AINEWS
The AI lab behind Claude has signed roughly $518 billion in cloud infrastructure commitments — one of the largest on record, most of it on take or pay terms, meaning the bill stays due even if usage falls short, with Amazon and Broadcom carrying
Anthropic, the San Francisco artificial-intelligence lab behind the Claude chatbot, filed a draft of its IPO paperwork with U.S. regulators in late September. The confidential prospectus, first reported by Reuters, lists at least $518 billion in commitments to cloud and data-center providers across roughly a decade, with six partners. About 80% of the total is non-cancelable, or still requires payment if Anthropic's usage falls short, a structure cloud contracts call "take-or-pay." For context, $518 billion is roughly the annual economic output of Belgium.
The largest single line is with Amazon. Anthropic has committed more than $100 billion to Amazon Web Services, the retailer's cloud-computing arm, for the period May 2026 through April 2036, and the obligation runs regardless of whether Anthropic actually uses the capacity, according to Reuters reporting republished by KSL. Anthropic declined to comment on the September 28 and 29 reports.
Two other cloud partners carry similar exposure. Google has at least $111.1 billion in infrastructure service obligations through July 2033, and Microsoft has $31.4 billion in obligations through May 2033, non-cancelable except for Microsoft's uncured material breach, the same filing shows. Equipment leases with Broadcom add another roughly $161.2 billion that are largely non-cancelable. Most of the $518 billion sits outside Anthropic's optionality.
A separate deal with Elon Musk's xAI is the structural counterpoint. Up to $84.5 billion in capacity with Nvidia-based hardware runs through 2029, but the contract is largely cancelable with 90 days' notice. It is the only major line in the prospectus that gives Anthropic an off-ramp, Reuters reports.
The math that determines whether the bill is affordable is on the other side of the ledger. Anthropic booked nearly $4.6 billion of revenue in 2025 and spent $7.33 billion on compute and infrastructure that year, according to a separate Reuters report republished by CNBC. The 2025 net loss reached nearly $42 billion, but about $34 billion of that was financing-related accounting charges; operating losses excluding those writedowns came to more than $8 billion.
The growth rate is the part the prospectus is selling. The Motley Fool, citing the disclosures, reports a Q2 2026 revenue figure of $11.5 billion, up roughly 14 times year-over-year, though Anthropic has stopped disclosing full quarterly revenue in the draft filing. If that trajectory held, the 2025 revenue line would be a rounding error by 2027.
Nearly a quarter of 2025 revenue came from just two customers, and many of Anthropic's largest clients lack long-term contracts and could reduce or stop spending, the September 28 report says. That is a normal disclosure for a young enterprise business, but it is the variable the take-or-pay structure does not protect against.
The public-market question is now on the calendar. Anthropic's revenue has to keep compounding at near-vertical rates for the next decade to justify the cloud bill it has already signed, and if it does not, the suppliers carrying the most exposure are Amazon, and Broadcom, in roughly that order.