Anthropic is still grappling with the economics of compute

Compute costs for AI models have become significantly more expensive than initially anticipated.

Anthropic's latest financial filings reveal a worsening trend in its compute economics as the AI company approaches an IPO. Despite a substantial revenue increase to $17 billion in the third quarter, its operating loss also grew, largely driven by escalating compute costs. Compute expenses now account for 58% of revenue, up from 48% in the previous quarter. This margin deterioration is attributed to the company's need to secure compute capacity quickly, leading to higher costs for short-term commitments with "neocloud" providers like SpaceX, Nscale, Akamai, and Volta, rather than relying solely on long-term, lower-cost contracts with hyperscalers. Anthropic spent $10 billion on compute in the quarter, with half of that on these pricier short-term deals. The company acknowledges that compute and infrastructure expenses will continue to rise and remain its largest operational cost in the foreseeable future, even as it pursues its own compute infrastructure project, "Anthropic Direct."

Anthropic's increasing reliance on "neocloud" providers for compute at higher rates, despite soaring revenue, signals a critical bottleneck in the AI infrastructure supply chain. The company's immediate need for capacity, evidenced by $5 billion in short-term commitments, highlights the difficulty in scaling AI operations efficiently when faced with exponential demand growth. This dependency on more expensive, flexible compute sources directly impacts profitability and raises concerns for its upcoming IPO, as investors might scrutinize these escalating operational expenses.

The situation underscores the broader vulnerability of major AI labs to the trillion-dollar infrastructure buildout and potential disruptions within it. As Anthropic's own efforts like "Anthropic Direct" are outpaced by demand, the company's future financial health is increasingly tied to the stability and pricing of external compute providers. Any slowdown in data center construction or shifts in the bond market could exacerbate these compute cost challenges, making the economics of running large-scale AI models a persistent concern.