OpenAI Told Investors About $50 Billion. The Press Had Printed $70 Billion. The Gap Is a Definitions Problem — and a Market Mover.

On 8 October 2026 the Financial Times reported that OpenAI told investors its annualized revenue was approaching $50 billion at the end of September. Press reports built on earlier investor material had put the figure at about $70 billion. A $20 billion difference between two numbers attached to the same company looks like a collapse. Axios, which had itself carried the higher figure in September, says it is closer to a definitions problem, and that nothing in the reporting amounts to a confirmed decline. That is a plausible reading. It is also one that needs testing, because the arithmetic of the reported explanation is stricter than it sounds.

The three numbers and who reported them

There are three dated figures. First, roughly $70 billion, which Axios says was widely reported in September and came from information OpenAI shared with investors, framed to allow a more direct comparison with Anthropic. Charles Schwab’s 9 October market note cites the previously reported figure as $68 billion instead; the sources disagree by about $2 billion, and we use ‘about $70 billion’ for the circulated number. Second, approaching $50 billion at the end of September, which the FT says OpenAI told investors, as relayed by AI Weekly and Axios. We could not read the FT article itself. Third, Bloomberg reported on 9 October that OpenAI expects to reach $70 billion in annualized revenue by year-end, per Schwab. The sources we read do not say whether OpenAI itself has commented on the discrepancy.

Note what is and is not a revenue fall. Axios cautions that the gap should not be read as a confirmed $20 billion decline. Annualized revenue is a projection, a recent period of sales scaled up to twelve months, and public-market investors generally prefer revenue actually reported. Nothing here is audited.

Why ‘annualized revenue’ is not one number

Three conventions can move the headline. The first is the period being annualized: a figure built from the latest month times twelve rises faster, and falls faster, than one built from a quarter times four or a trailing year. The second is product mix, since consumer subscriptions, API usage and enterprise contracts behave differently and some are contracted while others are usage-based. The third, and the one the reporting points to, is gross versus net treatment of sales made through cloud partners.

Axios describes the difference this way. Both OpenAI and Anthropic comply with US accounting rules, but they treat cloud-partner sales differently. Anthropic counts revenue from such sales in its total on a gross basis: if a customer pays $100 through a cloud provider, Anthropic can record the full $100 as revenue and book the provider’s cut as an expense. OpenAI, by contrast, records only its own share of certain partner sales. Accounting professor Francine McKenna, quoted by Axios, says the right treatment depends on each company’s role in the transaction, including who controls the customer relationship and who delivers the product. Axios adds, citing unnamed sources, that the $70 billion figure was an effort to gross up OpenAI’s revenue to match Anthropic’s method. On that account the $50 billion is OpenAI’s own basis, and the $70 billion was a translation, not an error.

Testing the explanation with arithmetic

Here is a check the coverage did not make. Suppose a share of OpenAI’s revenue passes through cloud partners, and the partner keeps a fraction c of what the customer pays. Under net accounting OpenAI records what it keeps; under gross accounting it would record the full amount, so the recorded figure rises by a factor of 1 divided by (1 minus c). The gap equals the net partner revenue times c divided by (1 minus c). All numbers in this paragraph are hypothetical and chosen only to illustrate the structure.

For the gap to be $20 billion, the net partner revenue multiplied by that factor must be $20 billion. Even if every dollar of the $50 billion were partner-routed, c would have to be about 29% (20 divided by 70). If only half of the $50 billion were partner-routed, c would have to be about 44%. A $100 sale where the partner keeps $30 produces a gross-up of about 43% on the partner-routed portion, so a 30% cut applied to everything would almost close the gap on its own, but only if essentially all revenue flowed through partners. That is hard to square with a business that also sells subscriptions directly and sells API access directly. The implication, which is ours and not the sources’, is that cloud-partner gross-up may account for part of the $20 billion but that other differences, such as the annualization period, mix, or what counts as partner revenue, are probably in play too. We would want OpenAI or an audited document to say.

Why a definitions gap moved a market

Schwab’s 9 October note says the Nasdaq fell 1.3% on Thursday 8 October, its worst session since mid-August, as scrutiny of OpenAI’s revenue hit AI-linked stocks. CoreWeave fell nearly 8%, Arm more than 6%, Intel more than 5% and Sandisk about 5%. Bloomberg’s year-end $70 billion expectation, reported the next day, eased some of the concern. We repeat the causal link as Schwab’s reading. Other things were happening: Apple was reported to have told suppliers to cut production of some parts, which Schwab says weighed on that stock on the Friday, the 10-year Treasury yield was 5.26%, and Schwab notes that investors are focused on AI-related debt, with strong results no longer enough on their own to lift AI-linked shares. A revenue-definition story landing on top of that is the kind of catalyst that a nervous market amplifies.

The mechanism is that valuations of AI infrastructure companies lean on forecasts of customer spending, and OpenAI is a very large customer or counterparty for several of them. A lower revenue base for the largest buyer reduces the credibility of the demand curve behind every supplier, even if the gap is purely definitional, because investors cannot quickly tell which parts of the demand story rest on which definition.

The counter-argument: a misunderstanding that changes nothing

The strongest counter-view is that this is a labelling dispute. AI Weekly relays a Bluesky post, amplified by the journalist Mary Branscombe, calling it a simple misunderstanding that does not change the trajectory. Bloomberg’s report that OpenAI expects $70 billion annualized by year-end points the same way: on that basis the company would reach the higher number on its own accounting within a quarter. If the growth path holds, the gap would close without anyone restating anything. The posts are social-media commentary, not analysis, and the Bloomberg expectation is a company projection, so neither settles the matter. They do show why the sceptical reading and the relaxed reading can both be defended on today’s evidence.

What to ask before comparing AI labs

• Which period is being annualized, and is it contracted or usage-based revenue?
• Is revenue sold through cloud partners recorded gross or net, and what is the partner’s share?
• Does the figure include revenue from affiliates or investors who are also customers?
• Is the number audited, or a management figure shared in an investor presentation?
• Are the two companies being compared on the same basis? Axios says the $70 billion was an attempt to put them on one basis, which also means any comparison with Anthropic depends on which method is used.

The same discipline applies to profit claims. Axios notes the Financial Times reported that Anthropic told investors it made an operating profit in the second quarter, with figures that excluded stock-based compensation, and advises readers to be sceptical until formal financial documents appear. Our earlier analysis of Anthropic’s S-1 commitments makes the same point about reading unaudited management numbers against filed ones.

What the evidence does not establish

We have not read the FT article or any OpenAI statement, and we have not seen the investor documents behind either number. The 8 October FT figure, the Axios accounting explanation and the Schwab market data come from three different outlets, and the size of the circulated figure is itself given as $68 billion in one source and $70 billion in others. We cannot say how much of the $20 billion is accounting, how much is timing, and how much is real. This analysis is current as of 11 October 2026, and CPI data due on 14 October and any formal disclosures could change the picture.