Anthropic's S-1: What the IPO Filing Says About the Business Behind Claude

Until now, anyone trying to judge the economics of frontier AI had to rely on leaks, investor letters and run-rate figures nobody could check. On Monday, 26 October 2026, that changed for one of the big labs. Anthropic, PBC publicly filed a Form S-1 with the US Securities and Exchange Commission for an initial public offering of Class A common stock, to be listed on the Nasdaq Global Select Market under the symbol “ANT”. The company’s announcement says the number of shares and the price range “have not yet been determined”, and the registration is not yet effective, so nothing can be sold until the SEC finishes its review. Goldman Sachs and Morgan Stanley are named as the contacts for the preliminary prospectus.

This post goes through the prospectus itself (the filing on SEC EDGAR), checked against coverage by Fortune, the Wall Street Journal (via Yahoo Finance and Investing.com) and others. Where the press and the filing disagree, I go with the filing.

The numbers

Revenue, as stated in the S-1:

2023: $41 million. 2024: $386 million. 2025: $4.6 billion.
First half of 2026: $16.3 billion, up from $1.2 billion in the first half of 2025. The filing describes this as 1,300% year-over-year growth.
Q2 2026 alone: $11.6 billion, against $787 million a year earlier.

The filing then warns, in its own words: “We do not expect our revenue will continue to grow at comparable rates, and we cannot assure you that it will continue to grow at all.”

Profitability is where the headline number misleads:

Operating loss, H1 2026: $693 million, down from $2.1 billion a year earlier.
Operating income, Q2 2026: $1.0 billion, an 8% operating margin. A year earlier the margin was minus 152%.
Net loss, H1 2026: $80.7 billion. Of that, $72.4 billion is the “change in fair value of convertible notes, derivative liabilities, and SAFEs”. In other words, Anthropic’s valuation rose so much that what it owes holders of convertible instruments rose too, and accounting rules book that increase as a loss. It is a real liability, but no cash was spent and it says nothing about whether serving Claude makes money.

The expense table matters more for developers. In H1 2026, compute and infrastructure cost $8.9 billion, or 55% of revenue. In H1 2025 it was 201% of revenue. Sales, marketing and partnerships rose from $120 million to $3.2 billion, and personnel from $538 million to $3.2 billion. The 2025 figures also carry a one-off $1.5 billion “Bartz settlement”, the copyright case over training data.

Who pays

The filing says the number of customers paying $100,000 or more over a trailing twelve months grew from more than 950 to more than 11,900 between 30 September 2025 and 30 September 2026. No single customer accounted for more than 10% of revenue in the first nine months of 2026, but the top five together accounted for about 25%. The filing adds that “many of our top customers are not subject to long-term contracts” and can cut spending at any time. Anthropic sells mainly by usage (“pay-for-intelligence”, in the filing’s phrase), so revenue moves with how much customers actually use it. The prospectus gives one concrete example: in the two weeks to 2 January 2026, as engineering teams paused for holiday code freezes, total customer spend fell about 20%.

Claude Code has its own section. Anthropic calls it the product that “has most concretely defined what autonomous AI work looks like” and says it is “becoming the default choice inside many enterprise engineering organizations”. It describes adoption as bottom-up: individual engineers first, then teams, then the whole company. The filing gives no separate revenue figure for Claude Code, so claims in the press about its share of revenue are not something the prospectus supports.

The governance machinery

This is the most unusual part of the filing. After the IPO there will be several share classes:

Class A: the public shares, one vote each. Class N: economic shares with no vote. 5% of fully diluted shares in Class N go to a wholly owned Anthropic Public Benefit Fund.
Class F: held by a Founder LLC whose only members are, initially, the seven co-founders. It carries 50.1% of total voting power on general matters and has no economic rights. It ends through a 450-day transition period that starts once there have been fewer than three continuing founders for 90 days in a row.
Class T: held by the Long-Term Benefit Trust, a body of “financially disinterested trustees”. It elects the Class T directors, who make up a majority of the board.
Class E: issued to employees other than the co-founders, voting only on “certain limited matters”.

So public shareholders get economic exposure but little control. The founders control ordinary votes, and a safety-oriented trust controls the board majority. As a result Anthropic will be a Nasdaq “controlled company”. The prospectus states the trade-off directly: as a public benefit corporation, choices to “delay or restrict the release of models” for safety reasons may not maximise shareholder value. It gives an example that already happened: in April 2026 Mythos was released only as a gated preview through Project Glasswing, limited to cybersecurity organisations, because of its vulnerability-discovery capabilities.

Risk factors a developer should read

By one commentator’s count, about 80 pages of the document are risk factors. A few of them are notable for a company filing to go public:

Existential risk, stated in an SEC filing. Advanced AI could “pose catastrophic or existential risks to humanity”, and “our models are contributing to the pace of our own research”. The filing warns that those risks will grow as AI automates AI research and development.
The US government dispute. The filing describes the 2026 clash with the Department of War over limits on autonomous weapons and mass surveillance: a threat to invoke the Defense Production Act, the President directing federal agencies to stop using Anthropic’s technology, and the Department designating Anthropic a “supply-chain risk to national security”.
Insurance. No single insurer, and possibly not the insurance industry as a whole, may have the capacity to cover a catastrophic AI incident.
Revenue from abusers. Some reported revenue comes from customers later found to be violating the usage policy. That revenue disappears once Anthropic cuts them off.

The blank that matters most

The filing says Anthropic has entered into compute and lease commitments “with an aggregate committed spend of more than $ “, and the amount itself is left blank. Press figures vary widely. Citing the WSJ, Yahoo Finance reported more than $600 billion. Dealroom reported up to $750 billion, including about $111 billion with Google, $110 billion with AWS and $31 billion with Microsoft. An analysis syndicated on Yahoo Finance put it at $518 billion. None of these figures appears in the public S-1 as filed, so treat them as unconfirmed until an amended filing fills in the number. The filing does confirm the structure: take-or-pay contracts that oblige Anthropic to pay “regardless of actual utilization”.

Why it matters for IT

Three takeaways for people who build on these models:

Serving frontier models can make money. Compute fell from twice revenue to just over half of it in a year. That suggests inference prices are set above serving cost, not subsidised below it. That matters if you are betting a product on API prices staying where they are.
The lab’s costs are mostly fixed. Huge take-or-pay commitments against usage-based revenue explain the aggressive pricing, the push into agents (which use far more tokens) and the large partnership budgets. A slowdown in demand would hit a lab with fixed compute bills hard.
Release decisions are now disclosed governance. Gated launches like Glasswing are listed as a business risk, and a trust with a board majority oversees them. Developers should expect more staged and restricted releases, not fewer.

Open questions

The offering size, price and valuation are all missing. So are the run-rate figure as of 31 October and the total compute commitment. Reported valuation targets between $1.5 and $2 trillion come from press reports, not from the filing. The figures cover only to mid-2026 for full financial statements and to September for some metrics, and the company itself says the growth rate will not last. The governance design is also untested: no one yet knows how a trust-controlled board will respond when quarterly results and safety calls conflict in public.