable assumption enters it. landed_floor = observed_price + sales_tax(observed_price) [shipping ≥ 0 and unknown] $45,702.9 + $3,027.82 = $48,730.72 observed_price_cents 4570290 ¢ — Σ cheapest observed offer per part sales_tax_cents 302782 ¢ — SPEC §2: NJ sales tax 6.625% offers_without_shipping 7 offers — price_observations.shipping_cents IS NULL Sales tax rounded to the nearest cent, half-up (x.5¢ → x+1¢), in exact decimal arithmetic on the rate. Shipping is not published by every retailer here. It is omitted, not assumed to be zero — which is why this is a floor. lower bound — 7 of 7 offers do not publish a delivery charge (SabrePC, Newegg, SabrePC, SabrePC, ARCTIC, Newegg, Newegg), so this total is a lower bound on what the machine lands for, not an estimate of it.
Landed costprice + shipping + tax — what you would actually pay	$57,841 (derived) DERIVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Landed costArithmetic on observed inputs only. No tunable assumption enters it. rig_landed = Σ part landed_total $638 + $895 + $49,900 + $5,438 + $75 + $192 + $704 = $57,841 AMD EPYC 9015 63809 ¢ — SabrePC · ship ASRock Rack GENOAD8QM3-2T/BCM 89458 ¢ — Newegg · ship Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM 4990040 ¢ — Newegg · ship MSI GeForce RTX 5090 32G GAMING TRIO OC 543786 ¢ — Newegg · ship ARCTIC Freezer 4U-SP5 7463 ¢ — ARCTIC · ship Phanteks Enthoo Pro 2 Server Edition 19191 ¢ — Newegg · ship Seasonic PRIME PX-2200 70371 ¢ — Newegg · ship Some part prices are older than 36 h. LOWER BOUND: 7 of 7 parts come from a retailer that does not state a delivery charge (SabrePC, Newegg, Newegg, Newegg, ARCTIC, Newegg, Newegg), so the rig costs at least this and never less. BEST — the lowest landed cost of the machines compared	$71,596 (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Landed costArithmetic on observed inputs only. No tunable assumption enters it. rig_landed = Σ part landed_total $638 + $895 + $49,900 + $19,192 + $75 + $192 + $704 = $71,596 AMD EPYC 9015 63809 ¢ — SabrePC · ship ASRock Rack GENOAD8QM3-2T/BCM 89458 ¢ — Newegg · ship Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM 4990040 ¢ — Newegg · ship NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB 1919249 ¢ — Newegg · ship ARCTIC Freezer 4U-SP5 7463 ¢ — ARCTIC · ship Phanteks Enthoo Pro 2 Server Edition 19191 ¢ — Newegg · ship Seasonic PRIME PX-2200 70371 ¢ — Newegg · ship Some part prices are older than 36 h. LOWER BOUND: 7 of 7 parts come from a retailer that does not state a delivery charge (SabrePC, Newegg, Newegg, Newegg, ARCTIC, Newegg, Newegg), so the rig costs at least this and never less.	$65,124 (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Landed costArithmetic on observed inputs only. No tunable assumption enters it. rig_landed = Σ part landed_total $638 + $895 + $49,900 + $12,721 + $75 + $192 + $704 = $65,124 AMD EPYC 9015 63809 ¢ — SabrePC · ship ASRock Rack GENOAD8QM3-2T/BCM 89458 ¢ — Newegg · ship Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM 4990040 ¢ — Newegg · ship NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB 1272101 ¢ — SabrePC · ship ARCTIC Freezer 4U-SP5 7463 ¢ — ARCTIC · ship Phanteks Enthoo Pro 2 Server Edition 19191 ¢ — Newegg · ship Seasonic PRIME PX-2200 70371 ¢ — Newegg · ship Some part prices are older than 36 h. LOWER BOUND: 7 of 7 parts come from a retailer that does not state a delivery charge (SabrePC, Newegg, Newegg, SabrePC, ARCTIC, Newegg, Newegg), so the rig costs at least this and never less.
Cost per usable GBon the landed floor	$0.41 (derived) DERIVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Cost per usable GBArithmetic on observed inputs only. No tunable assumption enters it. cost_per_gb = landed_cost / usable_memory_gb = $41,448 / 1,003.84 GB = $41.29/GB landed_cost 41447.57 USD usable_memory_gb 1003.8399999999999 GB BEST — the least you pay per usable GB	$0.45 (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Cost per usable GBArithmetic on observed inputs only. No tunable assumption enters it. cost_per_gb = landed_cost / usable_memory_gb = $47,520 / 1,065.92 GB = $44.58/GB landed_cost 47520.37 USD usable_memory_gb 1065.92 GB	$0.46 (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Cost per usable GBArithmetic on observed inputs only. No tunable assumption enters it. cost_per_gb = landed_cost / usable_memory_gb = $48,731 / 1,065.92 GB = $45.72/GB landed_cost 48730.72 USD usable_memory_gb 1065.92 GB
Cost per GB/son the landed floor	$0.23 (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 = landed_cost / memory_bandwidth_gbs = $41,448 / 1,792 GB/s = $23.13/GB/s landed_cost 41447.57 USD memory_bandwidth_gbs 1792 GB/s BEST — the least you pay per GB/s of bandwidth	$0.27 (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 = landed_cost / memory_bandwidth_gbs = $47,520 / 1,792 GB/s = $26.52/GB/s landed_cost 47520.37 USD memory_bandwidth_gbs 1792 GB/s	$0.27 (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 = landed_cost / memory_bandwidth_gbs = $48,731 / 1,792 GB/s = $27.19/GB/s landed_cost 48730.72 USD memory_bandwidth_gbs 1792 GB/s
Cost per tok/s — GLM-5.3on the landed floor	$69 (derived) DERIVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — cost per tok/s on GLM-5.3Arithmetic on observed inputs only. No tunable assumption enters it. cost_per_tok_s = landed_cost / decode_tok_s = $41,448 / 5.97 tok/s = $6,940.07/tok/s landed_cost 41447.57 USD decode_tok_s 5.972210936466788 tok/s BEST — the least you pay per tok/s on GLM-5.3	$71 (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — cost per tok/s on GLM-5.3Arithmetic on observed inputs only. No tunable assumption enters it. cost_per_tok_s = landed_cost / decode_tok_s = $47,520 / 6.66 tok/s = $7,133.7/tok/s landed_cost 47520.37 USD decode_tok_s 6.661392885611014 tok/s	$73 (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — cost per tok/s on GLM-5.3Arithmetic on observed inputs only. No tunable assumption enters it. cost_per_tok_s = landed_cost / decode_tok_s = $48,731 / 6.66 tok/s = $7,315.39/tok/s landed_cost 48730.72 USD decode_tok_s 6.661392885611014 tok/s
Cost per tok/s — DeepSeek-V4-Pro-0813on the landed floor	$74 (derived) DERIVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — cost per tok/s on DeepSeek-V4-Pro-0813Arithmetic on observed inputs only. No tunable assumption enters it. cost_per_tok_s = landed_cost / decode_tok_s = $41,448 / 5.61 tok/s = $7,391.91/tok/s landed_cost 41447.57 USD decode_tok_s 5.607153782676166 tok/s BEST — the least you pay per tok/s on DeepSeek-V4-Pro-0813	$80 (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — cost per tok/s on DeepSeek-V4-Pro-0813Arithmetic on observed inputs only. No tunable assumption enters it. cost_per_tok_s = landed_cost / decode_tok_s = $47,520 / 5.92 tok/s = $8,026.53/tok/s landed_cost 47520.37 USD decode_tok_s 5.920413198217758 tok/s	$82 (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — cost per tok/s on DeepSeek-V4-Pro-0813Arithmetic on observed inputs only. No tunable assumption enters it. cost_per_tok_s = landed_cost / decode_tok_s = $48,731 / 5.92 tok/s = $8,230.97/tok/s landed_cost 48730.72 USD decode_tok_s 5.920413198217758 tok/s
Cost per tok/s — Qwen3-VL 32B Instructon the landed floor	$8 (derived) DERIVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — cost per tok/s on Qwen3-VL 32B InstructArithmetic on observed inputs only. No tunable assumption enters it. cost_per_tok_s = landed_cost / decode_tok_s = $41,448 / 52.67 tok/s = $786.95/tok/s landed_cost 41447.57 USD decode_tok_s 52.66829150819229 tok/s BEST — the least you pay per tok/s on Qwen3-VL 32B Instruct	$9 (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — cost per tok/s on Qwen3-VL 32B InstructArithmetic on observed inputs only. No tunable assumption enters it. cost_per_tok_s = landed_cost / decode_tok_s = $47,520 / 52.67 tok/s = $902.26/tok/s landed_cost 47520.37 USD decode_tok_s 52.66829150819229 tok/s	$9 (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — cost per tok/s on Qwen3-VL 32B InstructArithmetic on observed inputs only. No tunable assumption enters it. cost_per_tok_s = landed_cost / decode_tok_s = $48,731 / 52.67 tok/s = $925.24/tok/s landed_cost 48730.72 USD decode_tok_s 52.66829150819229 tok/s

