On 29 September, Axios reported that OpenAI's annual recurring revenue was approaching $70 billion. Within about a week, the Financial Times reported that OpenAI's September annualized revenue was closer to $50 billion, below the earlier estimates.
Same company. Same month. Twenty billion dollars apart. Neither outlet is careless, and nobody is lying.
That gap is worth more to you than either number, because it exposes something most AI coverage hides: "run rate" is a marketing unit, not an accounting one. Once you know what goes into it, the whole leaderboard of AI revenue figures, including the one we maintain on this site, reads differently, and so does every valuation built on top of it.
The numbers currently on the table
| Figure | Who reported it | What it refers to |
|---|---|---|
| ~$70B | Axios, 29 September 2026 | OpenAI annual recurring revenue |
| ~$50B | Financial Times, early October | OpenAI September annualized revenue |
| $65B | Anthropic to investors, reported 17 August | run rate at the end of July, roughly seven times a year earlier |
| $11.5B vs $6.7B | both companies, Q2 2026 | Anthropic's quarterly revenue against OpenAI's |
| ~$36B | OpenAI investor materials via the FT | revenue OpenAI expects to actually book in 2026 |
| $20B β $6B | same materials | OpenAI's run rate entering 2026, against 2024 |
Look at the last two rows next to the first. OpenAI expects to recognise around $36 billion of revenue across the whole of 2026, while its annualized run rate is quoted somewhere between $50 billion and $70 billion. Both can be true at once, and the reason why is the first thing to understand.
Why one company has two revenue numbers
Run rate is last month multiplied by twelve. If revenue is growing fast, the exit run rate massively exceeds the money that actually passed through the business during the year. A company that starts the year at $20 billion and ends it at $60 billion books something in the middle, not $60 billion. That alone explains most of the distance between $36 billion and the headline figures.
Different companies count partner sales differently. This is the part that makes cross-company comparison genuinely unsafe. Anthropic includes sales made through partners such as AWS and Google Cloud in its revenue, and pays those partners roughly 16 cents of every dollar earned that way. A Reuters analysis found such sales made up about half of Anthropic's revenue last year. OpenAI does not treat certain partner-generated sales the same way. Put a company that counts the gross channel next to one that does not, and the comparison is broken before you start.
None of it is audited. These are private companies. Every figure here comes from investor updates, internal materials or people briefed on them. There is no filing to check, no auditor's signature, and no obligation to define the term the same way twice.
| The number | What it actually tells you | How reliable |
| Annualized run rate | the size of the best recent month, times twelve | weakest; definition varies by company |
| Revenue booked in the year | what customers actually paid over twelve months | stronger, but private and unaudited |
| Quarterly revenue | the cleanest like-for-like between two companies | best available comparison |
| Cash flow | whether the business funds itself | the one that decides survival |
The comparison that does survive
Strip out the definitional noise and one pairing still means something: Anthropic's $11.5 billion of Q2 revenue against OpenAI's $6.7 billion. Same quarter, same unit, no annualizing. Even that carries the partner-channel caveat, but it is the closest thing to apples and apples anyone outside these companies has.
The more interesting divergence is further down the accounts. Anthropic posted an adjusted operating profit in Q2 and has told shareholders to expect another in Q3, the shift I wrote about when it became the first big lab to make money. OpenAI's own investor materials, as reported by the FT, project cumulative negative free cash flow of roughly $280 billion between 2026 and 2030, alongside revenue growing from about $36 billion this year to about $350 billion in 2030.
Those two sentences describe genuinely different companies, and no run-rate headline captures the difference.
Where the forecasts stop being facts
Be just as careful with the forward numbers, because they are where reporting gets loosest.
Anthropic's year-end run rate has been reported both as $100 billion (people who spoke to Reuters) and $120 billion (another account of the same trajectory). Both are projections of a number that is itself a projection, and the two sources disagree by the entire annual revenue of a large software company. OpenAI's $350 billion for 2030 is a plan presented to investors, not a forecast anyone is accountable for.
The honest way to hold all of it: the direction is real and extraordinary, the precision is not. A reader who remembers "Anthropic is somewhere around sixty-five billion and growing very fast" understands the situation better than one who can quote a figure to the decimal.
The same business, three headlines
It is easier to see with small numbers. Take one imaginary company that ends the year selling $5 million in December, having grown steadily all year, with 40% of sales going through a cloud marketplace that keeps 16 cents of every dollar.
| Headline you could honestly write | How you get there | Number |
| "Reaches $60M run rate" | December times twelve, gross | $60M |
| "Reaches $56M run rate" | the same, net of the channel fee | $56M |
| "Books $38M in revenue" | what customers actually paid across the year | $38M |
| "Grows 5x year over year" | January's $1M against December's $5M | 5x |
Four true statements, one business, and a spread of more than $20 million between the biggest and the smallest. Nobody has to lie. Somebody only has to choose.
Now scale that to companies selling tens of billions, add the fact that one of them puts the cloud channel in the number and the other does not, and the $50 billion against $70 billion gap stops looking like a contradiction and starts looking like arithmetic.
If you sell AI services, this is your pitch deck problem too
There is a smaller, more immediate version of this for anyone running an AI business, which is most of the people who read this site.
Clients and investors have been trained by these headlines to expect run-rate language, and the temptation to use the flattering version is strong. Your best month times twelve always sounds better than the money in the bank. The catch is that whoever you are pitching has also learned to ask the follow-up question, and the person who volunteers the definition before being asked looks like the one who understands their own business.
