The Number With the Missing Label

He may have some deep-seated recondite complaint. I can’t say. (The Physician, on Mr Merdle)

OpenAI’s annualized revenue circulated for a while at around $70 billion, and then, in early October, at around $50 billion. Twenty billion dollars appeared to leave the building between bulletins. It had gone, as far as anyone can tell, into a footnote.

Reuters reported on October 8 that the gap arose mainly from an attempt to compare OpenAI with Anthropic, whose figures treat partner-channel sales differently. That account supports a correction to a league table. It offers no evidence that customers walked away with $20 billion of business. A financing critique can be serious and still lose the argument on arithmetic. The harder obligations survive the correction untouched.

Whose sale is it?

An AI model sold through another company’s cloud marketplace produces two honest numbers. One is the full amount the customer pays. The other is what the model’s maker keeps after the platform takes its share. Accounting rules decide which one a company may call revenue, based on who provides the service, and reasonable firms land in different places. (A bookshop and a publisher can both say they sold a novel; only one of them should put the cover price on its income statement.)

Matching the dates therefore settles nothing when one top line includes partner fees and the other excludes them. A league table built across that boundary ranks the way businesses present themselves alongside the businesses themselves, and the press, fond of a footrace, prints it anyway.

An annualized figure adds a second trap. It often takes a recent month and multiplies it by twelve. A shop that has a roaring December can annualize its way to a very pleasant forecast, provided nobody mentions February. For a business growing as fast as these, the trap also runs the other way: Ed Zitron’s reporting on OpenAI’s audited private statements puts its 2025 revenue at $13.07 billion, a figure that sits awkwardly beside a run-rate of about $50 billion nine months later until one remembers that a run-rate measures the pace and a year measures the distance. Both can be true. Neither says anything about expenses, which is where the drama dwells.

The loss that grows with good news

The largest numbers in circulation are mostly paper.

Reuters’s review of Anthropic’s confidential prospectus puts its 2025 net loss near $42 billion. About $34 billion of that is a noncash charge on financing instruments; the operating loss is a little over $8 billion. Zitron’s figures for OpenAI have the same shape: an operating loss of $20.92 billion, a fair-value charge of $41.55 billion, and a net loss attributable to OpenAI of $38.53 billion after part of the loss is allocated to other holders. For both laboratories, the headline loss runs well past the cost of running the business: roughly five times the operating loss at Anthropic, nearly double at OpenAI on its attributable basis.

The mechanism is one of accounting’s better jokes. The convertible financing instruments in these accounts are carried at their estimated value, and when investors decide the company is worth more, that value rises and the increase is booked as a loss. The better the market feels about the company, the worse its income statement looks. No cash leaves the bank in the process. Anyone subtracting one of these charges from a net figure to calculate a “real” loss, meanwhile, is mixing stages of the accounts that were never meant to be stirred together, and anyone presenting either headline as money burned has confused a valuation event with a fire.

None of this makes the operating losses small. Eight billion dollars and twenty-one billion dollars are large sums to lose while selling a popular product, and the cash-flow statements (which neither company has yet published) will say more about the runway than any loss figure can.

The quarter, adjusted

Fresh evidence cuts the other way. Reuters reported on October 14 that people at Anthropic’s investor meeting described third-quarter revenue of $16.8 billion and adjusted profitability. Those are management’s figures, relayed by attendees, and “adjusted” is a word that comes with its own footnote: it means profit after the company has decided which costs not to count this time. The adjustment schedule will show whether the excluded items are one-offs or the cost of doing business. Until it appears, the adjusted figure cannot be set beside anyone else’s operating earnings.

Still, a company can improve its current operations while carrying enormous future obligations, and growth makes those obligations easier to meet. It can also tempt the company into another round of expansion, which resets the clock. A strong quarter cannot settle a decade of debts, and a bad year cannot prove permanent failure. The useful question is narrower and less exciting: in the period when each cost falls due, what income will be there to meet it?

The roadshow has its own genre. Reuters’s account of the investor meeting carried the headline “‘Thinking’ caps, surging revenue, no human extinction,” which is the safety company’s version of a valuation floor: the risks that justify the mission filed under marketing. Sam Altman’s reported position that listing below a trillion dollars is a non-starter belongs to the same genre from the other direction. Both laboratories are asking the public to price a story about the future; the difference is which future each story protects.

Who signs for the sum

There is a legal distinction the circulating figures cross without ceremony. Antifraud rules apply to securities offers and sales even when the offering is exempt or confidential, so a false statement made in offering securities carries exposure before any prospectus is published. What a public registration statement adds is common ground: a document every reader can inspect, and, once it takes effect, a defined civil remedy under Section 11 of the Securities Act for material misstatements, reaching the issuer and other specified parties. A leak gives the reader less access to the evidence and less certainty about who made the representation. Before the filing goes public, the figures sit inside the law and are simply harder to test.

Ledger, entry two. Commitment: an honest account of revenue and losses. Made by: the issuer, in its prospectus. Exposed if it fails: investors who buy on the headline, and a public argument conducted in the wrong units. Who can call it in: regulators under the applicable antifraud rules; purchasers through available securities-law remedies, including Section 11 for material defects in an effective registration statement.

Until the prospectus is public, every figure in that table is somebody’s word, carried by somebody else.

Bibliography

Dickens, Charles, Little Dorrit (London: Bradbury and Evans, 1857), Book the First, chapter 21.

Reuters, ‘Anthropic’s IPO prospectus shows sweeping AI vision, surging costs’, 28 September 2026.

Singh, Jaspreet, and Krystal Hu, ‘OpenAI’s September annualized revenue nears $50 billion, less than previously indicated, source says’, Reuters, 8 October 2026.

US House of Representatives, ‘Securities Act of 1933, Section 11’, 15 U.S.C. § 77k, and ‘Section 17’, 15 U.S.C. § 77q.

US Securities and Exchange Commission, ‘Frequently Asked Questions About Exempt Offerings’, noting that antifraud provisions apply to all securities transactions, including exempt transactions.

Vinn, Milana, and Krystal Hu, ‘At Anthropic pre-IPO meet: “Thinking” caps, surging revenue, no human extinction’, Reuters, 14 October 2026.

Zitron, Ed, ‘Exclusive: OpenAI Losses Increased Nearly 8X in 2025, With Spending Hitting $34 Billion’, Where’s Your Ed At, 15 June 2026.