AI inference glossary
Benchmark metrics

Modeled profit per gigawatt

Also known as profit per GW, annual modeled profit per utility GW

In plain English

This estimate subtracts modeled serving costs and applicable model-license fees from annual token revenue for a gigawatt of utility power.

Technical definition

Modeled profit per gigawatt is annual token revenue less the costs and license fees included in the model, normalized to an all-in utility gigawatt.

Typical unit

modeled USD profit per utility GW per year

Engineering details

The result inherits the revenue model’s interactivity, token prices, cache mix, and utilization assumptions. Compute expense uses the selected owning or rental cost basis. If a license fee is specified as a percentage of revenue, it is calculated on revenue rather than on the amount left after compute costs.

Why it matters

This is a defined economic estimate, not audited corporate net income. Costs outside the model can change realized profit, and weak demand or falling token prices can reduce earnings without changing the benchmark performance of the hardware.

How to read it in InferenceX

The Rubin article’s 75 TPS example assumes 60% utilization and no model-license fee for MIT-licensed DeepSeek V4 Pro. Its linear conversion from a GW to a smaller deployment holds the same assumptions; it is not a measured fleet-scale profit result.