Anthropic’s Compute Costs Climb to 58% of Revenue as It Develops FlashProve to Verify AI Model Outputs
By James Dargan · · 3 min read
Anthropic is offering a clearer view of the economics of running a frontier AI lab, as an updated IPO filing shows compute costs consuming a growing share of its fast-rising revenue, while separately the company has published early research on an open source system for verifying which AI model produced a given output.
In an amended S-1 filed on Monday, Anthropic extended its financial disclosures by a quarter to September 30. Revenue reached $17 billion in the quarter, up from $11.5 billion in the previous three months, but the operating loss widened, driven largely by rising compute costs. Compute accounted for 58% of revenue, up from 48% a quarter earlier, and revenue generated per dollar spent on compute fell from a peak of just over 2x to 1.72x.
The filing shows Anthropic’s compute bill rose to $10 billion for the quarter, including $5 billion in compute under short-term commitments, which the company said costs significantly more than capacity secured through multi-year agreements. With demand surging, Anthropic has turned to smaller, more flexible but pricier providers known as neoclouds. It has been paying SpaceX $1.25 billion a month for compute and has signed multibillion-dollar deals with Nscale, Akamai and Volta.
Anthropic said in the filing that it expects compute and infrastructure expenses to keep rising in absolute terms and remain its largest operating expense for the foreseeable future as more committed capacity comes online through multi-year arrangements with cloud service providers. The company is also working to lower costs by building its own compute base through a project called Anthropic Direct, though that effort is unlikely to keep pace with demand in the near term.
The margin decline may prove temporary, but it highlights how heavily AI labs depend on the broader infrastructure buildout and how exposed they are to any slowdown caused by nervous bond markets or growing local opposition to data centers.
In other company news, Anthropic published a blog post on Tuesday detailing FlashProve, an open source system designed to deliver provable inference, a cryptographic guarantee that a particular output came from a specific model. Working much like email signatures that confirm a sender’s identity, the system is intended to resist impostors disguising one model as another, including the lab operating the model. Researchers said it could stop AI companies from quietly substituting a cheaper model to save money or a stronger one to inflate performance results, and could help users verify third-party claims that tasks such as vulnerability scanning were carried out by a particular frontier model.
FlashProve requires labs to publish a fingerprint of each model that can later be matched cryptographically to outputs. The fingerprints do not expose model weights, meaning closed-model developers such as Anthropic and OpenAI could take part without revealing their most valuable asset. The system is lightweight enough for verification on a local device.
Anthropic released a technical specification and working code on GitHub under an open source license. The researchers stressed that the work represents intermediate progress rather than a finished agenda, with no current plans to deploy it at Anthropic or elsewhere, and said they are sharing it so others can build on and improve it.