Three money rows, not one, because they are three different things. Observed price is the sum of the cheapest price seen for each part — real, and complete. Landed floor adds NJ sales tax and leaves shipping out; it is therefore the least the machine can land for. Landed cost is price + shipping + tax, and it is what the app would rank on — it is unavailable for any machine whose retailer does not publish a delivery charge, and the cell says which. Every ratio below is computed on the FLOOR, so each is also a lower bound: the true cost per GB is higher by the shipping that is not published.

What each costs to own

over 60 months at 180 h a month (SPEC §8.5)

Net cost of ownership = purchase + electricity over the hold − projected resale. Two of those three terms are missing here.

Projected resale needs a depreciation curve fitted to observed used prices, and none of these machines has any — the used market has not been observed for them yet (SPEC §8.5 tracks it as a first-class price channel).
Electricity is computable wherever the parts publish their power draw, so it is shown as its own rows rather than being lost inside an unavailable total.

What each of the 3 machines costs to own over 60 months, and what renting the same capability would cost
Compared on	Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM7-part build · 3 blockers	Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM7-part build · 3 blockers	Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM7-part build · 3 blockers
Draw under loadpublished component figures	Unavailable — 5 parts of Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM publish no power figure, so its draw under load cannot be summed: ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition, Seasonic PRIME PX-2200. (missing: load_watts)	Unavailable — 5 parts of Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM publish no power figure, so its draw under load cannot be summed: ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition, Seasonic PRIME PX-2200. (missing: load_watts)	Unavailable — 5 parts of Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM publish no power figure, so its draw under load cannot be summed: ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition, Seasonic PRIME PX-2200. (missing: load_watts)
Draw at idlethe state it is in most of the day	Unavailable — 7 parts of Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM publish no idle figure, so the 18 idle hours a day cannot be costed: AMD EPYC 9015, ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, MSI GeForce RTX 5090 32G GAMING TRIO OC, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition, Seasonic PRIME PX-2200. (missing: idle_watts)	Unavailable — 7 parts of Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM publish no idle figure, so the 18 idle hours a day cannot be costed: AMD EPYC 9015, ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition, Seasonic PRIME PX-2200. (missing: idle_watts)	Unavailable — 7 parts of Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM publish no idle figure, so the 18 idle hours a day cannot be costed: AMD EPYC 9015, ASRock Rack GENOAD8QM3-2T/BCM, Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition, Seasonic PRIME PX-2200. (missing: idle_watts)
Energy a yearat the duty cycle in Settings	Unavailable — Annual energy needs the wall draw for every state the machine spends time in; unknown: load_wall_watts, idle_wall_watts. An unknown draw is not assumed to be 0 W. (missing: load_wall_watts, idle_wall_watts)	Unavailable — Annual energy needs the wall draw for every state the machine spends time in; unknown: load_wall_watts, idle_wall_watts. An unknown draw is not assumed to be 0 W. (missing: load_wall_watts, idle_wall_watts)	Unavailable — Annual energy needs the wall draw for every state the machine spends time in; unknown: load_wall_watts, idle_wall_watts. An unknown draw is not assumed to be 0 W. (missing: load_wall_watts, idle_wall_watts)
Electricity over 60 monthsat the NJ rate in Settings	Unavailable — annual_kwh is unavailable: Annual energy needs the wall draw for every state the machine spends time in; unknown: load_wall_watts, idle_wall_watts. An unknown draw is not assumed to be 0 W. (missing: load_wall_watts, idle_wall_watts)	Unavailable — annual_kwh is unavailable: Annual energy needs the wall draw for every state the machine spends time in; unknown: load_wall_watts, idle_wall_watts. An unknown draw is not assumed to be 0 W. (missing: load_wall_watts, idle_wall_watts)	Unavailable — annual_kwh is unavailable: Annual energy needs the wall draw for every state the machine spends time in; unknown: load_wall_watts, idle_wall_watts. An unknown draw is not assumed to be 0 W. (missing: load_wall_watts, idle_wall_watts)
Purchase + 5 years of powerSPEC §8.5 — the figure to show beside a sticker price	Unavailable — five_year_power_cents is unavailable: annual_kwh is unavailable: Annual energy needs the wall draw for every state the machine spends time in; unknown: load_wall_watts, idle_wall_watts. An unknown draw is not assumed to be 0 W. (missing: load_wall_watts, idle_wall_watts)	Unavailable — five_year_power_cents is unavailable: annual_kwh is unavailable: Annual energy needs the wall draw for every state the machine spends time in; unknown: load_wall_watts, idle_wall_watts. An unknown draw is not assumed to be 0 W. (missing: load_wall_watts, idle_wall_watts)	Unavailable — five_year_power_cents is unavailable: annual_kwh is unavailable: Annual energy needs the wall draw for every state the machine spends time in; unknown: load_wall_watts, idle_wall_watts. An unknown draw is not assumed to be 0 W. (missing: load_wall_watts, idle_wall_watts)
Projected resale at 60 monthswhat the used market would pay you back	Unavailable — Residual value cannot be computed. launch_price_cents is unavailable: No launch (list) price is recorded for Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, and the used market has not been observed for it, so there is no depreciation curve and no projected resale value. Residual value needs a depreciation curve: No used-price observations are on record for this machine, so no depreciation curve has been fitted and the resale value at the hold horizon cannot be projected. (missing: launch_price_cents, used price history)	Unavailable — Residual value cannot be computed. launch_price_cents is unavailable: No launch (list) price is recorded for Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, and the used market has not been observed for it, so there is no depreciation curve and no projected resale value. Residual value needs a depreciation curve: No used-price observations are on record for this machine, so no depreciation curve has been fitted and the resale value at the hold horizon cannot be projected. (missing: launch_price_cents, used price history)	Unavailable — Residual value cannot be computed. launch_price_cents is unavailable: No launch (list) price is recorded for Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, and the used market has not been observed for it, so there is no depreciation curve and no projected resale value. Residual value needs a depreciation curve: No used-price observations are on record for this machine, so no depreciation curve has been fitted and the resale value at the hold horizon cannot be projected. (missing: launch_price_cents, used price history)
