How much revenue does one gigawatt of AI compute earn?

• $110Bquoted per gigawatt-year, traced to a rent estimate with its conditions removed
• $10–15Bwhat labs actually realize per gigawatt-year, cross-checked two ways
• $11.6Bwhat the owner needs to earn for a 15% return, by Goldman Sachs's arithmetic

Somewhere in a board deck near you sits a line of arithmetic that has done more to justify AI capital expenditure than any earnings report. One gigawatt of AI compute, it reads, one hundred and ten billion dollars of annual recurring revenue. Since a one-gigawatt build runs to fifty or sixty billion, the building appears to repay itself in six months, which is the kind of number that ends a budget meeting.

It is wrong by roughly a factor of three. The interesting part is not the correction — any spreadsheet can do that — but what the error reveals: a rent figure has been circulating in a revenue costume for two years, underwriting decisions worth hundreds of billions.

Where the number came from

The traceable origin is an estimate by SemiAnalysis that OpenAI and Anthropic can earn more than $100 billion a year from a gigawatt of inference sold through their APIs. The sentence carried three qualifications, each of which fell off on the way to the conference stage: it is gross lab revenue before the cost of training a single model, so the building's owner never sees it; it assumes frontier API pricing, the richest rates in the market; and it describes a ceiling case, a cluster running flat out on paid inference. Slides hold conclusions, not conditions.

Six prices for the same gigawatt

Google publishes the only honest rate card in this market. Ironwood, its current flagship TPU, lists at $12 per chip-hour on demand, $8.40 on a one-year commitment and $5.40 on three years — identical silicon, a 55% discount for time. At the other end, Nebius reports about $40 billion per gigawatt-year on short contracts, and the SpaceX–Anthropic lease annualizes to roughly $50 billion.

From watts to revenue

A gigawatt measures power; an income statement counts tokens sold, at what price, with what share of the machine doing paid work. Operators plan around 65–75% schedulable utilization, then the workload mix takes its cut: 20–30% of a frontier fleet runs training, which earns nothing directly. Multiply what remains by the price of a token and JPMorgan's framework lands at $20–40 billion a year, against about $10 billion realized in 2025.

The cross-check the company ran on itself

In April 2026 Anthropic announced a run-rate above $30 billion, up from roughly $9 billion at the end of 2025. Assign all of it to inference, which flatters the result, and divide by the compute the company could actually serve: the answer lands between $10 and $15 billion per gigawatt, seven to twenty times below the number in circulation.

The company's own filings agree with the smaller figure. A prospectus Reuters reviewed in September lists $518 billion of commitments over a decade, about 80% of them non-cancellable or payable regardless of usage, against 2025 revenue of $4.6 billion. If a gigawatt earned $110 billion, a company holding eleven of them would treat those covenants as a rounding error, and the lawyers would not have spent the pages.

The ten-minute check

Before accepting any per-gigawatt revenue claim, three questions separate the defensible from the decorative: what fraction of the power runs paid inference, at what realized load, and what a token sells for after discounts. The public versions of those questions are cheaper to answer than most decks assume — build cost, rent by contract length, and the buyer's own run-rate.

This page is the public data companion to Two Tides Issue 006. The complete argument, including the contract structure and the 2027–28 oversupply case, is in the letter.