Anthropic can’t cancel most of its $518B compute bill

Anthropic has committed at least $518 billion to AI infrastructure over the next decade, and 80% of that, about $414 billion, sits in contracts it cannot cancel. That comes from the company’s confidential IPO prospectus, which Reuters reviewed and reported on on September 29. The commitments go to Google, Amazon, Microsoft and Broadcom.

With a non-cancelable commitment, the payment is owed whether or not Anthropic ends up using the capacity. If demand slows, the bill stays the same. How the $518 billion splits among the four partners has not been disclosed.

The same prospectus shows 2025 revenue of almost $4.6 billion and a net loss of $42 billion. Compute was the largest cost that year, at $7.33 billion, which was more than half of $12.65 billion in operating expenses.

Revenue has been growing quickly. In its announcement of the Google and Broadcom deal, Anthropic said its run-rate revenue had passed $30 billion, up from about $9 billion at the end of 2025. Run-rate revenue is current revenue scaled up to a full year. The company also said the number of business customers each spending more than $1 million a year rose from over 500 at its February Series G round to over 1,000.

The Google and Broadcom agreement covers multiple gigawatts of next-generation TPU capacity, expected to start coming online in 2027. Most of it will be in the United States, and Anthropic calls the deal an expansion of the $50 billion it pledged for American computing infrastructure in November 2025. Claude is already trained and served on AWS Trainium, Google TPUs and Nvidia GPUs, and Amazon remains Anthropic’s primary cloud provider and training partner.

CFO Krishna Rao called the TPU deal “our most significant compute commitment to date”.