Net cost of ownershippurchase + power − projected resale	Unavailable — Net cost of ownership cannot be computed. power_cents_over_horizon is unavailable: annual_kwh is unavailable: Annual energy needs the wall draw for every state the machine spends time in; unknown: load_wall_watts, idle_wall_watts. An unknown draw is not assumed to be 0 W. residual_cents is unavailable: Residual value cannot be computed. launch_price_cents is unavailable: No launch (list) price is recorded for Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, and the used market has not been observed for it, so there is no depreciation curve and no projected resale value. Residual value needs a depreciation curve: No used-price observations are on record for this machine, so no depreciation curve has been fitted and the resale value at the hold horizon cannot be projected. (missing: load_wall_watts, idle_wall_watts, launch_price_cents, used price history)	Unavailable — Net cost of ownership cannot be computed. power_cents_over_horizon is unavailable: annual_kwh is unavailable: Annual energy needs the wall draw for every state the machine spends time in; unknown: load_wall_watts, idle_wall_watts. An unknown draw is not assumed to be 0 W. residual_cents is unavailable: Residual value cannot be computed. launch_price_cents is unavailable: No launch (list) price is recorded for Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, and the used market has not been observed for it, so there is no depreciation curve and no projected resale value. Residual value needs a depreciation curve: No used-price observations are on record for this machine, so no depreciation curve has been fitted and the resale value at the hold horizon cannot be projected. (missing: load_wall_watts, idle_wall_watts, launch_price_cents, used price history)	Unavailable — Net cost of ownership cannot be computed. power_cents_over_horizon is unavailable: annual_kwh is unavailable: Annual energy needs the wall draw for every state the machine spends time in; unknown: load_wall_watts, idle_wall_watts. An unknown draw is not assumed to be 0 W. residual_cents is unavailable: Residual value cannot be computed. launch_price_cents is unavailable: No launch (list) price is recorded for Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, and the used market has not been observed for it, so there is no depreciation curve and no projected resale value. Residual value needs a depreciation curve: No used-price observations are on record for this machine, so no depreciation curve has been fitted and the resale value at the hold horizon cannot be projected. (missing: load_wall_watts, idle_wall_watts, launch_price_cents, used price history)
Nearest cloud instancethe same memory and bandwidth, rented	CoreWeave AMD Genoa 9654 (High Core) $7.54/h (observed) OBSERVEDCoreWeave AMD Genoa 9654 (High Core) hourly rateMeasured or published, passed through unchanged — not computed here. cloud_hourly_equiv = cheapest on-demand instance with memory ≥ build and bandwidth ≥ build CoreWeave AMD Genoa 9654 (High Core): 1,536 GB ≥ 1,004 GB, bandwidth unpublished; $7.54/h observed 2026-10-18 (cheapest of 22 matched on memory alone) build_memory_gb 1003.8399999999999 GB build_bandwidth_gbs 1792 GB/s instance_memory_gb 1536 GB — CoreWeave AMD Genoa 9654 (High Core) instance_rate_cents 754 ¢/h — https://www.coreweave.com/pricing instance_observed_at 2026-10-18T20:08:08.498Z — https://www.coreweave.com/pricing Cheapest ON-DEMAND instance, at its latest observed price, whose memory ≥ the build's memory (the model must fit) and whose bandwidth ≥ the build's bandwidth (decode at least as fast). Instances with unpublished bandwidth are used only if no instance with a verified bandwidth qualifies. Spot is excluded (pre-emptible: not equivalent availability); per-token APIs are compared separately. Rate observed 2026-10-18T20:08:08.498Z — the latest of 2 dated observation(s) of this instance on file; its age is the time since then. Instance memory is the memory a model can live in (total GPU memory for GPU instances); a model split across several GPUs needs a runtime that shards it. Instance bandwidth summed across GPUs is only reached with tensor parallelism; with a layer split, per-token bandwidth is one GPU's. List on-demand price; reserved, committed-use and marketplace discounts are not applied. CoreWeave AMD Genoa 9654 (High Core) does not publish its memory bandwidth: matched on memory alone, so it may decode slower than the build. Region: North America. matched on memory alone — this instance does not publish its memory bandwidth, so it was matched on capacity only and may decode more slowly than the machine it stands in for.	CoreWeave AMD Genoa 9654 (High Core) $7.54/h (observed) OBSERVEDCoreWeave AMD Genoa 9654 (High Core) hourly rateMeasured or published, passed through unchanged — not computed here. cloud_hourly_equiv = cheapest on-demand instance with memory ≥ build and bandwidth ≥ build CoreWeave AMD Genoa 9654 (High Core): 1,536 GB ≥ 1,066 GB, bandwidth unpublished; $7.54/h observed 2026-10-18 (cheapest of 22 matched on memory alone) build_memory_gb 1065.92 GB build_bandwidth_gbs 1792 GB/s instance_memory_gb 1536 GB — CoreWeave AMD Genoa 9654 (High Core) instance_rate_cents 754 ¢/h — https://www.coreweave.com/pricing instance_observed_at 2026-10-18T20:08:08.498Z — https://www.coreweave.com/pricing Cheapest ON-DEMAND instance, at its latest observed price, whose memory ≥ the build's memory (the model must fit) and whose bandwidth ≥ the build's bandwidth (decode at least as fast). Instances with unpublished bandwidth are used only if no instance with a verified bandwidth qualifies. Spot is excluded (pre-emptible: not equivalent availability); per-token APIs are compared separately. Rate observed 2026-10-18T20:08:08.498Z — the latest of 2 dated observation(s) of this instance on file; its age is the time since then. Instance memory is the memory a model can live in (total GPU memory for GPU instances); a model split across several GPUs needs a runtime that shards it. Instance bandwidth summed across GPUs is only reached with tensor parallelism; with a layer split, per-token bandwidth is one GPU's. List on-demand price; reserved, committed-use and marketplace discounts are not applied. CoreWeave AMD Genoa 9654 (High Core) does not publish its memory bandwidth: matched on memory alone, so it may decode slower than the build. Region: North America. matched on memory alone — this instance does not publish its memory bandwidth, so it was matched on capacity only and may decode more slowly than the machine it stands in for.	