Anthropic has been shopping for compute like it's going out of style. Over the past 11 months, it signed contracts for at least 14.8 gigawatts (GW) of compute capacity, to be used over the next several years, according to The Information. Based on the public agreements, the potential total cost comes to as much as $517 billion — most of it due in the next decade. Add the 1–2GW the company locked in before October of last year, and Anthropic has roughly 16GW in total compute commitments.
The expansion is being driven by demand for Claude. Growth in Claude Code and Cowork has exceeded Anthropic's expectations this year, pushing the company to grab capacity while it can. A new Amazon deal can provide up to 5GW; Google and Broadcom add another 5GW; Microsoft and Nvidia contribute up to about 1GW. Anthropic also leased the full capacity of SpaceX's Colossus 1, picking up more than 220,000 Nvidia GPUs in one shot — including H100, H200, and GB200.
Yet even this massive footprint may still be smaller than OpenAI's roadmap. Anthropic has around 16GW locked in, with quite a few contracts extending past 2030. OpenAI, meanwhile, has told investors it's targeting 30GW by 2030 and expects cumulative compute spending of roughly $750 billion by then.
The $517 billion figure isn't a guaranteed tab. The Information arrived at it as the potential maximum cost based on current cloud, chip, and data-center contracts. Some capacity could be delayed or never used, and some agreements allow for early cancellation.
Source: The Information
Join the 动察 Beating channel or the 动察 Beating discussion group.