Executive Summary

Key Figures

Four numbers. The first two are the OpenAI figure as corrected and the distance it fell from the number that had been circulating. The third is Anthropic's figure, the one most often set beside it, counted a different way. The last shows how quickly the confusion turned into a price.

Sources: TechCrunch (2026-10-08), SiliconANGLE (2026-10-08).

On October 8 the Financial Times reported that OpenAI's annualized revenue is approaching $50 billion. It is not a figure the company published. It sits in material circulated to investors. OpenAI is private and files no quarterly statements, so numbers used from outside to size its revenue generally arrive by that route.

The trouble was the figure the market had been holding until then. From September 29 several outlets reported that OpenAI's annualized revenue had come close to $70 billion, and for the better part of ten days that number served as a baseline. A roundup at Yahoo Finance put it at $68 billion and traced it back to reports at Axios and Reuters. Either way it is about $20 billion larger than the figure that ran on October 8.

The market moved the same day. The Nasdaq fell 1.25%, its worst day since the middle of August, and the S&P 500 fell 0.5%. Individual names fell further. Nvidia lost 2.9%, Oracle 5.5% and CoreWeave 8%, while Intel fell 5.3% and AMD and Broadcom each gave up about 4%. These are the names that have been priced on equipment and chips with OpenAI's revenue standing in for future demand.

The Financial Times also wrote where the figure came from. In the passage TechCrunch carried, the $70 billion arose from "attempts by OpenAI's own investors to produce a direct comparison with Anthropic's annualised revenues." It was a restatement made on the investor side rather than a number the company released.

The restatement had two ingredients. One was the baseline of about $40 billion known in August, the other the growth OpenAI had disclosed. Third-quarter revenue run rate was reported up 77%, and the enterprise segment up 107%. Put growth in the seventies on top of $40 billion and the result lands between $68 billion and $71 billion, which matches what was circulating.

The multiplication itself was sound. The error was in the units of the two things being multiplied. The $40 billion counted only OpenAI's own share, and investors applied Anthropic's method as they extended it by the growth rate. One outlet described the step as applying Anthropic's methodology when extrapolating the growth rate, in order to make OpenAI's net figure comparable with Anthropic's gross one. It is measuring a length in centimetres and then adding the increase in inches.

Dates make the two rulers easier to tell apart. The table below sets out when each figure appeared and from what position.

Read the table down the page and one company's revenue appears to have jumped from $40 billion to $70 billion in two months and then dropped to $50 billion. Read it across and a different picture comes out. The $40 billion and the $50 billion are two points measured with one ruler, and the $70 billion is a second ruler that entered between them.

The difference in what the two companies count reduces to one sentence. Suppose a customer uses a model through a cloud provider and pays one dollar. Anthropic records that whole dollar as its revenue and takes the provider's portion out separately as cost of revenue. OpenAI records only the portion that reaches it. The first is called a gross basis and the second a net basis, and neither term refers to gross margin, which is a separate ratio.

The route is not hypothetical. Anthropic's models are sold through AWS and Google Cloud, and OpenAI's through Microsoft Azure. This $20 billion gap opened over which basis is used to count the amounts written on that route.

Neither treatment is a dressing-up of the books. Whether a company records the full amount or only a commission turns on whether it acts as principal or as agent in the transaction, and the accounting standards leave room for that judgment. Two companies in the same line of business can land on different sides of it if their contracts are built differently. But once figures produced under two different judgments sit in one column, the difference in the size of the business and the difference in how it is written down get mixed inside a single number.

▲ Original Pebblous diagram. Counting method from TechCrunch (2026-10-08); Anthropic figure from SiliconANGLE (2026-10-08).

How much of the partner dollar reaches OpenAI cannot be known from outside. One outlet carried an analysis estimating roughly 20 cents, but the company has not confirmed that figure and it has wide room to vary by contract. Not knowing the ratio is the point. Without it there is no way to convert a figure measured on one basis onto the other.

So the comparison still does not hold after the correction. OpenAI's $50 billion sits on a net basis, and the $65 billion Anthropic was reported to have at the end of July sits on a gross basis. Subtracting one from the other to say Anthropic leads by $15 billion is equally wrong before and after October 8. The correction landed on the OpenAI figure, not on the act of setting the two side by side.

Annualized revenue is not a metric the accounting standards define. It is usually built by taking the most recent month or quarter of revenue and multiplying by twelve or by four. Which items go into that month's revenue, whether one-off contracts are excluded, which date the exchange rate is fixed at: the company decides all of it. The name leaves wide room for different formulas to sit underneath it.

Listed companies write this down in their reports. The passage where Upland Software defines ARR in its fiscal 2025 annual report is typical. Upland counts only recurring revenue of the subscription kind, so it is not the same metric as the OpenAI and Anthropic figures, but it belongs to the same family in that recent performance is multiplied up to a year. The filing gives the formula as monthly recurring revenue as of December 31 multiplied by twelve, and then adds this.