CoreWeave AMD Genoa 9654 (High Core) $7.54/h (observed) OBSERVEDCoreWeave AMD Genoa 9654 (High Core) hourly rateMeasured or published, passed through unchanged — not computed here. cloud_hourly_equiv = cheapest on-demand instance with memory ≥ build and bandwidth ≥ build CoreWeave AMD Genoa 9654 (High Core): 1,536 GB ≥ 1,066 GB, bandwidth unpublished; $7.54/h observed 2026-10-18 (cheapest of 22 matched on memory alone) build_memory_gb 1065.92 GB build_bandwidth_gbs 1792 GB/s instance_memory_gb 1536 GB — CoreWeave AMD Genoa 9654 (High Core) instance_rate_cents 754 ¢/h — https://www.coreweave.com/pricing instance_observed_at 2026-10-18T20:08:08.498Z — https://www.coreweave.com/pricing Cheapest ON-DEMAND instance, at its latest observed price, whose memory ≥ the build's memory (the model must fit) and whose bandwidth ≥ the build's bandwidth (decode at least as fast). Instances with unpublished bandwidth are used only if no instance with a verified bandwidth qualifies. Spot is excluded (pre-emptible: not equivalent availability); per-token APIs are compared separately. Rate observed 2026-10-18T20:08:08.498Z — the latest of 2 dated observation(s) of this instance on file; its age is the time since then. Instance memory is the memory a model can live in (total GPU memory for GPU instances); a model split across several GPUs needs a runtime that shards it. Instance bandwidth summed across GPUs is only reached with tensor parallelism; with a layer split, per-token bandwidth is one GPU's. List on-demand price; reserved, committed-use and marketplace discounts are not applied. CoreWeave AMD Genoa 9654 (High Core) does not publish its memory bandwidth: matched on memory alone, so it may decode slower than the build. Region: North America. matched on memory alone — this instance does not publish its memory bandwidth, so it was matched on capacity only and may decode more slowly than the machine it stands in for.
Renting instead, 60 months180 h a month, compute only	$81,432 (estimate) ESTIMATEHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Renting instead, 60 monthsArithmetic that also depends on a tunable assumption, listed below. rent_over_horizon = hourly_rate × monthly_usage_hours × hold_horizon_months $7.54/h × 180 h/month × 60 months = $81,432 hourly_cents 754 ¢/h — CoreWeave AMD Genoa 9654 (High Core) on-demand monthly_usage_hours 180 h/month — Settings → Power → monthly usage hours hold_horizon_months 60 months — Settings → Power → hold horizon Compute only: storage, egress and any data transfer are extra and are not in this figure. The rate is the latest observation of that instance; cloud prices move, and this one is not projected forward. Renting leaves you nothing at the end; buying leaves you a machine, which is what the residual term of net cost of ownership would capture.	$81,432 (estimate) ESTIMATEHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Renting instead, 60 monthsArithmetic that also depends on a tunable assumption, listed below. rent_over_horizon = hourly_rate × monthly_usage_hours × hold_horizon_months $7.54/h × 180 h/month × 60 months = $81,432 hourly_cents 754 ¢/h — CoreWeave AMD Genoa 9654 (High Core) on-demand monthly_usage_hours 180 h/month — Settings → Power → monthly usage hours hold_horizon_months 60 months — Settings → Power → hold horizon Compute only: storage, egress and any data transfer are extra and are not in this figure. The rate is the latest observation of that instance; cloud prices move, and this one is not projected forward. Renting leaves you nothing at the end; buying leaves you a machine, which is what the residual term of net cost of ownership would capture.	$81,432 (estimate) ESTIMATEHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM — Renting instead, 60 monthsArithmetic that also depends on a tunable assumption, listed below. rent_over_horizon = hourly_rate × monthly_usage_hours × hold_horizon_months $7.54/h × 180 h/month × 60 months = $81,432 hourly_cents 754 ¢/h — CoreWeave AMD Genoa 9654 (High Core) on-demand monthly_usage_hours 180 h/month — Settings → Power → monthly usage hours hold_horizon_months 60 months — Settings → Power → hold horizon Compute only: storage, egress and any data transfer are extra and are not in this figure. The rate is the latest observation of that instance; cloud prices move, and this one is not projected forward. Renting leaves you nothing at the end; buying leaves you a machine, which is what the residual term of net cost of ownership would capture.
Breakeven against rentingmonths of use before owning is cheaper	Unavailable — net_cost_of_ownership is unavailable: Net cost of ownership cannot be computed. power_cents_over_horizon is unavailable: annual_kwh is unavailable: Annual energy needs the wall draw for every state the machine spends time in; unknown: load_wall_watts, idle_wall_watts. An unknown draw is not assumed to be 0 W. residual_cents is unavailable: Residual value cannot be computed. launch_price_cents is unavailable: No launch (list) price is recorded for Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, and the used market has not been observed for it, so there is no depreciation curve and no projected resale value. Residual value needs a depreciation curve: No used-price observations are on record for this machine, so no depreciation curve has been fitted and the resale value at the hold horizon cannot be projected. (missing: load_wall_watts, idle_wall_watts, launch_price_cents, used price history)	Unavailable — net_cost_of_ownership is unavailable: Net cost of ownership cannot be computed. power_cents_over_horizon is unavailable: annual_kwh is unavailable: Annual energy needs the wall draw for every state the machine spends time in; unknown: load_wall_watts, idle_wall_watts. An unknown draw is not assumed to be 0 W. residual_cents is unavailable: Residual value cannot be computed. launch_price_cents is unavailable: No launch (list) price is recorded for Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, and the used market has not been observed for it, so there is no depreciation curve and no projected resale value. Residual value needs a depreciation curve: No used-price observations are on record for this machine, so no depreciation curve has been fitted and the resale value at the hold horizon cannot be projected. (missing: load_wall_watts, idle_wall_watts, launch_price_cents, used price history)	Unavailable — net_cost_of_ownership is unavailable: Net cost of ownership cannot be computed. power_cents_over_horizon is unavailable: annual_kwh is unavailable: Annual energy needs the wall draw for every state the machine spends time in; unknown: load_wall_watts, idle_wall_watts. An unknown draw is not assumed to be 0 W. residual_cents is unavailable: Residual value cannot be computed. launch_price_cents is unavailable: No launch (list) price is recorded for Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, and the used market has not been observed for it, so there is no depreciation curve and no projected resale value. Residual value needs a depreciation curve: No used-price observations are on record for this machine, so no depreciation curve has been fitted and the resale value at the hold horizon cannot be projected. (missing: load_wall_watts, idle_wall_watts, launch_price_cents, used price history)

