The Anthropic Filing: An $80 Billion Loss, And Seven Founders Who Cannot Be Outvoted

Anthropic's public S-1 shows 1,300 per cent revenue growth, an $80 billion loss that is mostly an accounting mark, an $800 billion commitment to computing, and a single share that keeps its founders in control.

The filing is the subject of this article: what the first full set of financial statements from a frontier AI laboratory actually contains, which of its numbers are cash and which are arithmetic, and what the structure of the company says about who will run it after the shares start trading.

1. What Was Filed

Anthropic filed its public Form S-1 with the Securities and Exchange Commission on Monday 26 October 2026, becoming the first of the large AI laboratories to reach a public filing. The company intends to list Class A stock on the Nasdaq Global Select Market under the ticker ANT. Goldman Sachs holds the lead advisory role, with Morgan Stanley alongside. The document runs to roughly 240,000 words.

Three things the coverage assumed are not in it. There is no valuation: the filing leaves the price range and the share count blank, so the reports of $1.5 trillion, $1.75 trillion and $2 trillion are expectations circulating around the deal rather than numbers the company has committed to. There are no results after 30 June, because the cut-off is the half year, which means the quarter that ended on 30 September is absent from the document that will be used to price it.

The third omission is time. The filing commits to a listing after the midterm elections of 3 November and before the Thanksgiving holiday, with trading possible from 16 November. Everything between now and then is a roadshow, a price range and a decision about what the largest new issue in American history is worth.

2. The Loss That Is Not Spending

The headline is a net loss of more than $80 billion for the first six months of 2026. It is the number that travelled around the world within minutes, and it is close to meaningless as a measure of how the business traded. More than $72 billion of that figure is a change in the fair value of convertible notes, derivative liabilities and simple agreements for future equity. That is an accounting mark, not a payment.

Just under $9 billion of the loss came from spending on computing power and cloud infrastructure. That is the part to treat as real cost, and it is roughly a ninth of the total. The remaining nearly nine tenths is the company telling its accountants what its own paper is now worth, which is a strange thing to report as a loss and a perfectly ordinary consequence of raising money through instruments that convert at ever higher prices.

The mechanism is worth following because it will recur. Anthropic sold convertible notes and future equity agreements at earlier, lower valuations. When the company's valuation rises, those instruments are worth more, and the increase has to be recognised. A rising valuation therefore produces a rising reported loss. The $80 billion is not evidence that the company is failing. It is evidence that the company's price went up.

3. Growth, And What It Is Made Of

The revenue line is the one that matters. Anthropic booked $16.3 billion in the first half of 2026, growth of 1,300 per cent on the same period a year earlier. Run that backwards and the first half of 2025 produced about $1.2 billion. A business that was worth roughly a billion a year in early 2025 was worth more than thirty billion annualised eighteen months later.

The composition is unusually concentrated. Anthropic reports more than 14,000 enterprise customers, of which about 250 generate $10 million or more in annual revenue each. Those 250 accounts alone are worth at least $2.5 billion a year, and probably considerably more given that $10 million is a floor rather than an average. The rest is spread across thousands of companies using the API and a developer tool that became the default for coding work.

The filing concedes the risk that follows from this. Much of the revenue depends on a limited number of products and a limited number of customers, and the largest of those customers are also the companies that funded the business. That is not unique to Anthropic, but it is the difference between a fast growing software company and one whose growth is partly bought by its own investors.

4. The Eight Hundred Billion Dollar Bet

The number that deserves more attention than the loss is the commitment to spend close to $800 billion on computing capacity over the next decade. Against $16.3 billion of revenue in six months, or about $32.6 billion annualised, the commitment is roughly 25 times current annual revenue and 49 times the last half year. It is the largest forward purchase of computing ever attached to a company of this age.

That commitment is the reason the loss exists at all. Compute is bought ahead of the revenue that justifies it, the revenue arrives on a lag, and the gap is financed by investors who accept paper today for capacity tomorrow. Gartner expects global spending on AI services, software and infrastructure to reach roughly $3.6 trillion in 2027, up from $1.8 trillion in 2025. Anthropic is betting a measurable fraction of that total will route through its models.

