Anthropic Files for IPO: $7 Secures 50.1% Voting Control, $800 Billion Compute Bill

On October 26, Anthropic officially filed its IPO application publicly with the U.S. Securities and Exchange Commission under the ticker symbol ANT, with a target valuation of $1.5 trillion to $2 trillion. Tech media have already written over and over about "$16.2 billion in half-year revenue," "$80.7 billion in net loss," and "a valuation of $1.5 trillion to $2 trillion, expected to surpass the historical record set by SpaceX." If we repeated the numbers again, readers would probably just yawn and scroll past.

But this prospectus, over 300 pages long, is worth reading slowly. Because what it answers is not "how much money Anthropic makes," but a more fundamental question: when an AI lab grows to trillion-dollar scale, what kind of company does it actually become?

After flipping through it, the impression is: it looks more like a combination of three companies—a compute "energy" purchaser, a B2B enterprise software company, and a mission experiment with a special governance structure.

In the prospectus, Anthropic calls compute a "critical input," with total signed contracts exceeding $800 billion, extending as far as 2045. The brokerage tally based on the contract table comes to at least $657 billion; the difference comes from the scope of counting—Google's additional $125 billion to $150 billion in TPU purchases from 2028 to 2031 was not even included in the contractual commitment table.

A few of the big ones include: Broadcom AI accelerator equipment leasing of about $161.2 billion, non-cancelable over five years; Google at least $114.5 billion, through July 2033; Amazon at least $110 billion, through April 2036; Microsoft at least $44.9 billion, through April 2036; SpaceX's AI subsidiary, up to $120.5 billion, through 2029, with an estimated roughly $3.1 billion per month once all compute is online and at full load in December; AMD, with Helios purchases starting next year, expected to exceed $20 billion, plus options that could double it; Nscale about $44.6 billion, Lambda about $41.3 billion.

The key point is not the large total, but the terms. The contracts with Google, Amazon, and Microsoft all contain the same phrase: "regardless of actual usage"—whether you actually use it or not, you have to pay. Google and Amazon also agreed that if usage falls short, the difference must be made up.

This kind of "take-or-pay" clause was previously most commonly found in long-term contracts for natural gas and electricity.

And what about Anthropic's revenue side? About 80% is settled based on actual usage—customers pay for how much they call. Costs are locked in more than a year in advance, while revenue fluctuates with call volume. This is a typical heavy-asset infrastructure financial structure: extremely high operating leverage, but once demand falls short of expectations, long-term contracts become a hard constraint.

So Anthropic is essentially purchasing compute as energy: long-term contracts, take-or-pay, costs locked in advance. The financial structure of large-model companies is increasingly resembling that of a heavy-asset infrastructure company.

Claude bans accounts so harshly, and the prospectus has a ledger for it

In the eyes of Chinese developers, Claude is probably one of the AI products with the strictest account bans and the most aggressive risk controls. After reading the prospectus, one finds that this is not product arrogance, but determined by the books.

In the first half of 2026, Anthropic's revenue was $16.286 billion. Of that, usage-based revenue was about $13.1 billion, accounting for 80%; subscription revenue was about $3.2 billion, accounting for only 20%. And that 20% of subscriptions also includes enterprise Team and Enterprise plans—the share of individual Pro and Max can only be lower.

Image source: Anthropic S-1 prospectus

So where does that 80% of the money come from? Enterprises and cloud service provider channels. 48% of revenue is sold through the marketplaces of AWS, Google Cloud, and Microsoft Azure; customers spending more than $100,000 annually grew from 950 in September 2025 to 1,500 in December 2025, then surged to more than 11,900 by September 2026.

Image source: Anthropic S-1 prospectus

What best illustrates the attitude is the net revenue retention rate: about 900%, meaning customers who spent $1 a year earlier spent an average of $9 in the same period this year. But when calculating this metric, individual Pro and Max subscribers are excluded on the grounds that "fixed-price subscriptions cannot reflect expansion behavior."

Even more direct is free users. The prospectus records the compute consumed by free access under "sales and marketing" expenses—$3.168 billion in the first half of 2026, including channels, marketing, personnel, and free compute. In translation: free users are part of customer acquisition cost on the books.

Anthropic coined a term for individual paying users: "prosumers"—researchers, developers, analysts, writers. Their strategic value is to serve as "kindling": first use Claude themselves, then go back to their companies and push for Enterprise procurement. But under the premise that compute is always insufficient, with fixed monthly individual subscribers on one side and Claude Code enterprise users spending 20 times more per month on the other, it is not hard to calculate who gets resources first.

As for mainland Chinese users, the prospectus gives only one institutional statement: the company has a "supported regions policy" and implements IP-based geographic blocking. They were never within the scope of service to begin with.

$7 for 50.1% of voting power

This is probably the most counterintuitive design in the entire prospectus.

