{ "generated_at": "2026-05-08T05:29:24.238915Z", "electricity_usd_per_kwh": 0.3148, "assumptions": { "amortization_24x7_hours": 26280, "amortization_40h_hours": 6240, "beastmode_note": "BeastMode cost-per-token and hourly models are lower-bound capital-only estimates because per-host power telemetry is not exposed cleanly through the ESXi management path used here." }, "study": { "title": "Calculating the Total Cost of a GPU Cluster", "pricing_snapshot": "February 2026", "scenario_1_ratio": { "Nebius": 1.0, "AWS": 1.09, "Silver-tier": 1.08 }, "scenario_2_ratio": { "Nebius": 1.0, "AWS": 1.43, "Silver-tier": 1.08 }, "scenario_3_ratio": { "Nebius": 1.0, "AWS": 2.13, "Silver-tier": 1.04 }, "scenario_3_note": "The study attributes the AWS premium in inference to GPU instance price, support, storage, and setup overhead even after assuming a large discount off p5en list price." }, "local": { "mac_mini": { "label": "Mac mini", "price_usd": 1609.0, "price_note": "Apple Certified Refurbished current-market anchor for an M4 Pro / 48GB / 512GB class system; the cited Apple listing uses 10GbE, which slightly overstates this exact box.", "power_w": 155, "power_note": "Apple maximum continuous power from current Mac mini technical specifications.", "specs": "Apple M4 Pro, 12 CPU cores (8P+4E), 48GB unified memory, 273 GB/s memory bandwidth.", "sample_model": "qwen2.5:7b", "sample_tokens_per_s": 47.76560945427192, "sample_prompt_tokens": 52, "sample_output_tokens": 20, "sample_total_s": 3.953212584, "interactive_rps": 0.265, "interactive_note": "Observed at concurrency 4 in the controlled text-capacity benchmark." }, "dgx_spark": { "label": "DGX Spark", "price_usd": 4699.0, "price_note": "Current NVIDIA Marketplace listed price for DGX Spark bundle.", "power_w": 240, "power_note": "Conservative upper bound using the bundled 240W external power supply; actual sustained draw is often lower.", "specs": "NVIDIA GB10 Grace Blackwell Superchip, up to 1 PFLOP FP4, 128GB unified memory, 20-core Arm CPU, 273 GB/s memory bandwidth.", "sample_model": "qwen3:8b", "sample_tokens_per_s": 41.974065544911795, "sample_prompt_tokens": 39, "sample_output_tokens": 29, "sample_total_s": 2.735895899, "interactive_rps": 0.135, "interactive_note": "Observed at concurrency 2 in the controlled text-capacity benchmark comfort band." } }, "beastmode": { "chewbacuh": { "label": "Chewbacuh", "host_price_usd": 977.0, "host_price_note": "Current Newegg market listing for a similar R730xd with 2x E5-2620 v3 and 128GB RAM; host share is conservative because the lane uses only part of the box.", "share_fraction": 0.3333333333333333, "vm_vcpu": 8, "vm_ram_gib": 48, "sample_model": "qwen3:8b", "sample_tokens_per_s": 3.316102255849126, "benchmark_rps": 0.098, "benchmark_note": "Observed plateau throughput across concurrency 1-4." }, "lil_beastly": { "label": "LiL-Beastly", "host_price_usd": 1339.7, "host_price_note": "Current TechMikeNY market listing for a 512GB R730xd-class system; lane share priced by vCPU fraction because the host is shared infrastructure.", "share_fraction": 0.5, "vm_vcpu": 12, "vm_ram_gib": 96, "sample_model": "qwen3:14b", "sample_tokens_per_s": 1.776796715325552, "benchmark_rps": 0.051, "benchmark_note": "Observed plateau throughput across concurrency 1-4." } }, "cloud": { "runpod_l4": { "label": "Runpod L4", "hourly_usd": 0.39, "tier": "Budget inference GPU", "spec_note": "24GB L4 on Secure Cloud, on-demand starting price." }, "nebius_l40s": { "label": "Nebius L40S", "hourly_usd": 1.35, "tier": "Mid-range enterprise GPU", "spec_note": "Per-GPU price; CPU and RAM are additional on L40S VM pricing." }, "nebius_h200": { "label": "Nebius H200", "hourly_usd": 3.5, "tier": "Premium inference GPU", "spec_note": "Unified per-GPU hour price including vCPU and RAM on H200 VMs." }, "aws_p5en_h200": { "label": "AWS p5en H200", "hourly_usd": 5.721, "tier": "Hyperscaler premium GPU", "spec_note": "Effective per-accelerator hourly rate from the current Capacity Blocks for ML pricing table." }, "nebius_cpu": { "label": "Nebius CPU-only equivalent", "cpu_hour_usd": 0.012, "ram_gibhour_usd": 