Anthropic's $750 Billion Cloud Bill Exposes the AI Capacity War

Anthropic cloud capacity worries just went public in dramatic fashion: a leaked draft of the company's S-1 filing, reported October 20, says the Claude-maker believes it could spend more than $750 billion on cloud contracts — including deals still under negotiation — and still not have enough compute.

The filing names the tension plainly. AWS, Google, and Microsoft, Anthropic's three biggest cloud suppliers, "increasingly see the Claude-maker as a competitor rather than a customer." When your landlords are building rival products, the lease negotiations get interesting.

The Anthropic cloud capacity crunch is a landlord problem

The contracted numbers are staggering. Per the filing: $111 billion committed to Google Cloud, $110 billion to AWS, and $31 billion to Microsoft. Then there's up to $86 billion for renting Nvidia chips from smaller AI clouds like Nscale and Lambda through 2033–34, plus more than $160 billion in Google TPU obligations through an arrangement with Broadcom. Add it up and Anthropic has over $600 billion locked in — noncancelable, the filing says, regardless of actual usage.

Pause on that last part. Noncancelable regardless of usage means Anthropic pays whether it fills the capacity or not. That's the cloud industry's dream contract and the customer's nightmare: commit to a decade of GPU rent and bet the entire company on growth filling it. It's a structure that only makes sense if you believe AI demand is going in one direction forever.

The workaround is equally telling: Anthropic is leasing its own data centers through third-party operators — a path the filing flags as having "limited operating history." Translation: the plan B for when your suppliers become rivals is to become your own customer, with partners who haven't done this at scale before.

$600B you can't cancel, in plain terms

Cloud capacity is the AI industry's actual bottleneck — not model architecture, not talent, not hype. Training runs and inference at Anthropic's scale consume power and hardware measured in gigawatts and hundreds of thousands of chips. The filing makes the economics visible:

• Concentration risk. Three hyperscalers plus Broadcom-controlled TPU supply means Anthropic's compute future runs through a handful of boardrooms.
• Rivalry risk. Those boardrooms now compete with Claude directly. AWS, Google, and Microsoft all ship their own models; every contract negotiation doubles as a competitive intelligence exercise.
• Demand risk. Noncancelable contracts bet that usage grows into the commitment. If growth stalls, the bills don't.
• Build risk. The self-leasing route depends on third-party operators with, in the filing's own words, limited operating history at this game.

Data center demand has been the loudest signal in cloud for two years — and power and cooling the bottleneck underneath it. Anthropic's filing is the moment a frontier lab admits the bottleneck has a name and a price tag.

What the financials say (and don't)

The filing also sketches the business behind the bills: first-half 2026 revenue of $16.3 billion against a net loss of $80.7 billion — a loss the Wall Street Journal attributes largely to an accounting adjustment on financing agreements, not cash going out the door. Two customers account for roughly 20% of sales, which is its own concentration story. Third-quarter revenue reportedly hit $16.8 billion, up 45% quarter over quarter.

Anthropic is said to be targeting a November IPO, possibly at a $2 trillion valuation. Whether that number survives contact with public markets is a Business-desk question; the Cloud-desk question is simpler. The filing also discloses that liability risk from harmful AI behavior "would exceed available insurance coverage" — a line that deserves its own article, but in cloud terms it's one more reminder that the risks of scale don't stop at the power bill.

Here's the takeaway: the AI boom's scarcest resource was never cleverness. It's concrete, steel, transformers, and chips — and the contracts to secure them are now measured in hundreds of billions, signed with companies that would love to eat your lunch. Anthropic just put the number on the page: $750 billion, and it still might not be enough.