The biggest AI scare of the week was not a model mishap — it was a spreadsheet. The Financial Times reported Thursday that OpenAI told prospective investors its annualized revenue run-rate was approaching $50 billion at the end of September, about $20 billion below the $70 billion figure that had been circulating among investors. AI infrastructure stocks slid on the news: Nvidia fell nearly 3%, Oracle shed close to 6%, and Arm dropped more than 6%.
The missing $20 billion was not lost, stolen, or hidden. It was never there. The gap came from two companies using two different methods to count money that flows through cloud partners — and investors who tried to make the figures comparable and made a compounding arithmetic error in the process.
Both OpenAI and Anthropic sell AI models directly and through cloud platforms such as Microsoft Azure, Amazon Web Services, and Google Cloud. When a sale passes through a partner, a company can book the gross amount the customer paid, or the net amount it keeps after the partner takes its share — a choice governed by the principal-versus-agent framework in U.S. accounting rules. Anthropic includes revenue from its cloud partners in its figures; OpenAI does not. The $70 billion estimate was built by investors who took Anthropic's accounting approach and applied it to OpenAI's revenue, incorrectly, to produce an apples-to-apples comparison.
Importantly, nothing in the report suggests OpenAI's business is deteriorating. In the same materials, OpenAI said its overall revenue run-rate grew 77% in the third quarter, with enterprise business run-rate growth reaching 107%. The newly disclosed figure is not a decline — it is a measurement clarified.
Still, the timing is sensitive. OpenAI confidentially filed for an IPO in June, with executives signaling a listing could come as early as 2027, and the company closed a record $122 billion funding round in March. Chief Financial Officer Sarah Friar said last week that the company remains "very well capitalized." But as both OpenAI and Anthropic march toward public markets, investors are getting their first lesson in how differently two frontier labs can count the same dollar — and how much those definitions matter at trillion-dollar valuations.