The bet is also the part of the filing that is hardest to price. A commitment to buy capacity is not the same as a contract to buy it, and the difference between an intention and an obligation is worth hundreds of billions. Investors will want to know which one this is. The filing describes the plan; the pricing documents will have to describe the terms.

5. Seven Founders, One Share

The governance is the most unusual thing in the document, and it is not modelled closely on any large technology company that came before it. Immediately before the offering, Anthropic will issue a single share of Class F stock to Anthropic Founders, LLC, whose members are the seven co-founders. That one share carries 50.1 per cent of total voting power and, in the filing's own words, will not have any economic rights.

The founders will hold roughly 10 per cent of the economics and control of the company. Class A voting power is capped at 49.9 per cent, and Class F power is not diluted by future issuance, so the arrangement cannot be eroded by raising more money. In practice the founders decide the election of the non-trust directors, certain changes to the charter, and any merger or sale. A public company in the legal sense is a controlled company in every other sense.

The board is split along a second line. The Long-Term Benefit Trust holds one Class T share and elects four of the seven directors, while Class F and Class A together elect the other three. A founder leaves the LLC only by falling below 2 per cent of their IPO date holdings, which means someone could sell 98 per cent of their stake and keep their vote. Amazon and Alphabet, which carry the largest economic exposure outside the founders, hold non-voting Class N stock.

6. Why The Loss Grows When The Company Succeeds

Anthropic was last valued at $965 billion after raising $65 billion from investors in May 2026. Every instrument issued before that round was struck at a lower price, and every mark applied after it moved. Alphabet recorded a $99 billion gain on its stake and Amazon booked $53 billion of pre-tax income in the second quarter largely because of the same revaluation. The loss on Anthropic's income statement and the gains on its investors' statements are two entries describing one event.

For anyone pricing the shares, that event is the only clean denominator in the document. At a $1.75 trillion valuation, annualised first half revenue of $32.6 billion implies a multiple of about 54 times. At $2 trillion it is 61 times. At the last private mark of $965 billion it is 30 times. Whatever number the banks print in November will sit somewhere in that range, and each step changes the story from expensive to extraordinary.

The comparison that matters is not with software companies trading at twenty times revenue. It is with what the market is willing to pay for a business growing at triple digit rates with a nine figure loss that is mostly paper. Anthropic has given investors the first regulated set of accounts from a frontier laboratory, and those accounts are simultaneously the best and the least flattering in the sector.

7. What The Filing Concedes

The risk section runs to 93 pages, against roughly 30 for Snap and 50 for Uber at their filings. Anthropic groups the danger into four categories: misuse, including the design of biological weapons and cyber attacks; loss of human control; authoritarian abuse, where the technology becomes infrastructure for surveillance states; and economic disruption. It warns that liability from extreme harm could exceed what its insurance covers, and that its models may take harmful actions that no user or third party directed.

The commercial risks are more conventional and more immediate. Compute supply may not keep pace with ambition. Regulation is uncertain in every major market. Revenue is concentrated in a small number of products and customers. And the company has to reconcile a founding mission about safety with the disclosure obligations of a public issuer, a tension the filing acknowledges rather than resolves.

One absence is worth noting. OpenAI, the closest competitor by reputation, has reportedly pushed its own listing into next year. That leaves Anthropic as the sector's test case rather than one of several, which raises the stakes for everyone holding AI exposure through other names.

8. The Pricing Test

Three things will decide how this is read. The first is the price range, which arrives in November and converts the reported valuations into an offer. The second is the September quarter, which the filing omits and which the banks will have to describe. The third is the language around the $800 billion commitment: whether it is a contract with penalties or an ambition with a schedule determines whether the balance sheet carries an obligation or a plan.

The wider consequence is not about Anthropic. The first public multiple for frontier AI becomes the reference for every private round, every data centre financing and every fund with artificial intelligence exposure. If ANT prices at 54 times revenue and trades up, the private market has a new floor. If it prices and breaks, the same market has a new question about what the last two years of valuations were based on.

The $80 billion loss will be forgotten within a week, because it is a mark that moves with the share price. The multiple will not be forgotten, because it is the number that every other company in the sector will be measured against. That is the sentence to remember from a 240,000 word document filed on a Monday in October.

Tags

#ai #business