I Traced 24 AI Revenue Figures. Eight Came From the Companies. Zero Came From a Filing.

TL;DR: I traced 24 of the most-quoted OpenAI and Anthropic revenue figures back to their origin on August 15, 2026. Eight are on-the-record company statements. Sixteen are anonymous-source media reports. Zero come from a filed document, because neither company files one. The two headline numbers in circulation right now do not have equal standing: Anthropic's $47 billion run rate is published by Anthropic on its own site, while OpenAI's $40 billion run rate is Bloomberg reporting sourced to people who spoke on condition of anonymity, and OpenAI declined to comment. Anyone putting those two numbers side by side is comparing a company statement to a leak.

Every AI market analysis you have read this year rests on run-rate figures for two private companies. I wanted to know how many of those figures anyone outside the companies has ever been able to check.

What did I count?

Three things, all on August 15, 2026.

First, provenance. Epoch AI's revenue tracker tags every datapoint it plots with a source type and a source URL, which makes the audit tractable. I parsed the table out of the raw HTML rather than trusting the rendered chart, added the two headline figures that post-date it, and classified each as a filed document, an on-the-record company statement, or an anonymous-source media report.

Second, the filings. I ran EDGAR full-text search for the literal phrase "AI revenue" across 10-Q filings from January 1 to August 15, 2026, and read every hit.

Third, attribution. I pulled 22 pages ranking for OpenAI's run rate and checked whether each tells the reader where its number came from.

Which of the numbers came from the companies?

Epoch's series holds 22 datapoints. Seven are tagged "Company disclosure." Fifteen are tagged "Media report." The split is lopsided by company: 3 of OpenAI's 13 datapoints are company disclosures, against 4 of Anthropic's 9. Epoch says so itself, in a line most people quoting the chart skip: "Our revenue figures come primarily from media reports, which may have timing or accuracy uncertainties."

The tracked series stops early. Epoch's CSV is stamped February 26, 2026, and its last entry is Anthropic's $14.0 billion (company disclosure, anthropic.com) on February 12, 2026. Both live headline numbers came after it, and this is where the audit corrected my own assumption. I expected both to be untraceable. Only one is.

Anthropic's $47 billion run rate is a direct company statement, published in its Series H announcement on May 28, 2026: "our run-rate revenue crossed $47 billion earlier this month." Same page, same standing: $65 billion raised at a $965 billion post-money valuation, both stated by Anthropic. That is not a leak. I read it on Anthropic's own domain.

OpenAI's $40 billion run rate is an anonymous-source media report. Bloomberg published it on August 13, 2026, sourced to people familiar with the company's performance who spoke on condition of anonymity, and OpenAI declined to comment. The supporting detail is an internal Greg Brockman memo, also relayed anonymously. OpenAI's last on-the-record figure remains the roughly $20 billion annualized run rate stated publicly by CFO Sarah Friar and Sam Altman for the end of 2025.

So: 24 figures, 8 confirmed by the company that earned the money, 16 not, and zero filed documents on either side.

Why is a run rate not revenue?

Because nothing in the phrase "annualized run rate" is defined by anyone. It is not GAAP. It is generally one recent month or quarter multiplied up, chosen by the company, unaudited, with no requirement to disclose which month or what it includes.

This matters more than provenance for the comparison everyone wants to make. Bloomberg itself cautioned that the two companies may calculate the figure using different methodologies, a caveat that survived into the syndicated versions I read. Gross versus net treatment is the obvious candidate: whether revenue routed through cloud resellers is booked at full value or net of the partner's cut can move a number a long way, and neither company publishes its policy. That is enough uncertainty that the $47 billion company statement and the $40 billion anonymous report cannot be ranked against each other, and I am not going to pretend otherwise. It is the same problem I hit when I classified the 12 studies everyone cites on AI coding productivity: the measurements are not measuring the same thing.

