Cost and equipment schedule
All figures derived, none quotedThe price is low for one structural reason: we own the accelerators instead of renting them, at $3.45 a GPU hour against $5.50 to rent the same board. Below is the whole calculation, not a summary of one — change an assumption and every figure on this site moves with it.
Calculation note A.1 — cost of one GPU hour
The same accelerator rents for about $5.50 an hour. Renting is not a small premium on a large number — it is the number, marked up. That margin is the one thing a better base model cannot take away from us.
Calculation note A.2 — Qwen3-Coder-480B-A35B at 2 GPU × 4,160 tok/s
Output price against published closed-model list
11× below frontier list for the same unit of output, on a model with comparable coding ability. The cheapest generative tier goes further — Devstral-Small lists at $0.58 per million. Closed list prices are market observations, not our figures; ours are derived above and move when the assumptions move.
Calculation note A.3 — why this board
A newer accelerator costs more per hour and less per token. It holds more memory, so a model needs fewer boards; and it serves faster, so those boards produce more. Both effects work against the higher capex, and they win.
| Board | Cooling | HBM | Board | Capex/GPU | $/GPU-hr | Replica | Throughput | List /M |
|---|---|---|---|---|---|---|---|---|
| H100 SXM5The prior generation. Cheaper per board, worse per token. | Air | 80 GB | 700 W | $31k | $2.21 | 4 GPU | 3.2k tok/s | $2.30 |
| IN SERVICEB200 SXMThe fleet standard. 192 GB holds most open models on fewer boards. | Direct liquid | 192 GB | 1000 W | $51k | $3.45 | 2 GPU | 4.2k tok/s | $1.38 |
| B300 Blackwell UltraMore HBM again — provisional pricing, reserved for the largest models. | Direct liquid | 288 GB | 1400 W | $61k | $4.09 | 2 GPU | 5.3k tok/s | $1.29 |
Reference workload is Qwen3-Coder-480B-A35B at the size it takes on 4 baseline boards. Blackwell pricing is negotiated per deal and moves quickly; the B300 row is provisional. Only the board changes between rows — life, utilisation, cost of capital, tariff and margin are held constant.
Equipment schedule — models in service
21 marksOpen weights, run on our own boards. Cost is what the calculation above produces; list is that cost at 3.0×. Every figure is derived — none of it is a price we picked.
| Mark | Model | Parameters | Licence | Context | Replica | Throughput | Cost | List | Per |
|---|---|---|---|---|---|---|---|---|---|
| H-01CodeSoftware engineering agents | |||||||||
| M-01 | Qwen3-Coder-480B-A35BDefault planner and editor for the code agent. | 480B / 35B act | Apache-2.0 | 262.144k | 2 GPU | 4.2k tok/s | $0.461 | $1.38 | 1M tok |
| M-02 | DeepSeek-V3.2Long-horizon reasoning; used for repository-wide refactors. | 685B / 37B act | MIT | 163.84k | 4 GPU | 7.0k tok/s | $0.546 | $1.64 | 1M tok |
| M-03 | Llama-4-MaverickWhole-codebase context without chunking. | 400B / 17B act | Llama 4 Community | 1,000k | 4 GPU | 7.9k tok/s | $0.483 | $1.45 | 1M tok |
| M-04 | Kimi-K2 (1T-A32B)Strongest open agentic tool-use; reserved for hard runs. | 1000B / 32B act | Modified MIT | 131.072k | 8 GPU | 6.8k tok/s | $1.13 | $3.40 | 1M tok |
| M-05 | GLM-4.6Balanced cost/quality workhorse. | 355B / 32B act | MIT | 200k | 4 GPU | 5.7k tok/s | $0.670 | $2.01 | 1M tok |
| M-06 | Qwen3-32BDense, latency-sensitive edits and completions. | 32B | Apache-2.0 | 131.072k | 1 GPU | 3.1k tok/s | $0.307 | $0.921 | 1M tok |
| M-07 | Devstral-SmallSingle-GPU tier for cheap loops and CI checks. | 24B | Apache-2.0 | 131.072k | 1 GPU | 4.9k tok/s | $0.194 | $0.582 | 1M tok |
| H-02Biological designProtein language models | |||||||||
| M-08 | ESM-C 6BRepresentation model — embeddings and variant effect scoring. | 6B | EvolutionaryScale Community | — | 1 GPU | 39.0k res/s | $0.025 | $0.074 | 1M res |
| M-09 | ESM3-open 1.4BGenerative PLM — sequence, structure and function conditioning. | 1.4B | EvolutionaryScale Community | — | 1 GPU | 10.4k res/s | $0.092 | $0.276 | 1M res |
| M-10 | ProGen2-xlargeFamily-conditioned de novo sequence generation. | 6.4B | Apache-2.0 | — | 1 GPU | 3.9k res/s | $0.246 | $0.737 | 1M res |
| M-11 | Boltz-2Co-folding and binding affinity in one pass. | 1.2B | MIT | — | 1 GPU | 0.13 struct/s | $7.37 | $22.11 | 1k struct |
| M-12 | RFdiffusionBackbone design for binders and scaffolds. | 60M | BSD-3-Clause | — | 1 GPU | 0.52 backbone/s | $1.84 | $5.53 | 1k backbone |
| H-03VisionVision-language models | |||||||||
| M-13 | Qwen3-VL-235B-A22BFrontier-class open VLM; documents, video, GUI grounding. | 235B / 22B act | Apache-2.0 | 262.144k | 4 GPU | 5.1k tok/s | $0.756 | $2.27 | 1M tok |
| M-14 | InternVL3-78BHigh-resolution tiling for dense diagrams and scans. | 78B | MIT | 32.768k | 2 GPU | 2.9k tok/s | $0.670 | $2.01 | 1M tok |
| M-15 | Molmo-72BPointing and spatial grounding with open training data. | 72B | Apache-2.0 | 32.768k | 2 GPU | 2.6k tok/s | $0.737 | $2.21 | 1M tok |
| M-16 | SAM 3Promptable segmentation and tracking. | 900M | SAM Licence | — | 1 GPU | 78 img/s | $0.012 | $0.037 | 1k img |
| M-17 | dots.ocrLayout-preserving document parsing to structured text. | 1.7B | MIT | — | 1 GPU | 6.5 page/s | $0.147 | $0.442 | 1k page |
| H-04KnowledgeRetrieval, embedding, synthesis | |||||||||
| M-18 | Qwen3-Embedding-8BMultilingual retrieval embeddings. | 8B | Apache-2.0 | 32.768k | 1 GPU | 249.6k tok/s | $0.004 | $0.012 | 1M tok |
| M-19 | BGE-M3Dense, sparse and multi-vector retrieval in one pass. | 570M | MIT | 8.192k | 1 GPU | 546.0k tok/s | $0.002 | $0.005 | 1M tok |
| M-20 | Qwen3-Reranker-4BCross-encoder rerank over retrieved candidates. | 4B | Apache-2.0 | 32.768k | 1 GPU | 166.4k tok/s | $0.006 | $0.017 | 1M tok |
| M-21 | Llama-4-ScoutTen-million-token context for whole-corpus synthesis. | 109B / 17B act | Llama 4 Community | 10,000k | 2 GPU | 4.4k tok/s | $0.433 | $1.30 | 1M tok |
21 models in service. Input billed at 25% of the listed output rate. Licences are as published by each model's authors — check them before commercial use. Prices are derived from the cost model in note A.1 and change when it changes.