"ARR does not have any standardized meaning and may not be comparable to similarly titled measures presented by other companies."

— Upland Software, Inc., Form 10-K for fiscal year 2025

The sentence is a disclosure, not a rhetorical flourish. A company using a measure from outside the accounting standards in a filing to the U.S. Securities and Exchange Commission has to give the definition and state its limits. Dozens of 10-K filings across different companies carry the same wording. The absence of a standard has itself become something to be disclosed.

But the side that has to attach that warning is the listed company carrying a disclosure obligation. OpenAI and Anthropic are both private, and there is no filing for such a sentence to sit in. Their figures come out through investor material and news reports with the number travelling and the definition left behind. Unless the receiving side asks what was counted, there is nowhere for the question to be put.

Conversations about data quality usually start from the values. Whether a cell is missing, whether a row is duplicated, whether a format is off. In this case nothing was missing and nothing was duplicated. The $50 billion and the $65 billion are each exact on their own definitions. And still a $20 billion misreading appeared the moment one of them was converted onto the other's method so they could be compared directly. The place that failed was not the values but the layer above them, where what a value counts is written down.

That layer always gets pushed back. It is invisible, and when it is wrong nothing shows immediately. A numerical error surfaces when a total fails to add up or a chart spikes, while a definition mismatch rides all the way into the conclusion with the numbers looking fine. This one was caught in ten days because the Financial Times looked at that spot, and far more numbers never get a reporter.

Inside a company's own metric dashboard the same thing happens more quietly and more often. Whether an active user is anyone who logged in once or someone who used a core feature, whether the denominator of a conversion rate is visitors or sign-ups, whether model accuracy is an average over everything or a figure with the hardest slice taken out. Teams settle these differently and then put them on one screen side by side, and whoever is looking takes them for values measured the same way. The screen is built that way.

So three questions follow any pair of numbers, outside or inside, before they are set next to each other.

• Where is this metric's definition written down? If it lives only in someone's memory or in a where clause inside a query, it is not written down.
• Do the two numbers being compared use the same definition? Sharing a name is not evidence. As here, a shared name is often what produces the comparison in the first place.
• Who finds out when the definition changes? Swapping a denominator is one line of code, and a report that compared the last quarter without knowing about that line keeps circulating.

All three have to be answered before the data is gathered. Upland can write the definition and the limits into its report partly because of the disclosure obligation, but also because it fixed in a document how the metric would be calculated in the first place. A definition is not recoverable after the fact. It survives only where someone decided to write it that way from the start.

Thank you for reading this far. The revenue figures and the account of counting method in this article come from the TechCrunch report and the Investing.com summary, and the disclosure wording for the non-standard metric was checked separately in Upland Software's annual report. We would be glad to hear where your team keeps the definitions of its core metrics.

Official documents

• 1.Upland Software, Inc. (2026). "Form 10-K, fiscal year 2025." U.S. Securities and Exchange Commission. — Primary source for the disclosure quoted in section 4. The filing defines the metric as monthly recurring revenue as of December 31 multiplied by twelve, then states that it has no standardized meaning and may not be comparable to similarly titled measures at other companies.

Industry and press

• 2.Wiggers, K. (2026). "OpenAI's revenue is reportedly $20 billion less than previously projected." TechCrunch, 2026-10-08. — Coverage of the Financial Times report. Source for annualized revenue approaching $50 billion, for the $70 billion arising from attempts by OpenAI's own investors to produce a direct comparison with Anthropic's annualised revenues, and for the counting difference in which Anthropic includes sales made by its cloud partners and OpenAI does not.
• 3.Investing.com. (2026). "OpenAI annualized revenue at $50bn, far below reports - FT." 2026-10-08. — Source for the August baseline of about $40 billion, for whether revenue routed through partners such as AWS and Google Cloud is included, and for growth above 70% over the period.
• 4.Deutscher, M. (2026). "AI stocks crumble on report that OpenAI's annualized revenue is much lower than previously believed." SiliconANGLE, 2026-10-08. — Source for the share price moves on the day (Nasdaq 1.25%, S&P 500 0.5%, Nvidia 2.9%, Oracle 5.5%, CoreWeave 8%, Intel 5.3%, AMD and Broadcom about 4%), for third-quarter run rate growth of 77% and enterprise growth of 107%, and for Anthropic's figure of more than $65 billion at the end of July.
• 5.Yahoo Finance. (2026). "OpenAI revenue appears $20 billion lower than reported." 2026-10-08. — Roundup that puts the earlier circulating figure at $68 billion and traces the September 29 reporting back to Axios and Reuters. Also the source for the gap being a measurement difference rather than a demand shortfall.
• 6.TechTimes. (2026). "Investors built $70B OpenAI revenue estimate using wrong method." 2026-10-09. — Source for the calculation path in which Anthropic's methodology was applied when extrapolating the growth rate from the $40 billion baseline. The 20-cent estimate introduced in section 3 as a single analysis also comes from this outlet and has not been confirmed by the company.