0.0032, "tier": "CPU cloud baseline", "spec_note": "Non-GPU AMD EPYC Genoa VM pricing." } }, "sources": [ { "label": "SemiAnalysis / Nebius study PDF used for TCO context", "url": "file:///Users/nation/Downloads/nebius-semianalysis-real-cost-of-gpu-clusters.pdf", "note": "Local study provided by user; pricing snapshot in the study is February 2026." }, { "label": "Apple Certified Refurbished Mac mini (M4 Pro / 48GB / 512GB / 10GbE)", "url": "https://www.apple.com/shop/product/g1kzmll/a/Refurbished-Mac-mini-Apple-M4-Pro-Chip-with-12-Core-CPU-and-16-Core-GPU-10Gb-Ethernet-", "note": "Used as a current-market anchor for the Mac mini capital cost." }, { "label": "Apple Mac mini technical specifications", "url": "https://www.apple.com/mac-mini/specs/", "note": "Used for M4 Pro architecture details and maximum continuous power." }, { "label": "NVIDIA DGX Spark marketplace listing", "url": "https://marketplace.nvidia.com/en-us/enterprise/personal-ai-supercomputers/dgx-spark/", "note": "Used for current published DGX Spark price." }, { "label": "NVIDIA DGX Spark hardware overview", "url": "https://docs.nvidia.com/dgx/dgx-spark/hardware.html", "note": "Used for DGX Spark system specs and power envelope." }, { "label": "Runpod L4 cloud pricing", "url": "https://www.runpod.io/gpu-models/l4", "note": "Used for budget L4 cloud baseline." }, { "label": "Nebius AI Cloud compute pricing", "url": "https://docs.nebius.com/compute/resources/pricing", "note": "Used for L40S, H200, and CPU-only cloud baselines." }, { "label": "AWS EC2 Capacity Blocks for ML pricing", "url": "https://aws.amazon.com/ec2/capacityblocks/pricing/", "note": "Used for p5en H200 per-accelerator pricing." }, { "label": "U.S. EIA electricity price table, California YTD through Feb 2026", "url": "https://www.eia.gov/electricity/monthly/epm_table_grapher.php?t=epmt_5_06_b", "note": "Used for local power-cost assumption of $0.3148/kWh." }, { "label": "Newegg Dell R730xd market listing", "url": "https://www.newegg.com/dell-poweredge-r730xd-rack/p/2NS-0008-703C5", "note": "Used as the vm1/chewbacuh host replacement-cost anchor." }, { "label": "TechMikeNY Dell R730xd market listing", "url": "https://techmikeny.com/products/dell-poweredge-r730xd-server-24-bay-2-20ghz-40-core-512gb-ram-26x-caddies", "note": "Used as the vm2/LiL-Beastly host replacement-cost anchor." } ], "computed": { "local_rows": [ { "id": "mac_mini", "label": "Mac mini", "hourly_24x7": 0.11001926636225268, "hourly_40h": 0.30664656410256413, "capex_24x7": 0.061225266362252666, "capex_40h": 0.25785256410256413, "tokens_per_s": 47.76560945427192, "million_tokens_per_hour": 0.1719561940353789, "cost_per_mtok_24x7": 0.6398098479640513, "cost_per_mtok_40h": 1.7832830380013733 }, { "id": "dgx_spark", "label": "DGX Spark", "hourly_24x7": 0.2543571750380518, "hourly_40h": 0.8285968717948717, "capex_24x7": 0.17880517503805174, "capex_40h": 0.7530448717948718, "tokens_per_s": 41.974065544911795, "million_tokens_per_hour": 0.15110663596168245, "cost_per_mtok_24x7": 1.6832958620199285, "cost_per_mtok_40h": 5.483524045925997 } ], "beast_rows": [ { "id": "chewbacuh", "label": "Chewbacuh", "lane_capex_usd": 325.66666666666663, "capex_24x7": 0.012392186707255198, "capex_40h": 0.052190170940170935, "tokens_per_s": 3.316102255849126, "million_tokens_per_hour": 0.011937968121056854, "cost_per_mtok_24x7": 1.038048232462371, "cost_per_mtok_40h": 4.371780055947293, "equivalent_cloud_hourly": 0.24960000000000002, "breakeven_hours_vs_cpu_cloud": 1304.7542735042732 }, { "id": "lil_beastly", "label": "LiL-Beastly", "lane_capex_usd": 669.85, "capex_24x7": 0.025488964992389652, "capex_40h": 0.10734775641025641, "tokens_per_s": 1.776796715325552, "million_tokens_per_hour": 0.006396468175171987, "cost_per_mtok_24x7": 3.984849809981945, "cost_per_mtok_40h": 16.78234823819319, "equivalent_cloud_hourly": 0.45120000000000005, "breakeven_hours_vs_cpu_cloud": 1484.5966312056737 } ], "pdf": "committee-review-cloud-economics-20260507.pdf", "charts": [ "beastmode-cpu-cloud.png", "breakeven-days.png", "local-token-cost.png", "local-vs-cloud-hourly.png" ] } }