Anthropic's compute margins worsened even as revenue surged

Anthropic is edging toward an IPO, and a Monday filing updating its S-1 is giving outsiders their clearest look yet at how much of the AI lab’s booming revenue is being consumed by compute. The update extends the prior filing’s disclosures, which ran only through June 30, out to September 30 — and the extra quarter shows compute margins getting worse just as a would-be IPO investor might hope to see them improving. The top line climbed to $17 billion from $11.5 billion a quarter earlier, a sign of durable growth, yet Anthropic’s operating loss widened at the same time, driven largely by rising compute costs. Compute alone consumed 58% of revenue in the new quarter, up from 48% in the prior one; a chart in the filing tracking revenue generated per dollar of compute spent shows that multiple topping out above 2x last quarter before easing back to 1.72x.

The filing pins much of the pressure on Anthropic’s scramble to keep up with demand: its compute bill hit $10 billion for the quarter, including $5.0 billion of compute bought under short-term commitments that cost significantly more than the multi-year deals the company prefers. Several of those arrangements are no secret: Anthropic sends SpaceX a monthly $1.25 billion compute payment, and it has separate multibillion-dollar contracts with Nscale, Akamai, and Volta — smaller, more flexible “neocloud” providers that charge higher rates than the major hyperscalers. The filing says compute and infrastructure costs will keep rising in absolute terms and remain the company’s largest operating expense for the foreseeable future, even as more committed multi-year compute comes online.

Anthropic is also trying to build its own compute base through a project called “Anthropic Direct,” but the filing suggests that effort can’t scale fast enough to match current demand growth, leaving neoclouds as the company’s best near-term option. Many indications suggest the margin dip is temporary rather than structural, and with revenue growing at nearly 50% a quarter, investors have shown little concern so far over the high compute costs.

Why it matters

The disclosure underscores how exposed major AI labs remain to swings in compute supply and pricing, a dependence that could turn costly if the broader data-center buildout slows on bond-market jitters or local opposition.