Anthropic discloses $5 billion of short-term compute Anthropic discloses $5 billion of short-term compute

Buying time for dedicated infrastructure

The filing describes Anthropic Direct, a procurement approach that separates data-center capacity from the chips installed inside it. Facilities take longer to prepare, so Anthropic can secure space earlier and defer accelerator commitments until its needs are clearer. It expects dedicated infrastructure to replace temporary capacity over time. Total purchase and lease commitments reach $895.4 billion, substantially for compute and infrastructure. The commitments extend through 2045.

Amazon’s expansion has several stages

The supply timetable predates the filing. On April 20, Anthropic announced an Amazon agreement covering up to five gigawatts of new capacity and more than $100 billion of AWS spending over ten years. It expected nearly one gigawatt of combined Trainium2 and Trainium3 capacity by the end of 2026. The agreement spans successive chip generations, with options for future hardware, rather than one immediate delivery. Anthropic also said it already used more than a million Trainium2 chips for training and serving Claude. The planned expansion included additional inference capacity in Asia and Europe. At that announcement, the company acknowledged that rapid consumer growth had affected reliability and performance for Free, Pro, Max and Team users, particularly at peak hours.

Google’s next expansion starts later

Google’s April 6 announcement put a different part of the buildout on a later schedule. Multiple gigawatts of additional TPU capacity were expected to begin coming online in 2027. The arrangement combines Google Cloud services with access to Google-designed processors supplied through Broadcom. Google described this expansion as supporting Anthropic’s growing model workloads. Those deliveries follow a different timetable from the Claude services that customers already access through Google Cloud. The supplier also identified BigQuery, Cloud Run and AlloyDB as services supporting Anthropic’s data, development and applications. It named Cursor, Replit and Shopify among customers accessing Claude through Google Cloud. Those examples show the cloud relationship includes application delivery as well as future processor supply.

A supplier describes the same pressure

Amazon provided a separate indication of pressure on immediately available hardware in an October 30 post reproducing earnings-call remarks. Chief executive Andy Jassy said short-term capacity on older Trainium generations sold at a premium to the first half of the year. The company described inference as accounting for well over half of AWS’s AI revenue. It intends to offer Trainium outside AWS in late 2027. The distinction between training and serving matters here: demand for current inference capacity continues while suppliers prepare later chip generations.