Anthropic has signed a multi-year compute deal with CoreWeave, making it the ninth of the ten largest AI model providers to run frontier training workloads on the specialised hyperscaler. The announcement comes one day after CoreWeave closed a $21 billion compute agreement with Meta — the single largest training deal in the company’s history.
The deal
Financial terms were not disclosed. Sources briefed on the agreement describe it as a “multi-year, multi-gigawatt” commitment that complements Anthropic’s previously announced 3.5 GW compute arrangement with Google and Broadcom. CoreWeave will provide dedicated GPU capacity — a mix of Nvidia H200 and Blackwell-class systems — at undisclosed US data centers.
CoreWeave’s run
The addition of Anthropic means CoreWeave now serves 9 of the top 10 AI model providers — a remarkable concentration for a company that was essentially a crypto-mining GPU reseller three years ago. Its customer list now spans OpenAI, Microsoft, Meta, Anthropic, Mistral, Cohere, Inflection, and several undisclosed frontier labs. Only one of the top ten — believed to be a Chinese national champion — is not on the roster.
The Meta precedent
The $21 billion Meta deal announced on Monday set the price floor. Meta committed to CoreWeave for multi-year GPU capacity to train the next generation of Llama and its emerging agent stack. That deal — widely seen as the largest training agreement ever signed — made the Anthropic deal easier to price.
Why this matters
The concentration of frontier training on a single hyperscaler is both efficient and fragile. Efficient because CoreWeave can optimise for the narrow workload of frontier model training in ways the general-purpose cloud cannot. Fragile because a single operational, geopolitical, or regulatory event at CoreWeave now touches 90 percent of the AI industry. The model labs know this. Most are quietly diversifying — Anthropic’s Google-Broadcom deal is itself evidence of exactly that.
Leave a Reply