Net cost of ownership is purchase + electricity over the hold − projected resale (SPEC §8.5), all undiscounted. The hold horizon (60 months), the duty cycle, the monthly usage hours and the electricity rate are Settings inputs, not constants — the rate cited here is: PSE&G Rate Schedule RS (residential), BGS-RSCP default supply, all-in incl. 6.625% NJ SUT; tariff B.P.U.N.J. No. 17 and BGS-RSCP Price to Compare, both effective 2026-07-15. Derived always-on marginal annual rate = (4 x 0.299218 [Jun-Sep >600 kWh] + 8 x 0.253411 [Oct-May]) / 12 = $0.26868/kWh = 26.868 c/kWh (26.416 c if summer bills stay <=600 kWh); excludes the $6.00/month fixed service charge. 08550 interior point is PSE&G territory; ~18% of the ZIP (SE part) is JCP&L - confirm on the bill. Sources: https://nj.pseg.com/-/media/pseg/public-site/documents/current-electric-tariff/electric-tariff-17-transmission-effective-20260715.ashx ; https://nj.pseg.com/-/media/pseg/public-site/documents/current-electric-tariff/electric-ptc-website-20260715-bgs-nits--tec.ashx (retrieved 2026-09-19; docs/research/13 §A2, verifier-confirmed to 6 decimals). A row that is unavailable names the term that is missing rather than dropping it: a net cost without a resale value would understate what owning costs, and a zero in its place would be a fabrication.

