Datacenter Dynamics reports that Anthropic is considering a $10 billion compute-leasing arrangement with Meta, a figure that — if accurate — would rank among the largest bespoke compute deals in the AI industry to date. The catch: neither company has confirmed terms, so treat that $10bn number as a placeholder until a filing or an official statement nails it down.
What’s notable is the structural shift implied here: Meta, long an AI lab that buys and hoards its own GPUs, would be renting out surplus capacity to a frontier competitor rather than just using it internally. That’s the same playbook Microsoft, Google, and Oracle have run for years, and it suggests Meta’s Llama-era infrastructure buildout has outpaced its own model roadmap enough to monetize the excess.
Meta selling compute to Anthropic isn’t just a capacity deal — it’s Meta hedging its AI bet by becoming everyone else’s landlord.
For data and compute markets, this is the story to watch: as training runs get more expensive, even well-funded labs like Anthropic are shopping for capacity wherever it’s cheapest or most available, and hyperscalers with spare silicon are happy to become de facto GPU brokers. Expect more of these arrangements to surface as compute, not data, becomes the binding constraint on frontier AI progress — assuming this particular deal actually closes.
As Meta increasingly looks set to become a cloud provider