The practical rule for your own numbers is the one I would apply to OpenAI's: state the unit, state the month, state whether it is gross or net of platform fees. If you are putting together income claims of any kind, the same discipline runs through the nine ways people make money with AI, where every figure is a range tied to a model rather than a single flattering headline.
Five questions to ask of any AI revenue claim
This is the part worth keeping, because these headlines will keep coming.
Is it run rate or revenue? If run rate, it is one month times twelve, and the company chose the month.
Which month? A run rate from July quoted in October is three months stale in a market moving this fast.
Does it include partner channels? If the company sells through AWS, Azure or Google Cloud, ask whether those sales are in the number and at gross or net. A 16% channel fee is the difference between a good margin and an ordinary one.
Is there a profit line anywhere? Revenue growth with no path to cash generation is the thing that actually breaks, as the Gartner data on AI returns keeps showing at the customer end.
Who is the source, and what are they raising? Nearly every one of these figures surfaces while the company is raising money or preparing a listing. That does not make them false. It does tell you which direction the rounding goes.
What this means for you
If you might buy one of these IPOs, this is the single most important habit to build before the prospectus lands. Both companies are heading for public markets, with Anthropic reportedly positioning for a listing at an eye-watering valuation and OpenAI chasing a trillion-dollar debut. The moment they file, the numbers become audited and defined, and some of the figures circulating today will be restated. Decide now which metric you will value them on, because the first prospectus will offer you several.
If you build on these APIs, your supplier's cash position is your pricing risk. A company burning tens of billions a year has to find margin eventually, and the obvious place is the price of tokens. That cuts both ways right now, because competition has been pushing prices down hard, as it did when both labs cut frontier prices on the same day. Keep your stack portable and do not build a business whose unit economics only work at today's promotional rates.
If you invest in the theme more broadly, treat every private AI revenue figure as a directional signal rather than a data point. The aggregate picture, which I keep updated in the AI revenue leaderboard, is reliable about who is big and who is growing. It cannot be reliable about who is exactly how big, and anyone presenting it with false precision is selling something. That is the same discipline behind how AI companies actually get valued and the reason the bubble question stays unresolved.
What to watch
An IPO filing from either company. The first audited income statement in this sector will be the most informative document of the decade for AI finance, and it will settle definitions that are currently arguments. The listings pipeline is where to watch for it.
Whether anyone adopts a standard. If the labs start reporting revenue on a comparable basis, voluntarily or because bankers make them, the leaderboards suddenly mean something. Until then, treat every ranking, including ours, as a best effort over inconsistent inputs.
OpenAI's burn against its plan. A projected $280 billion of negative free cash flow through 2030 is the real story under the revenue headlines. Whether the actual figures track that plan, run ahead of it or fall behind will tell you more than any run-rate update.
Anthropic's third quarter. A second consecutive adjusted operating profit would turn one quarter into a trend, and a trend is what a public market will pay for.
The honest take
What the $50 billion versus $70 billion gap really shows is that the most quoted numbers in the most capitalised industry on Earth are produced by the companies themselves, in units they define, released when it suits them. That is normal for private companies. It stops being normal when those figures underpin valuations in the hundreds of billions and shape how ordinary people think about where the economy is going.
None of this means the growth is fake. Anthropic going from roughly $9 billion to $65 billion in a year is extraordinary by any definition you choose, and OpenAI's business at $36 billion of actual booked revenue would still be one of the fastest scaling companies in history. The caution is narrower and more useful than scepticism: know which unit you are being handed, and notice that the person handing it to you picked it.
So before the next headline number lands, decide what you would need to see to believe it. If the answer is "an audited statement", then everything before the IPO filing is weather, not climate.
Frequently asked questions
How much money does OpenAI actually make?
It depends which unit you want. Around $36 billion of revenue is expected to be booked across 2026, per investor materials reported by the FT. Its annualized run rate during September was reported between roughly $50 billion and $70 billion depending on the source and the method. All of these are unaudited.
Is Anthropic bigger than OpenAI now?
On quarterly revenue in Q2 2026, yes: $11.5 billion against $6.7 billion. On run rate the comparison is unsafe, because Anthropic counts sales through cloud partners such as AWS and Google Cloud, which made up about half its revenue last year, and OpenAI does not treat certain partner sales the same way.
Is OpenAI profitable?
No. Its own investor materials project cumulative negative free cash flow of roughly $280 billion between 2026 and 2030, while revenue is projected to climb to about $350 billion by the end of that period.
Is Anthropic profitable?
It reported an adjusted operating profit in Q2 2026 and has told shareholders to expect another in Q3. "Adjusted" is doing real work in that phrase, and the figures are unaudited, but it is further than any other frontier lab has got.
Why do different sites list different AI revenue numbers?
Because run rate has no standard definition, companies pick which month to annualize, some count gross channel sales and some do not, and none of the figures are audited. Any ranking, including ours, is a best effort over inconsistent inputs.
This article is general information, not financial advice. Do your own research and consider a licensed professional before making investment decisions.
Sources: Axios: OpenAI's ARR nears $70B; Analytics Insight on the FT's revised September figure; CNBC: Anthropic's run rate climbed to $65B in July; Techstrong on OpenAI's projected cash burn.