Where every price came from

provenance and age are permanent, not hover · stale past 36 h

Where each machine's price came from, how old it is, when it could be delivered, and whether it is buildable
Compared on	Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM7-part build · 3 blockers	Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM7-part build · 3 blockers	Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM7-part build · 3 blockers
Priced fromretailer · condition · access tier / route	SabrePCcondition newaccess tier 2, route sabrepc_ngstate Neweggcondition newaccess tier 3, route newegg_realtime SabrePCcondition newaccess tier 2, route sabrepc_ngstate Neweggcondition newaccess tier 3, route newegg_realtime ARCTICcondition newaccess tier 1, route arctic_jsonld Neweggcondition newaccess tier 3, route newegg_realtime Neweggcondition newaccess tier 3, route newegg_realtime	SabrePCcondition newaccess tier 2, route sabrepc_ngstate Neweggcondition newaccess tier 3, route newegg_realtime SabrePCcondition newaccess tier 2, route sabrepc_ngstate SabrePCcondition newaccess tier 2, route sabrepc_ngstate ARCTICcondition newaccess tier 1, route arctic_jsonld Neweggcondition newaccess tier 3, route newegg_realtime Neweggcondition newaccess tier 3, route newegg_realtime	SabrePCcondition newaccess tier 2, route sabrepc_ngstate Neweggcondition newaccess tier 3, route newegg_realtime SabrePCcondition newaccess tier 2, route sabrepc_ngstate SabrePCcondition newaccess tier 2, route sabrepc_ngstate ARCTICcondition newaccess tier 1, route arctic_jsonld Neweggcondition newaccess tier 3, route newegg_realtime Neweggcondition newaccess tier 3, route newegg_realtime
Price observedcaution past the 36 h freshness window	stale 40 h — past the 36 hour freshness window 27 h stale 40 h — past the 36 hour freshness window 27 h stale 41 h — past the 36 hour freshness window 27 h 27 h	stale 40 h — past the 36 hour freshness window 27 h stale 40 h — past the 36 hour freshness window stale 41 h — past the 36 hour freshness window stale 41 h — past the 36 hour freshness window 27 h 27 h	stale 40 h — past the 36 hour freshness window 27 h stale 40 h — past the 36 hour freshness window stale 41 h — past the 36 hour freshness window stale 41 h — past the 36 hour freshness window 27 h 27 h
When you could have itavailability is its own axis, never folded into price	Not every part in stock SabrePC: out of stock Newegg: out of stock	Not every part in stock SabrePC: out of stock Newegg: out of stock	Not every part in stock SabrePC: out of stock SabrePC: out of stock
CompatibilitySPEC §9.2 — errors block, warnings do not	0 errors, 2 warnings Board wires 8 of the CPU's 12 memory channels — AMD EPYC 9015 has 12 channels; ASRock Rack GENOAD8QM3-2T/BCM wires 8. Peak system-RAM bandwidth is capped at 307.2 GB/s instead of 460.8 GB/s (−33% decode on RAM-resident models), whatever memory is installed. Choose a board that wires all 12 channels. Electrical feasibility not confirmed — The 15 A / 120 V circuit check cannot be completed: 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. The parts with data draw at least 760.87 W at the wall (1,440 W usable). Add the missing power figures (load or TDP watts from the maker's spec sheet) so the check can run.	0 errors, 2 warnings Board wires 8 of the CPU's 12 memory channels — AMD EPYC 9015 has 12 channels; ASRock Rack GENOAD8QM3-2T/BCM wires 8. Peak system-RAM bandwidth is capped at 307.2 GB/s instead of 460.8 GB/s (−33% decode on RAM-resident models), whatever memory is installed. Choose a board that wires all 12 channels. Electrical feasibility not confirmed — The 15 A / 120 V circuit check cannot be completed: 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. The parts with data draw at least 788.04 W at the wall (1,440 W usable). Add the missing power figures (load or TDP watts from the maker's spec sheet) so the check can run.	0 errors, 2 warnings Board wires 8 of the CPU's 12 memory channels — AMD EPYC 9015 has 12 channels; ASRock Rack GENOAD8QM3-2T/BCM wires 8. Peak system-RAM bandwidth is capped at 307.2 GB/s instead of 460.8 GB/s (−33% decode on RAM-resident models), whatever memory is installed. Choose a board that wires all 12 channels. Electrical feasibility not confirmed — The 15 A / 120 V circuit check cannot be completed: 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. The parts with data draw at least 461.96 W at the wall (1,440 W usable). Add the missing power figures (load or TDP watts from the maker's spec sheet) so the check can run.
Wall draw and the circuita build over the circuit is a hard error, not a cost	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)	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)	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)

