SHEET 02

Cost derivation

All figures derived, none quoted

The price is low for one structural reason: we own the accelerators instead of renting them, at $1.53 a GPU hour against $2.49 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 02.1 — cost of one GPU hour

Accelerator — H100 SXM5$21k

Host share (CPU, RAM, NVMe, chassis)$5k

Fabric share (switch, optics, cable)$2k

Fit-out share (power, cooling, rack)$2k

Capital employed per GPU$31k

Depreciation — 5 yr straight line$0.977 /h

Capital charge — 11% on average balance$0.269 /h

Energy — 0.89 kW × PUE 1.18 × $0.058/kWh$0.061 /h

Operations — staff, transit, spares$0.222 /h

Cost per GPU hour, owned$1.528 /h

The same accelerator rents for about $2.49 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 02.2 — Qwen3-Coder-480B-A35B at 4 GPU × 3,200 tok/s

Replica cost$6.11 /h

Output produced11.5 M tok /h

Cost to serve$0.531 /M tok

List at 2.0× cost$1.06 /M tok

Output price against published closed-model list

14× 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 $0.45 per million. Closed list prices are market observations, not our figures; ours are derived above and move when the assumptions move.

Rev	Issue	Halls	GPU
A	Pod 01 energised, code hall only	1	384
B	Vision and knowledge halls in service	3	1,152
C	Biological design hall in service	4	1,536

Cost derivation

Issue and revision

Bring a workload. We will show you what it costs here, and what it runs on, before you move it.