The largest number in Anthropic’s registration statement is not the valuation. It is the bill.
The company filed its S-1 with the Securities and Exchange Commission on Oct. 26, setting up a Nasdaq listing under the ticker ANT in the week of Nov. 16, according to people familiar with the timetable. Investors involved in early conversations have discussed a valuation between $1.5 trillion and $2 trillion, a range that would make the offering the largest in the history of the public markets if it prices near the top.
A valuation is not a raise. Saudi Aramco’s 2019 offering raised $29.4 billion and still holds the record for proceeds. But a float of even 5% at the low end of Anthropic’s range would raise more than $70 billion, and bankers working on large technology deals rarely settle for a float that thin.
What the filing makes visible is the shape of the business. Anthropic counted more than 11,900 enterprise customers paying at least $100,000 a year as of the filing, a customer base that looks like a software company’s rather than a consumer app’s. Revenue for the six months through June was $16.29 billion, against $1.2 billion a year earlier.
The loss lines need reading in order. The net loss was $80.75 billion, a figure larger than revenue and one the company attributes mostly to fair-value changes on convertible notes, about $72.4 billion of it non-cash. The operating loss was $693 million. On an adjusted basis the company reported operating profit of $1.62 billion. Three numbers, three different pictures of the same six months, and the roadshow will be spent explaining which one to use.
Then comes the commitment schedule. Anthropic has signed compute and infrastructure agreements totaling between $542 billion and $567 billion. The largest, executed on Dec. 31, 2025, obliges the company to pay at least $114.5 billion to Google between April 2026 and July 2033, at least $110 billion to Amazon between May 2026 and April 2036, and at least $44.9 billion to Microsoft between November 2026 and May 2033. The filing states that the commitments do not shrink if Anthropic uses less capacity than it reserved.
Read those dates as a calendar of obligation. The Google agreement begins paying out this spring and runs seven years. The Microsoft agreement starts in November 2026 and runs into 2033. The Amazon agreement runs longest, a full decade, ending in April 2036. A company that stumbles in 2029 still owes for capacity scheduled into 2036.
Financing arrangements around the commitments are disclosed in the same section. Broadcom, described as the company’s largest partner for 2026, has arranged potential financing of $42 billion, with Google at $20 billion and Amazon at $15 billion. Anthropic also plans to build its own data centers, a step that converts a variable cost into capital expenditure and a construction schedule.
The structure has become standard among model developers, who stopped renting computing by the hour and started reserving multiyear blocks of data-center space, power and silicon. The reasoning is that a training run which begins late is worth less than one that begins on time, and that the largest buyer gets the allocation. Those agreements also appear in a prospectus as purchase obligations rather than debt, a distinction that matters to a credit analyst and rather less to a treasurer, since both consume cash in the quarter they come due.
Control is settled before the stock trades. Seven co-founders hold 50.1% of voting power through an entity, which keeps the founding group in charge of decisions public shareholders cannot outvote. At the $58.90 per-share price of the most recent private round, each founder’s stake of roughly 240 million shares would be worth about $14 billion.
The comparison that will occupy analysts is with OpenAI, which has stayed private and raised at a valuation near $1.25 trillion, according to people briefed on those talks, and with the smaller developers that have listed at fractions of Anthropic’s scale. A public price for Anthropic hands every private negotiation in the sector a reference point, and hands employees at every rival a number to quote in a recruiting conversation.
There is also the matter of what the customer list implies. Selling to nearly 12,000 accounts that each spend six figures is a different business from selling subscriptions to individuals. Those contracts renew on procurement calendars, they carry security reviews, and they concentrate. Analysts who have reviewed comparable filings said the risk sections in this generation of prospectuses read as descriptions of the industry rather than of one company: dependence on a handful of cloud partners, concentration among enterprise buyers, and compute costs that rise with capability.
The offering will be judged on whether buyers accept that a company can carry a decade of fixed obligations and a research budget that scales with them, and still convert revenue into cash at the end. The filing puts revenue, an operating profit, a net loss produced by accounting and a $567 billion ceiling on future spending on consecutive pages, audited to one standard. Whoever prices the deal has to hold all four in mind at once.
