Meta Weighs Renting Up to $10B of AI Compute to Anthropic

Meta is reportedly weighing a deal that would put a competitor's models on its own hardware. According to a New York Times report, Anthropic first proposed the arrangement in June, offering to rent computing power from Meta's AI data centers in a deal that could reach up to about $10 billion over two years. The talks are described as early, and Meta is still considering.

The context is that Meta has spent the last two years building AI capacity faster than its own model roadmap has been able to consume, and the company is now looking for outside tenants to fill the gap. Reporting has that effort organized under a nascent unit called Meta Compute, led by infrastructure chief Santosh Janardhan, Meta Superintelligence Labs leader Daniel Gross, and president Dina Powell McCormick. Two service shapes are on the table: raw GPU capacity, or a Bedrock-style layer that sells hosted access to models running on Meta hardware, including Meta's closed-weight Muse Spark model.

If the Anthropic deal closes, it would be the second time in a year that Anthropic has moved to lock down a rival's excess capacity. TechCrunch's summary of the broader trend notes SpaceX, via xAI, has already leased Anthropic the entire Colossus 1 data center in Memphis for around $1.25 billion per month through May 2029. The pattern is that model labs are running so far ahead of their own infrastructure that they will now rent from anyone with the racks, including companies they compete with for talent and users.

The honest caveats are worth pulling out. The reporting is single-sourced to the Times, the talks are described as early, and none of the specifics that would actually matter to a buyer appear settled: which GPU generation, which site, whether Meta keeps rights to reclaim capacity for its own training runs. Anthropic's separate commitments to Fluidstack and TeraWulf sit alongside this, so the total compute footprint the company is contracting for is bigger than any single headline suggests.

What is worth watching is whether Meta actually goes ahead and becomes a compute landlord to a direct model rival. If it does, the AI infrastructure market gets a fourth serious hyperscale seller, which is a structural change that matters more than any one contract number.