Anthropic is inching toward an IPO, and its latest S-1 update is finally providing more visibility into the economics of compute — specifically, how much of its fast-growing revenue is lost to compute costs. The filing, submitted Monday as an update to the S-1 released last week, extends financial disclosures through September 30; the earlier version stopped at June 30. The quarter itself was strong: revenue reached $17 billion, up from $11.5 billion in the previous quarter, suggesting durable growth. At the same time, operating losses grew, thanks in large part to rising compute costs.

Compute margins got worse rather than better — a problem at a moment when the company could win over IPO investors by showing improvement. Compute consumed 58% of revenue in the latest quarter, up from 48% the quarter before. A chart in the filing tracks revenue per dollar spent on compute: the ratio grew through 2025, peaked at just over 2x during the quarter, then fell to 1.72x. Anthropic‘s compute bill reached $10 billion for the quarter, including $5.0 billion of compute under short-term commitments, which the filing says currently costs significantly more than comparable compute procured under multi-year commitments. Some of that short-term capacity is already public: Anthropic has been paying SpaceX $1.25 billion a month for compute, alongside multi-billion-dollar deals with Nscale, Akamai, and Volta. These are neoclouds — smaller outfits that offer flexible compute deals at higher rates than hyperscalers.

The filing projects the pressure will continue: Anthropic expects compute and infrastructure expenses to increase in absolute dollars and remain its largest operating expense for the foreseeable future, as additional committed compute becomes available under multi-year arrangements with cloud providers and as it continues to invest. Anthropic is trying to reduce its dependence on expensive short-term capacity through an internal program called Anthropic Direct, which aims to build its own compute base. But with demand growing so fast, the article says those efforts are unlikely to keep pace, and neoclouds will remain the best option for the near term.

Many indications suggest the margin shift may be a blip rather than a long-term problem. Revenue is growing at nearly 50% a quarter, so investors are unlikely to worry about high compute costs for now; if demand slows, the compute picture could look much sunnier. Still, the episode is a reminder of how dependent AI labs are on the trillion-dollar data center buildout. If that buildout slows over the next year — because bond markets get skittish or local opposition gets more intense — Anthropic has little ability to counter it, and a messy compute picture would get messier. For the moment, Anthropic and other labs remain at the mercy of compute prices.