Provenance is permanent, not a tooltip: each price states the retailer, the condition, and the access tier and route key that produced it (SPEC §5.4), and its age is a first-class state — past the 36 h window in Settings it renders in caution ink with the word “stale” and a clock. Availability is its own row and never folded into price: a machine that is cheaper but ships in eight weeks is a different decision (SPEC §8.5).

Parts, machine by machine

every part, its observed price, and the offer it came from

Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM7 parts · 7 priced · $38,872 observed

Every part of Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, with the observed price and where it came from
Part	Observed unit price	Shipping	Where the price came from	Observed
AMD EPYC 90151 unit	$598	Unavailable — SabrePC does not publish a delivery charge for this part. It is left out of the total rather than assumed to be zero, which is why the landed cost is unavailable. (missing: shipping_cents)	SabrePCcondition newaccess tier 2, route sabrepc_ngstate	stale 40 h — past the 36 hour freshness window
ASRock Rack GENOAD8QM3-2T/BCM1 unit	$839	Unavailable — Newegg does not publish a delivery charge for this part. It is left out of the total rather than assumed to be zero, which is why the landed cost is unavailable. (missing: shipping_cents)	Neweggcondition newaccess tier 3, route newegg_realtime	27 h
Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM8 units	$3,928	Unavailable — SabrePC does not publish a delivery charge for this part. It is left out of the total rather than assumed to be zero, which is why the landed cost is unavailable. (missing: shipping_cents)	SabrePCcondition newaccess tier 2, route sabrepc_ngstate	stale 40 h — past the 36 hour freshness window
MSI GeForce RTX 5090 32G GAMING TRIO OC1 unit	$5,100	Unavailable — Newegg does not publish a delivery charge for this part. It is left out of the total rather than assumed to be zero, which is why the landed cost is unavailable. (missing: shipping_cents)	Neweggcondition newaccess tier 3, route newegg_realtime	27 h
ARCTIC Freezer 4U-SP51 unit	$70	Unavailable — ARCTIC does not publish a delivery charge for this part. It is left out of the total rather than assumed to be zero, which is why the landed cost is unavailable. (missing: shipping_cents)	ARCTICcondition newaccess tier 1, route arctic_jsonld	stale 41 h — past the 36 hour freshness window
Phanteks Enthoo Pro 2 Server Edition1 unit	$180	Unavailable — Newegg does not publish a delivery charge for this part. It is left out of the total rather than assumed to be zero, which is why the landed cost is unavailable. (missing: shipping_cents)	Neweggcondition newaccess tier 3, route newegg_realtime	27 h
Seasonic PRIME PX-22001 unit	$660	Unavailable — Newegg does not publish a delivery charge for this part. It is left out of the total rather than assumed to be zero, which is why the landed cost is unavailable. (missing: shipping_cents)	Neweggcondition newaccess tier 3, route newegg_realtime	27 h

Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM7 parts · 7 priced · $44,568 observed

