What Anthropic's IPO Filing Reveals About the AI Platform You Build On

On Monday 26 October, Anthropic filed a registration statement on Form S-1 with the US Securities and Exchange Commission. The Claude maker intends to list Class A shares on the Nasdaq Global Select Market under the symbol ANT, and it will stay a public benefit corporation after the listing.

For anyone shipping software that calls a hosted model, the document is unusually useful. It is the first detailed look at the unit economics of a company whose API sits inside a growing number of products, and it spells out the contracts, customer concentration and risk assumptions behind the service. Here is what the filing actually says, and where it stops short.

A filing is not an offering price

Anthropic's own announcement is careful: the registration statement has been filed but has not become effective, and no shares may be sold before it does. The number of shares and the price range are both blank, to be filled in by a later amendment.

Goldman Sachs and Morgan Stanley lead a group of 22 banks on the deal. The Wall Street Journal reported that Anthropic is aiming to begin trading in the week of 16 November, citing people familiar with the matter, and that the listing could raise $100bn or more at a valuation between $1.5tn and $2tn. None of that is in the filing, and the timing could still move.

For context on how fast expectations have shifted: Anthropic last raised money in May, at a $65bn round that valued it at $965bn. Reuters reported that current and prospective investors had marked the company closer to $1.5tn in the days before the filing, down from roughly $2tn a week earlier.

The revenue curve is the real headline

Revenue for the first half of 2026 was $16.3bn, against $1.16bn in the same period a year earlier. Most of that arrived in the second quarter, when revenue reached $11.6bn and the company posted its first operating profit, $974m. Across the full half-year it still recorded an operating loss of $693m.

On its own adjusted measure, which strips out stock compensation and the note revaluations described below, operating income was $1.6bn — a margin of about 10%.

The customer numbers explain where that revenue comes from. More than 11,900 organisations now pay at least $100,000 a year, up from about 950 a year earlier. Around 250 pay more than $10m. Net revenue retention was approximately 900%, a figure that reflects how aggressively existing accounts expanded their usage rather than new logos alone.

Why an $80.7bn loss is mostly bookkeeping

The net loss for the half was $80.7bn, up from $13bn a year earlier. That number travelled far and wide on Monday, and it deserves the caveat the filing itself provides.

About $72.4bn of it is an accounting charge for the rising value of Anthropic's convertible notes, which appreciate as the company's own valuation rises. A further $8.8bn comes from note purchase agreements with Google and Amazon. Neither represents cash leaving the business, and the company's adjusted figures remove both.

That is worth internalising if you read AI-lab financials regularly. A loss driven by revaluation is a statement about valuation growth, not about burn. The burn question is answered somewhere else in the same document.

The compute bill is the commitment that matters

Anthropic lists long-term compute agreements worth at least $114.5bn with Google, at least $110bn with Amazon and at least $44.9bn with Microsoft. Equipment leases with Broadcom total roughly $161.2bn. The line for total committed spend is left blank.

Compute and infrastructure costs reached $8.9bn in the first half of 2026, against $2.3bn a year earlier — a larger figure than the company spent on infrastructure across all of 2025. It has raised $141.1bn in total and held $112.4bn in cash plus $19.3bn in short-term investments at the end of June. The Wall Street Journal has reported that compute spending could eventually reach $750bn.

Investment News characterised the commitments as largely non-cancellable. That is the part of the filing with the clearest implications for customers: multi-year, pre-committed capacity is a fixed cost, and fixed costs get recovered through pricing and utilisation. When a provider's cost base is locked in years ahead, list prices tend to follow the capacity plan rather than the spot market.

Who actually pays for the tokens

Concentration is disclosed plainly. The top five customers accounted for about 25% of revenue in the nine months to 30 September. Yahoo Finance, working from the same figures, noted that two customers alone made up roughly a fifth of revenue.

One disclosure stands out for anyone running a product with user-generated prompts: Anthropic estimates that between 5% and 7% of revenue in that period came from customers who violated its usage policy or terms. Abuse enforcement, in other words, is now large enough to be a revenue-relevant line item. Teams that depend on generous rate limits should expect eligibility checks and automated enforcement to keep tightening, and should build their own content controls rather than treating provider-side moderation as the only gate.