Strategic shareholders like Amazon and Google poured in huge sums, and SBF spent $500 million back then to buy Anthropic stock—what they received were all Class N shares, with no voting rights. So who holds the votes in this company?

The answer lies in Article 10 of an LLC agreement in the appendix: seven people on the list, each contributing $1 to the company, $7 in total, to subscribe for 1 share of Class F stock. This 1 share has no economic interest and cannot receive a single cent of profit, but its voting power is fixed at 50.1% of all the company's voting power, and it cannot be diluted even by future issuance of new shares.

The seven people are Dario Amodei, Daniela Amodei, plus Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, and Chris Olah—the seven co-founders. The prospectus describes their role as "guardians of our mission."

There is another layer above that: the Long-Term Benefit Trust (LTBT) generally has the right to elect and remove four of the seven directors, and Dario and Daniela cannot be removed by the trust. The company is also a "controlled company" under Nasdaq rules and can be exempt from some governance requirements.

To translate: you can buy Anthropic stock, but it is very hard to buy Anthropic's steering wheel. The capital market bears stock price fluctuations, while the founding team controls the strategic direction. This is a governance structure designed for the "AI safety mission," and it is also a precondition that ordinary shareholders must accept.

Not panicking over an $80.7 billion loss—the real bill is the $800+ billion

A net loss of $80.748 billion in the first half sounds terrifying. Broken down, about $72.388 billion of it came from fair value changes in convertible notes, derivative liabilities, and SAFEs.

The mechanism is not complicated: part of Amazon's investment is in convertible notes, equivalent to an IOU to "convert into stock at an agreed price later." Anthropic's valuation was $380 billion in February this year and rose to $965 billion in May, so the IOUs appreciated substantially, and accounting rules require the appreciation to be recorded as a company loss. The prospectus notes that such changes were mainly driven by valuation increases and did not involve cash outflows.

On a non-GAAP basis, adjusted net profit for the first half was a positive $2.671 billion; second-quarter operating profit was $974 million, with an operating margin of 8% and an adjusted operating margin of 19%; net cash inflow from operating activities in the first half was $3.9 billion.

But "paper losses" are not the core risk. The real bill is the $800+ billion compute commitment and the depreciation that comes with it. The prospectus itself warns that performance may show "significant period-to-period volatility" due to compute spending.

On the positive side, efficiency is improving: revenue generated per $1 of compute investment rose from $0.40 in Q1 2025 to $2.08 in Q2 2026. Reasons include improved training efficiency, improved inference efficiency, simultaneous scheduling across four types of chips—Amazon Trainium, Google TPU, Nvidia GPU, and AMD GPU—and price cuts on new models.

"Revenue per $1 of compute" in the prospectus Image source: Anthropic S-1 prospectus

To translate: the loss is terrifying but not bleeding cash; what you really need to watch is not the $80.7 billion on the income statement, but the $800+ billion in the contract table.

Valuation and risk, written into the same document for the first time

The risk factors section of the prospectus runs more than 90 pages, about three times the comparable disclosure by SpaceX, and is very different from the usual boilerplate corporate disclosures.

The document warns that its technology may pose a "catastrophic or existential risk" to humanity and may "substantially assist malicious actors intent on developing chemical, biological, or other dangerous weapons." The loss-of-control scenarios described include the model potentially exhibiting "unexpected self-preservation behavior" in the future, such as resisting shutdown, concealing or manipulating information, and possibly generating coercive outputs resembling "extortion."

As for whether insurance can cover it, the prospectus says: no single insurer, or even the entire insurance industry combined, may be able to underwrite it.

A few details are worth noting: in the first nine months ended September 30, about 5%–7% of revenue came from customers later determined to have violated usage policies or commercial terms, because there was a time lag between discovery, investigation, and account suspension, and revenue for services already provided before suspension was recognized as usual—roughly $1.6 billion to $2.3 billion. Bans were stepped up in August and September, and the company cautioned that this could weigh on future revenue and growth.

The seven founders advise potential investors in the document "not to view your purchase purely as a trade," and acknowledge that company decisions "may run counter to investor expectations," because maximizing profit must be weighed against the "public interest."

A company preparing to list at a trillion-dollar-scale valuation is at the same time publicly warning that its technology may bring existential risk. Valuation pricing and risk pricing have been put into the same document for the first time.

Final thoughts

Closing the prospectus, Anthropic's true face is already clear: revenue like enterprise software, costs like long-term energy contracts, governance like a foundation, and risks that are systemic, spillover, and public in nature.

As for whether this option is worth $1.5 trillion to $2 trillion, the next things to watch are October's ARR (Run-Rate Revenue), how the compute contracts are implemented, and whether enterprise customers' token usage can continue to support that steep curve.

The hype has passed, but the questions in this document are only just beginning to be priced.