Same company, same year, revenue figures 63% apart

TL;DR. OpenAI's booked full-year 2025 revenue was reported at $13.1 billion, against a year-end annualised run rate of $21.4 billion. Same company, same year, 63% apart, and both figures are correct, because run rate annualises the final month of a growing year while booked revenue counts what actually arrived. That gap is the smallest problem here. These are private companies, so there is no filing, no auditor and no segment reporting. Independent trackers covering the same period disagree by multiples: one placed a lab at $6.5 to $7.5 billion in mid-2026 while another recorded a company disclosure of $47 billion weeks earlier. Nobody outside these firms can state what AI revenue is, and the hyperscalers do not break it out either, with one reporting AI revenue inside Cloud and Workspace and disclosing no product-specific figure at all.

---

Status: contested, and deliberately unresolved. This article does not adjudicate between conflicting figures, because the sources are not comparable. All figures are attributed to their reporting basis. It is descriptive and is not investment advice. The companies discussed include the maker of the model used in drafting parts of this site, and no figure here is presented as favouring any of them.

---

The 63% gap, which is the easy part

OpenAI's booked revenue for calendar 2025 was reported at $13.1 billion. Its annualised run rate at the end of 2025 was $21.4 billion.

Both are accurate and they measure different things.

Booked revenue is what was recognised across twelve months. Run rate takes the most recent month and multiplies by twelve. In a company growing quickly, the second is always larger, and the faster the growth the larger the gap.

So a 63% difference is not a discrepancy. It is the arithmetic of growth, and quoting either without saying which is being used makes the figure uninterpretable.

This is the well-behaved case, where one company's own disclosures separate the two and the difference is explicable.

The part that does not resolve

These are private companies. No 10-K. No audited statements. No segment reporting. No obligation to disclose anything, and no standard definition when they choose to.

The result is visible in the trackers.

One serious dataset records a company disclosure putting a lab's run-rate revenue at $14 billion in February 2026, and another company disclosure putting it at $47 billion in May 2026. Those are company statements attached to funding announcements, which is the strongest available basis.

A different analysis, published in June 2026, placed the same lab at $6.5 to $7.5 billion in mid-2026, describing its figures as compiled from media reporting, revenue leak disclosures, management commentary on partner calls, and triangulation from hyperscaler segment disclosures.

Those cannot both be right, and this article cannot tell you which is. The definitions differ, the bases differ, and neither is auditable from outside.

What can be said is that a range from $6.5 billion to $47 billion for one company in one quarter is not a measurement of anything. It is several different quantities wearing the same label.

Why the definitions do not line up

Four distinct choices, each defensible, each producing a different number.

Run rate against booked revenue. Established above, worth 63% in one documented case.

Full company against product line. A figure for "the company" and a figure for one product are both quoted as revenue, and coverage rarely says which.

Gross against net of revenue share. Where one firm pays another a share of revenue, or buys compute from an investor, the same dollar can appear in more than one company's figure or in neither, depending on treatment. One reconstruction of a single year's payment between two firms found two reported values differing by more than $6 billion, with the difference between compute credits, research spending and cloud charges not reconciled anywhere public.

And point-in-time against period. A run rate quoted with a date is a snapshot. Quoted without one, it silently becomes an annual figure.

The hyperscalers do not resolve it either

The obvious response is to look at the public companies, which do file.

They do not break out AI revenue. One reports AI revenue inside its Cloud and Workspace segments and discloses no product-specific figure. Others reference AI contribution in commentary without a defined, auditable line.

Which means the comparison everyone wants to make cannot be made. Capital expenditure is disclosed and enormous, with 2026 guidance across four firms summing to roughly $700 billion. AI revenue is not disclosed in a form that can be set against it.

Analysts construct the comparison anyway by triangulating from cloud segment growth, which is a defensible method producing an estimate, and the estimate then circulates as though it were a disclosure.

That is citation decay with an unusually short chain: the estimate is often one hop from the reader, and the hop is still invisible.

Three things this establishes

Run rate is not revenue and the difference is large. A documented 63% gap in one company in one year, and the gap widens with growth. Any figure quoted without its basis is uninterpretable, and most are quoted without it.

Private-company figures have no verification layer. Nothing here suggests any company has misstated anything. The point is structural: a company disclosure attached to a funding announcement is the strongest available evidence, and it is still an unaudited statement with a self-selected definition.

And the capex-to-revenue comparison cannot currently be made honestly. One side is filed and audited; the other is estimated, undefined and inconsistent between sources. Presenting a ratio built from those two as a finding attaches the credibility of the first number to the second.

What it does not establish

That any figure is wrong. Sources conflict because they measure different things on different bases at different dates, which is the normal result of no standard existing.

That the businesses are or are not viable. This article makes no claim in either direction, because the inputs required to make one are not public.

That disclosure is being withheld improperly. Private companies have no obligation to disclose, and voluntary disclosure at funding events is more than the law requires.

And nothing about any valuation. Nothing here evaluates any company or security.

What is unresolved

Whether any standard emerges. Nothing requires a definition of AI revenue, and the firms have no incentive to adopt one that constrains them.

