Anthropic promised $518 billion. Most of it is owed either way.

The number the wires led with was the loss. The number that decides what Anthropic is actually worth is the one it has to pay whether or not anyone shows up to use it.

Anthropic's confidential IPO prospectus, reviewed by Reuters, says the company expects to spend at least $518 billion over roughly a decade building AI infrastructure with six partners, and that about 80 percent of it — roughly $414 billion on the filing's own numbers — is non-cancelable or payable regardless of usage. That includes a clause worth reading twice: if Anthropic's actual spend with Google falls short of its commitment, it must pay Google the difference. The company declined to comment, and the filing has not been made public, so everything here is as reported rather than as confirmed by Anthropic.

Set that against the rest of the document and the halves do not look alike. On the income statement: revenue grew twelvefold in 2025 to nearly $4.6 billion, operating loss widened to $8.06 billion from $2.98 billion, and net loss reached $42 billion — about $34 billion of which is a non-cash charge tied to financing that could later convert into shares. Compute and infrastructure alone ran $7.33 billion last year, a threefold increase and more than half of $12.65 billion in total operating expenses. At the end of 2025 the company held $20.28 billion in cash and short-term investments.

What the commitments actually say

The largest line is Google: at least $111.1 billion covering April 2026 through July 2033, with the shortfall clause attached. Amazon is next at $110 billion, running May 2026 to April 2036. Microsoft adds $31.4 billion from November 2026 to May 2033, non-cancelable except in the event of Microsoft's own uncured material breach. Separately, Anthropic is carrying about $161.2 billion of Broadcom-related equipment lease obligations that are largely non-cancelable by either party outside a default — the chip-supply half of the buildout, financed rather than bought.

Two of the six partners are on much softer terms, and the difference is the tell. A deal with xAI could reach $84.5 billion of Nvidia-based capacity through 2029 but is largely cancelable on 90 days' notice, and AMD has committed to buy up to $5 billion of Anthropic stock against compute capacity expected to exceed $20 billion. So the market has already priced these contracts in two tiers: multi-year take-or-pay for the hyperscalers and Broadcom, and a rolling arrangement with a competitor for capacity that can be walked away from.

The filing's own risk language concedes the awkward part. Google, Amazon and Microsoft are simultaneously investors, customers, cloud providers, distributors and competitors, so their incentives "may not be fully aligned" with Anthropic's — and if third-party compute is "curtailed, repriced, or terminated," the company says its results could suffer. That multi-role structure is the loop we have tracked across the sector all year, from Nvidia's $10 billion anchor closing the loop on circular AI money to the hyperscaler side of it. Anthropic's justification for signing anyway is that compute availability, not demand, is the constraint on AI development, and that demand from advanced systems will exceed supply.

A mortgage on the buy side, a rolling contract on the sell side

Here is the mismatch a public investor has to underwrite. Roughly $414 billion of obligations behave like a decade of fixed cost. The revenue that has to service them does not behave like a decade of fixed revenue.

Nearly a quarter of 2025's revenue came from just two customers — and the filing warns that many of its largest clients have not signed long-term contracts and could reduce or stop spending. Anthropic has not named the two. It is worth being blunt about the shape rather than the names: the counterparties most likely to be both are also the counterparties on the other side of the $252.5 billion of Google, Amazon and Microsoft obligations. Customer concentration and supplier concentration are the same concentration here, which is a structural fact about the AI stack rather than a suspicion about any one company.

The bull case on the mismatch is real and Anthropic makes it repeatedly. Through the first half of 2026 the growth has been faster than the losses: second-quarter revenue reached $11.5 billion, up from $4.73 billion in the first quarter, and the company is on track for a second consecutive quarter of adjusted operating profit. If that curve holds, a fixed compute base signed at today's prices is an advantage, not a liability — it locks in capacity ahead of everyone still negotiating for it, at a valuation that could exceed $2 trillion against the $965 billion it was priced at in May.

But the growth story and the commitment story pull in opposite directions for one specific reason: revenue is metered and cancellable, and cost is not. A lab whose customers can leave next quarter has sold itself the right to pay for a decade of capacity either way. It is the same asymmetry that defined OpenAI's leaked numbers when OpenAI's cash gap shrank while its compute bill grew — the burn improved, the amount it had promised went up.

There is a second, harder question sitting under the first. A commitment is only as good as the cash behind it, and the $1 trillion AI build-out is running into something cash alone cannot fix — the power, land and transmission the compute depends on. Anthropic can sign for $518 billion of capacity; the megawatts have to be built either way.

What the skeptics are right about

The most common dismissal of AI-lab losses is that they belong to a growth company, so they are noise. That argument works better for a business whose costs scale with its revenue than for one holding fixed obligations. Uber and Amazon could pull back on spend when demand softened; a take-or-pay clause means the bill arrives even in the quarter where the model release lands late, adoption stalls, or a cheaper open-weight competitor resets prices. The clause that forces Anthropic to pay Google the difference if it under-spends is precisely the mechanism that removes the option to slow down.

There is also a finance-side reading of the Broadcom leases that deserves more attention than it has received. Equipment lease obligations of $161.2 billion mean Anthropic is moving from renting compute on a cloud bill to owning or directly leasing chips — the shift the filing itself describes. That reduces the reps of dependency on any one cloud and increases the fixed-charge load. It is the same trade every time: more control over the supply, less flexibility on the cost.

And one caution on the whole picture. This is a leak of a confidential draft reviewed by one wire service, with no second independent read of the document. The numbers are internally consistent and several outlets have now repeated them, but nobody outside the filing has seen the 261 pages — including which 80 pages are risk factors and which 48 describe the business. Treat the totals as reported, not as audited.

What to watch

• Whether the public S-1 holds these numbers or adds to them. The filing must be released at least 15 days before a roadshow, and the offering is expected after the November midterms, so the commitments disclosed in the draft are unlikely to be the final set.
• Whether "pay the difference" becomes standard. If Google's shortfall clause appears in other labs' filings, take-or-pay has become the price of priority access — and the constraint on AI development stops being chips and becomes balance-sheet room.
• Whether the adjusted operating profit streak survives the fixed base. Two quarters is a trend; four quarters with $414 billion of committed spend landing behind it is evidence.

If a lab has to pay for compute whether or not it uses it, who is really taking the demand risk — the company, or the investors buying the IPO? Tell us in the comments.

Sources: Reuters — Anthropic's IPO prospectus shows sweeping AI vision, surging costs · Reuters via CNA — Anthropic's $518 billion buildout hinges on deals that cannot be canceled · Fortune — Anthropic's IPO income statements · CNBC — Anthropic warns AI may pose 'existential risks to humanity' · CoinDesk — the $34 billion non-cash charge inside the $42 billion loss