Every part of Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, with the observed price and where it came from
Part	Observed unit price	Shipping	Where the price came from	Observed
AMD EPYC 90151 unit	$598	Unavailable — SabrePC does not publish a delivery charge for this part. It is left out of the total rather than assumed to be zero, which is why the landed cost is unavailable. (missing: shipping_cents)	SabrePCcondition newaccess tier 2, route sabrepc_ngstate	stale 40 h — past the 36 hour freshness window
ASRock Rack GENOAD8QM3-2T/BCM1 unit	$839	Unavailable — Newegg does not publish a delivery charge for this part. It is left out of the total rather than assumed to be zero, which is why the landed cost is unavailable. (missing: shipping_cents)	Neweggcondition newaccess tier 3, route newegg_realtime	27 h
Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM8 units	$3,928	Unavailable — SabrePC does not publish a delivery charge for this part. It is left out of the total rather than assumed to be zero, which is why the landed cost is unavailable. (missing: shipping_cents)	SabrePCcondition newaccess tier 2, route sabrepc_ngstate	stale 40 h — past the 36 hour freshness window
NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB1 unit	$10,795	Unavailable — SabrePC does not publish a delivery charge for this part. It is left out of the total rather than assumed to be zero, which is why the landed cost is unavailable. (missing: shipping_cents)	SabrePCcondition newaccess tier 2, route sabrepc_ngstate	stale 41 h — past the 36 hour freshness window
ARCTIC Freezer 4U-SP51 unit	$70	Unavailable — ARCTIC does not publish a delivery charge for this part. It is left out of the total rather than assumed to be zero, which is why the landed cost is unavailable. (missing: shipping_cents)	ARCTICcondition newaccess tier 1, route arctic_jsonld	stale 41 h — past the 36 hour freshness window
Phanteks Enthoo Pro 2 Server Edition1 unit	$180	Unavailable — Newegg does not publish a delivery charge for this part. It is left out of the total rather than assumed to be zero, which is why the landed cost is unavailable. (missing: shipping_cents)	Neweggcondition newaccess tier 3, route newegg_realtime	27 h
Seasonic PRIME PX-22001 unit	$660	Unavailable — Newegg does not publish a delivery charge for this part. It is left out of the total rather than assumed to be zero, which is why the landed cost is unavailable. (missing: shipping_cents)	Neweggcondition newaccess tier 3, route newegg_realtime	27 h

Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM7 parts · 7 priced · $45,703 observed

Every part of Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM, with the observed price and where it came from
Part	Observed unit price	Shipping	Where the price came from	Observed
AMD EPYC 90151 unit	$598	Unavailable — SabrePC does not publish a delivery charge for this part. It is left out of the total rather than assumed to be zero, which is why the landed cost is unavailable. (missing: shipping_cents)	SabrePCcondition newaccess tier 2, route sabrepc_ngstate	stale 40 h — past the 36 hour freshness window
ASRock Rack GENOAD8QM3-2T/BCM1 unit	$839	Unavailable — Newegg does not publish a delivery charge for this part. It is left out of the total rather than assumed to be zero, which is why the landed cost is unavailable. (missing: shipping_cents)	Neweggcondition newaccess tier 3, route newegg_realtime	27 h
Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM8 units	$3,928	Unavailable — SabrePC does not publish a delivery charge for this part. It is left out of the total rather than assumed to be zero, which is why the landed cost is unavailable. (missing: shipping_cents)	SabrePCcondition newaccess tier 2, route sabrepc_ngstate	stale 40 h — past the 36 hour freshness window
NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB1 unit	$11,931	Unavailable — SabrePC does not publish a delivery charge for this part. It is left out of the total rather than assumed to be zero, which is why the landed cost is unavailable. (missing: shipping_cents)	SabrePCcondition newaccess tier 2, route sabrepc_ngstate	stale 41 h — past the 36 hour freshness window
ARCTIC Freezer 4U-SP51 unit	$70	Unavailable — ARCTIC does not publish a delivery charge for this part. It is left out of the total rather than assumed to be zero, which is why the landed cost is unavailable. (missing: shipping_cents)	ARCTICcondition newaccess tier 1, route arctic_jsonld	stale 41 h — past the 36 hour freshness window
Phanteks Enthoo Pro 2 Server Edition1 unit	$180	Unavailable — Newegg does not publish a delivery charge for this part. It is left out of the total rather than assumed to be zero, which is why the landed cost is unavailable. (missing: shipping_cents)	Neweggcondition newaccess tier 3, route newegg_realtime	27 h
Seasonic PRIME PX-22001 unit	$660	Unavailable — Newegg does not publish a delivery charge for this part. It is left out of the total rather than assumed to be zero, which is why the landed cost is unavailable. (missing: shipping_cents)	Neweggcondition newaccess tier 3, route newegg_realtime	27 h

How these numbers were produced

the assumptions behind every cell above

Every machine is evaluated by the same function the Plan, reverse mode, the nightly fitness job and the acceptance gate call, so a tok/s here is the tok/s scored everywhere else in this app.
KV cache at 2 bytes per element, runtime overhead 8% of weights, prefill MFU 0.35. All three are Settings inputs, not constants.
Sales tax 6.625% — SPEC §2: NJ sales tax 6.625%.
Duty cycle 6 h at load and 18 h idle a day; hold horizon 60 months; 180 h a month of use for the rent comparison. All Settings inputs.
Prices are the cheapest monitored, non-quarantined observation per part. A quarantined observation never prices a machine, and an offer that does not publish shipping is left out of the landed cost rather than having shipping assumed to be zero.

Machines in this comparison