Whether hyperscalers begin breaking it out. Segment reporting follows how management runs the business, so a breakout would signal AI being managed as a distinct unit rather than an input to existing products.

How much revenue is circular. Where a chip vendor holds equity in customers, or a cloud provider invests in a company that buys its compute, the same dollar can be counted more than once across the ecosystem. The magnitude of that is not publicly quantifiable.

And what the trackers should do. The best of them, such as the one recording dated company disclosures with source links and confidence flags, are doing the honest thing available. They still cannot make incomparable figures comparable.

The counter-argument

Demanding audited figures from private companies is a category error. They are private precisely so they need not report, and the alternative to voluntary disclosure at funding events is silence. Criticising the quality of numbers nobody is obliged to publish sets a standard that would produce less information rather than better information.

Run rate is the right metric for a fast-growing company. Booked revenue for a year in which a company tripled understates its position at year end by construction. Investors use run rate because it answers the question they have, and calling it misleading assumes a reader who wanted the other number.

The conflicting figures may be less conflicting than presented. A source reporting $6.5 to $7.5 billion may be estimating recognised quarterly revenue while another reports annualised run rate, in which case the gap is definitional rather than substantive. This article treats an unreconciled difference as a finding, and some of it will reconcile.

And the capex comparison may be legitimate even if imprecise. Directionally, capital expenditure across the sector clearly exceeds attributable revenue by a wide margin, and refusing to state that because the second figure is unaudited is its own kind of evasion.

The short version

OpenAI's booked 2025 revenue was reported at $13.1 billion against a year-end run rate of $21.4 billion. Same company, same year, 63% apart, both correct, because run rate annualises the final month of a growing year.

That is the well-behaved case. These are private companies with no filing, no auditor and no segment reporting, and independent trackers covering the same period differ by multiples: one records dated company disclosures of $14 billion in February 2026 and $47 billion in May 2026 for a lab that another analysis places at $6.5 to $7.5 billion in mid-2026, compiled from media reports, leaks and triangulation.

A range from $6.5 billion to $47 billion for one company in one quarter is not a measurement. It is several quantities sharing a label, separated by four defensible choices: run rate against booked, full company against product line, gross against net of revenue share, and point-in-time against period.

The public companies do not settle it either, since AI revenue is reported inside existing segments with no product-specific line, which means the comparison everyone wants, capital expenditure against AI revenue, cannot currently be made honestly. One side is filed and audited. The other is estimated and inconsistent. Building a ratio from both lends the credibility of the first to the second.

None of which means any figure is wrong. It means no standard exists, nobody is obliged to create one, and the numbers in circulation are doing more work than their basis supports.

Common questions

What is the difference between run rate and revenue? Booked revenue is what was recognised over a full period, usually twelve months. Annualised run rate takes the most recent month or quarter and multiplies it out. In a fast-growing company the second is always larger, and the faster the growth the larger the gap. One documented case shows booked 2025 revenue of $13.1 billion against a year-end run rate of $21.4 billion for the same company: 63% apart, both accurate, measuring different things.

Why do different sources give such different figures? Because four defensible choices each change the number, and coverage rarely states which was made. Run rate against booked revenue; the full company against a single product line; gross revenue against revenue net of any share paid to a partner; and a point-in-time snapshot against a period total. A source reporting one basis and a source reporting another will differ by multiples while both being internally consistent.

Can these figures be verified? Not from outside. The leading AI labs are private companies with no obligation to file audited statements, no segment reporting, and no standard definition when they do disclose. The strongest available evidence is a company statement attached to a funding announcement, which is unaudited and uses a self-selected basis. That is more than the law requires and less than verification.

Do the public hyperscalers break out AI revenue? No. AI revenue is reported inside existing segments, with one company disclosing it within Cloud and Workspace and giving no product-specific figure. Analysts construct estimates by triangulating from segment growth, which is a reasonable method that produces an estimate, and the estimate then circulates as though it were a disclosure.

So can capital expenditure be compared to AI revenue? Not honestly at present. Capex is filed and audited, with 2026 guidance across four firms summing to roughly $700 billion. AI revenue is estimated, undefined and inconsistent between sources. Building a ratio from those two lends the credibility of the audited figure to the estimated one. The directional statement that sector capital expenditure exceeds attributable revenue by a wide margin is defensible; a specific ratio is not.

Does this mean the numbers are being manipulated? No, and nothing here suggests any company has misstated anything. The problem is structural rather than behavioural: no standard exists, private companies are not obliged to create one, and voluntary disclosure with a self-selected basis is the best available evidence. Sources conflict because they measure different things, not because someone is lying.

Is run rate a bad metric? Not inherently, and this is the strongest objection to the article's framing. For a company that tripled during a year, booked revenue understates its position at year end by construction, and run rate answers the question an investor actually has. It becomes misleading only when quoted without its basis or its date, which is how it usually travels.

What would improve this? A stated basis attached to every figure, which costs nothing and would resolve most apparent conflicts. Beyond that, segment breakout by the public companies would allow the comparison people keep attempting, though segment reporting follows how management runs a business, so a breakout would itself signal that AI is being managed as a distinct unit rather than as an input to existing products.