Price the whole AI factory, not the GPU.
Turn owned utilization, throughput, PUE, power, electricity, and capex assumptions into a comparable cost per million tokens. Every number stays inspectable.
Owned scenario CSV
Replace the fictional sample with your own assumptions. Percent utilization is entered as a decimal; all costs are annualized USD.
Ready for owned assumptions.
Scenario evidence
Computed from the same parsed rows shown above. No vendor or market data is fetched.
| Scenario | Util. | PUE | Annual tokens | Energy MWh | Energy cost | Annual cost | Cost / 1M |
|---|
Arithmetic and units
Annual tokens = GPUs × utilization × tokens per second per GPU × 31,536,000. Facility energy = GPUs × accelerator watts × utilization × PUE ÷ 1,000 × 8,760. Cost per million = (energy cost + annualized capex) ÷ annual tokens × 1,000,000.
What the sensitivity proves
Improving utilization can increase productive token output faster than facility energy grows, while PUE reduces overhead. Neither variable should be interpreted alone: throughput, electricity price, accelerator power, and capital cost remain in the denominator and numerator.
Boundaries
This is a fictional sample and an owned-input calculator, not a forecast. It makes no live claim about IREN, Horizon, Microsoft, NVIDIA, delivery timing, contracts, performance, security prices, returns, or investment suitability.