Glossary: the terms behind AI compute economics

Six terms carry most of the disagreement about AI economics. None of them are difficult; all of them are routinely used loosely, usually by people who have never read a capacity contract.

Gigawatt-year

One gigawatt sustained for one year: 8,760 GWh. The unit that makes rent comparable to build cost, and the one that disappears whenever a slide says "$110 billion per gigawatt".

Take-or-pay

A clause obliging the buyer to pay for capacity whether or not it uses it. In a market where the supplier finances a five-year asset, this is what makes the debt possible — and what moves utilization risk onto the buyer.

Spot premium

The gap between on-demand and committed rates for identical silicon. At the top of the market it reaches four to six times the long-contract rate, which tells you what urgency costs.

Utilization

The share of installed capacity doing paid work. Every revenue-per-gigawatt figure depends on it, and almost none of them state it.

Die yield

The percentage of started dies that become working chips. It multiplies every die on the line, which makes it the only variable that scales a fab's entire output at once.

Nominal capacity

What lines can release if yields and allocations hold. Deliverable chips run below it, and the gap is where forecasts go wrong.