The AI economy’s most repeated revenue figure was wrong by definition, and the difference is 26%.
OpenAI told investors it closed September with about 50 billion dollars of annualized revenue. The 68 billion circulating at the end of September was gross revenue, including partners’. The difference is not growth nor loss: it is an accounting criterion.
The source and the criterion
The correction came from CNBC, with a person familiar with the matter, after the Financial Times broke the news first. The numbers came from an investor presentation.
And the rest of the package: 77% run-rate growth in the third quarter, and 107% in the enterprise business. The company is under pressure to justify an 852-billion-dollar valuation, with a confidential prospectus filed in June and a targeted stock debut in 2027. In the same package, the Anthropic comparison: 65 billion annualized at the end of July, 2025 revenue of 4.6 billion, and a 42-billion net loss, per Reuters, citing a leaked prospectus copy.
The market repriced the premise
The reaction came the same day: Nvidia down 3%, Oracle nearly 6%, CoreWeave nearly 8%, AMD 4%, Broadcom 4%, Intel 5%, and Super Micro nearly 5%.
The entire AI chain repriced because the revenue premise changed 26%. The market lesson is the year’s most expensive: a premise is the most expensive asset to have wrong. The whole chain, from chip maker to data center vendor, was priced on a number whose criterion nobody had checked.
Gross revenue is not revenue
It is the FinOps lesson that sinks payrolls when confused: billing and recognized revenue are different numbers, and the difference is not a detail when the valuation is 852 billion.
The same criterion holds at any company’s scale: gross AI spend (subscriptions, API, licenses, partners) is not the spend sustaining production. Per-completed-task measurement is what separates billing from value, at your operation’s size. And the same criterion rigor missing in AI’s economy is missing in the internal reports nobody audits.
Conclusion
Two numbers, 26% apart, and the whole explanation fits in one sentence: gross revenue is not revenue.
For anyone deciding with AI numbers (invoice, report, pitch), the permanent question is the case’s: gross or net, under which criterion, measured by whom? The answer separates decision from headline at a billion-dollar lab’s scale and at a monthly spreadsheet’s scale. Accounting criterion is the cheapest FinOps asset to own and the most expensive to lack.