Anthropic in talks to lease up to $10B of Meta AI compute over two years

Anthropic is in talks to lease up to $10 billion of Meta's AI compute capacity over two years, according to reports from IBTimes and others. The deal would be a landmark on two fronts: it would give Anthropic a major new source of training and inference capacity as it races against OpenAI and Google, and it would mark Meta's transformation from an internal-only infrastructure operator into a commercial cloud provider.

That commercial pivot has a name — 'Meta Compute' — a business Meta reportedly plans to build to sell excess AI capacity, potentially anchored by the Anthropic lease. Meta shares jumped 10.1% on the cloud-business reports, per Kavout, and its Louisiana data center is scaling rapidly to support Llama models and now, potentially, third-party workloads. The move would put Meta in direct competition with AWS, Azure, and Google Cloud for AI compute rental.

Separately, PitchBook's Q3 2026 analysis argues Anthropic is winning on task-completion value rather than token price — a framing that matters as Chinese open-weight models (Kimi K3, Qwen 3.8) trigger a race-to-the-bottom on per-token pricing. Anthropic's bet is that enterprises pay for reliable task completion, not the cheapest tokens.

The deal mechanics remain unconfirmed and no contract is signed. Skeptics question whether Anthropic wants dependence on a competitor's infrastructure, and whether Meta can operationalize a cloud business at hyperscaler standards. Competitively, the arrangement is unusual — Meta open-sources Llama while renting compute to a closed-model rival. Readers should watch for confirmation of terms, whether other labs sign 'Meta Compute' deals, and how this reshapes the compute-supplier landscape long dominated by the big three clouds.