OpenAI's Revenue Was A Third Smaller Than Reported, And The Whole AI Chain Fell

In this storyOpenAINVIDIA

Shares in Nvidia, Oracle, CoreWeave and other artificial intelligence companies fell on Thursday after OpenAI told investors its annualised revenue run rate was about $50 billion at the end of September, below the $68 billion figure widely reported late last month.

Nvidia fell 3 per cent, Oracle 6 per cent and CoreWeave 8 per cent in intraday trading. Advanced Micro Devices and Broadcom each fell 5 per cent, with Intel and Super Micro Computer down 6 per cent.

The gap is a measurement difference rather than lost sales. A person familiar with the matter said the $68 billion figure included gross revenue from OpenAI's partners, and that the $50 billion number allows a more direct comparison with Anthropic.

OpenAI had told investors late last month that its annualised run rate had risen more than 70 per cent since the third quarter began, with enterprise sales more than doubling since July. Reports at the time put the figure near $70 billion. The company has since indicated growth of 77 per cent in the overall run rate during the quarter and 107 per cent in its corporate business.

OpenAI recorded a net loss of $38.5 billion in 2025 on revenue of $13.07 billion.

The company is under pressure to justify a valuation of around $852 billion. It filed a prospectus confidentially with regulators in June and executives have signalled a listing in 2027. It is in early-stage discussions about a new funding round that could raise roughly $30 billion, with no term sheet finalised, after closing a $122 billion round in March. Chief financial officer Sarah Friar said last week the company remains well capitalised.

No customer left, no contract was cancelled and no forecast was cut. What happened is that a number was measured differently, and roughly $20 billion of apparent revenue turned out to belong to other companies.

That it moved Nvidia, Oracle, CoreWeave, AMD, Broadcom, Intel and Super Micro in a single session says something about how this market is valued. These are suppliers priced on the assumption that demand from a small number of AI labs will keep compounding. When the clearest available measure of that demand is revised down by a third on a definitional point, the whole chain reprices, because the chain was never priced on its own order books.

The distinction being corrected is itself worth understanding. Gross revenue including partners counts money that passes through a business; net revenue counts what it keeps. The difference matters enormously for a company whose partners include its own cloud suppliers, and the comparison being drawn was with Anthropic, which had presumably been quoting on the narrower basis.

Against that, the underlying business is not in question. Seventy-seven per cent growth in a quarter and a corporate line more than doubling are the figures of a company compounding fast. A $38.5 billion loss on $13.07 billion of 2025 revenue is the figure of one spending far ahead of it.

Both can be true. The $852 billion valuation requires the first to continue for long enough to outrun the second, and this week the market got a slightly clearer view of the starting line.

The reason a definitional correction moved semiconductor stocks is that nobody has a better number to use.

Nvidia, Broadcom, AMD, Oracle and CoreWeave all sell into AI infrastructure demand, and the largest single source of that demand is a private company that discloses what it chooses, when it chooses. There are no quarterly filings, no audited statements and no obligation to reconcile figures that circulate in the press. Investors are pricing the supply chain off numbers they cannot verify.

That is the actual finding here, and it is more interesting than the revenue itself. A $50 billion run rate growing 77 per cent in a quarter is a remarkable business by any standard. The problem is that the market learned the size of it from an unnamed person correcting a figure that had been reported for a fortnight.

The listing would change that. A public OpenAI reporting quarterly would give the entire AI supply chain a disclosed demand signal, which is why the 2027 timetable matters well beyond OpenAI's own shareholders. DeepSeek is targeting early 2027 too, and whichever lists first creates the first audited reference point for the category.

For India the read-through is narrower but real. Indian data centre operators, including the Nxtra and Neysa projects that raised a combined $1.6 billion this year, are building capacity on demand assumptions that ultimately trace back to the same handful of American labs. When the demand signal from those labs turns out to have been overstated by a third because of a definitional choice, every downstream capacity plan deserves re-examination.

For daily, sharp analysis of the biggest moves in the Indian business and startup ecosystem, follow StartupFox.