Anthropic’s S-1 makes clear it’s on the hook for billions in compute. Its customers can walk at any time

Anthropic’s new S-1 reads like a growth story with a trapdoor underneath it. The company has been selling AI mostly on usage, and that usage is exploding: first-half revenue hit $16.28 billion, up 1,300% from a year earlier. Roughly $13.1 billion of that came from enterprise customers paying by consumption, not by monthly seat. That is a strong business right up until the bill for the underlying compute arrives.

And the bill is enormous. Anthropic says it has committed at least $475 billion in future cloud compute purchases with partners including Amazon, Google parent Alphabet, SpaceXAI, Microsoft, Nscale and Lambda. If chip-leasing deals are included, the total reaches as much as $750 billion. Much of that spending is locked in regardless of how much customers actually use the models.

The problem is on the other side of the ledger. Anthropic’s contracts often let customers reduce usage or walk away with limited or no notice. The company also says it has a habit of underestimating its compute needs, which has left it capacity constrained before. In the filing, it warns about periods of excess or underused compute and cash outflows that do not produce immediate revenue. It even cites a stretch ending Jan. 2 when customer spend fell about 20% from the prior two weeks before rebounding.

For now, the math still works. Anthropic says it makes more than two dollars of revenue for every dollar spent on computing infrastructure, and in the second quarter it reported $974 million in operating income. It also had $112.4 billion in cash and cash equivalents at the end of June. But the company is warning that as new compute comes online, margins will get squeezed, especially if it has to lower token prices or if growth slows.

That is why Anthropic is pushing enterprise buyers into multiyear, noncancelable deals with fixed minimum spends. It had about $10.8 billion in remaining performance obligations under those contracts as of June 2026, though 75% of that revenue is expected to be recognized over the next year. The catch is obvious: the revenue gets booked relatively soon, while the data center commitments keep running far beyond that.