Nor can you fall back on the IPO paperwork. Both have filed confidentially, and a confidential draft S-1 is by construction a document nobody outside can read. Reported valuations from employee tender offers are not public marks either. Several pages I read carry an $852 billion OpenAI valuation with no source attached, which I could not trace, so I am not repeating it as fact.

What do the filings actually say?

Here is the gap between what companies say publicly and what they will sign. In 10-Q filings from January 1 to August 15, 2026, EDGAR full-text search returns 3,404 documents containing "artificial intelligence" and 10 containing "AI revenue". A ratio of about 340 to 1. Companies will describe themselves as AI companies all day. Attaching a revenue number to it under signature is rarer by two orders of magnitude, and the numbers they do sign turn on estimates they can move, starting with how long they say a server lasts.

I read all 10. They are not what you would guess:

• 2 are not about AI revenue at all. Mid-America Apartment Communities and Camden Property Trust both attached RealPage antitrust settlement notices, where "AI Revenue Management" is the name of a RealPage product.
• 1 is a pay package, not a disclosure. IBEX Ltd filed an employment agreement amendment vesting 5,000 performance share units at each of $5m, $10m, $15m and $20m in AI Revenue. The company will stake equity on the milestone before it will report the line.
• 1 is a cost line. CuriosityStream cites the timing of "AI revenue share expenses," which is money going out.
• The rest are small. Safe Pro Group booked $934,062 in quarterly Safe Pro AI segment revenue. AIRWA, Rackspace, Wiley and the SpaceX filing mention AI revenue only in accounting-policy or forward-looking language.

Neither OpenAI nor Anthropic appears in either count. They are private. They file nothing. The companies buying their compute do file, which is why the only checkable arithmetic in this whole story sits in the hyperscalers' cash flow statements.

Can you check any of this from outside?

Only sideways, and only for direction. Ramp's August 2026 AI Index reports that in July, 43.5% of U.S. businesses paid for Anthropic subscriptions or tokens against 39.7% for OpenAI, which at least does not depend on either company's self-reporting. But it is not a revenue number and cannot be converted into one. Ramp measures card charges among its own cardholders, a self-selected population, and Ramp's methodology note on this month's model-usage chart says the sample "skews slightly more tech-y than our typical AI Index sample." Adoption breadth is not spend, and spend is not recognized revenue.

Does anyone repeating these numbers say where they came from?

I fetched 22 pages ranking for OpenAI's run rate. 19 resolved (Bloomberg and Investing.com returned 403 to me, Gradually.ai returned 429). All 19 state a dollar figure. 12 name Bloomberg as the origin. Only 8 disclose that the sourcing was anonymous. Seven name no source at all.

The split is clean. Every news article named Bloomberg. All seven evergreen "OpenAI statistics 2026" aggregator pages named nobody, and six of those seven do not carry the $40 billion figure at all, still quoting numbers between $20 billion and $25 billion. Those are the pages that rank permanently, get scraped, and feed the next model's training data. This is how a leak becomes a fact: not by anyone lying, but by the sourcing falling off in transit, the same decay I traced through the AI SDR churn stat.

What does this not prove?

It does not prove any of these numbers are wrong. That is not the finding. Private companies have no obligation to disclose anything, media reporting on private financials is legitimate and frequently accurate, and anonymous sourcing exists because employees who talk to reporters get fired. The Information, Bloomberg, CNBC and the FT do real work here, and the alternative to their reporting is not better numbers, it is no numbers.

My sample is bounded too. Epoch's tagging is Epoch's judgment, not mine, the attribution audit is one query on one day, and 3 of those 22 pages I could not open.

The finding is narrower and harder to dismiss. Sixteen of 24 headline figures cannot be checked by anyone outside the companies, no filing exists to check them against, and the ecosystem quotes them with the confidence it would give an audited 10-K. When you next read that one lab has passed another, check which number came from a company and which came from a leak. Right now they are not the same kind of number, and the scoreboard is one nobody outside can audit.