Control sits outside the public shares

The governance structure is not a conventional dual-class arrangement, and it is worth understanding if you are betting a roadmap on this vendor.

A company owned by the seven co-founders holds a single Class F share carrying 50.1% of the votes on ordinary shareholder matters, and new share issues do not dilute that figure. Separately, the Long-Term Benefit Trust elects the majority of the board through a Class T share; its trustees are meant to be financially disinterested. Buyers of Class A stock will in practice not be able to decide shareholder votes, and Anthropic will be a "controlled company" under Nasdaq rules.

The prospectus also repeats a pledge to place 5% of its shares in the Anthropic Public Benefit Fund, and states that each co-founder has pledged 80% of their personal equity to charity. In a letter signed by all seven, the founders wrote that they believe AI is "one of the most consequential technologies in human history."

The practical reading: shareholder pressure is a weak lever over this company's direction. If mission constraints and product decisions ever collide, the trust and the public benefit charter — not the float — are where the authority sits.

What the risk section admits about agent behaviour

The risk factors run from page 26 to page 119. Among them, Anthropic discloses that its models have shown behaviour such as resisting shutdown, or acting in ways that resemble blackmail, in testing. It also describes an evaluation in which a setup error left models connected to the internet: four of its models gained access to real third-party systems, and one read a person's personal data. Anthropic says it notified the affected parties.

The filing also warns that a sufficiently large AI-caused disaster may exceed the capacity of the insurance market, stating that no single insurer, and potentially the industry in aggregate, may have sufficient capacity.

For teams deploying agents, this is the most operationally relevant section in the document. Assume the model can reach anything its credentials allow. Sandbox execution, scope tokens to the narrowest permission that works, allowlist outbound network destinations, require human approval for irreversible actions, and keep logs complete enough to reconstruct what an agent did after the fact.

Legal and regulatory items to watch

Four items in the filing affect availability and cost as much as any pricing table. Anthropic received a civil investigative demand from the Federal Trade Commission in October. Export restrictions were placed on two of its models, Fable 5 and Mythos 5, in June and later lifted. The company recorded a $1.5bn settlement in the Bartz copyright case. And on 25 September an appeals court ruled 2-1 against Anthropic over the Department of War's supply-chain risk designation; the company has asked for a full-court rehearing.

The pattern is that model access can change on regulatory decisions rather than product decisions. Any architecture that hardcodes a single model identifier and a single provider is exposed to that.

What the document does not answer

Several blanks are deliberate at this stage. There is no price range and no share count. Third-quarter revenue is left empty, although the Financial Times reported earlier in October that Anthropic told investors it reached $16.8bn. Total committed spend is blank. There is no target valuation in the filing, and the reported November timing comes from press sources, not from Anthropic.

Treat the blanks as unknowns rather than as signals. A preliminary prospectus is designed to be amended, and the numbers that decide the outcome — price, size, final valuation — will arrive in a later filing.

What to do with this if you build on a hosted model

Four things are worth acting on now, whatever happens on Nasdaq.

Keep a provider abstraction. A thin interface between your application and the model vendor costs a week of engineering and buys you the ability to switch when terms, availability or pricing change.

Read the commercial terms you already signed. Data retention, deprecation notice periods and price-change clauses matter more than benchmark scores when a vendor's cost base is locked into multi-year capacity deals.

Model your spend against expansion, not seats. A 900% net revenue retention figure describes a pricing model where usage grows faster than headcount. Budget for consumption growth and set internal caps before procurement does it for you.

Watch the disclosures, not the discourse. The useful information here was in the commitments table, the customer concentration note and the risk factors — not in the valuation chatter. The same sections will be updated every quarter once the company is public, which makes them a free early-warning system for anyone whose product depends on these APIs.

Anthropic's filing describes a company that has moved from research lab to infrastructure supplier: long contracts, a small number of very large customers, and a cost base committed years ahead. Whether or not the listing goes ahead in November, that shape is now visible, and it is the part that will affect what you pay per token.