Meta-Anthropic AI compute lease talks could hit $10bn

The New York Times and Reuters say Meta is in early talks to lease data-centre compute to Anthropic, in a possible arrangement worth up to $10 billion over two years. The discussions are still preliminary and could fall over. The size alone explains the attention: for frontier AI companies, chips, power and fast networking are now as contested as the models themselves.

No binding deal has been announced by either company.

The structure under discussion is an infrastructure lease, not a product tie-up or an equity investment. Anthropic would be buying time on Meta capacity rather than waiting for new data centres to come online under its own name. That gives the reported figure some weight, even if the talks never result in a signed contract. It is a rough market price for scarce AI compute.

CNBC reported that Meta has projected as much as $145 billion in 2026 capital expenditure, with AI infrastructure accounting for much of the spend. Anthropic, meanwhile, is still funding a costly model race against OpenAI, Google and other developers. For Australian enterprise buyers, there is no immediate local launch or price change to watch. The practical point is procurement: capacity guarantees are becoming part of what customers buy when they choose an AI supplier.

Meta has also been testing whether its own build-out can become a revenue line. Earlier this month, Reuters reported that Meta was building a cloud offering to sell excess AI computing capacity. A lease to Anthropic would fit that plan, moving the company further into a market now served by hyperscalers and specialist GPU lessors.

Zuckerberg had already left the door open. In comments carried by Reuters, he said firms were approaching Meta to buy model access or spare capacity “almost every week”. “It’s definitely on the table,” he said.

What the talks could mean for AI infrastructure

If Meta rents large blocks of capacity to Anthropic, the first pressure may land on specialist infrastructure groups that have benefited from the AI build-out. Reuters said CoreWeave shares fell 10.8 per cent and Nebius dropped 12.4 per cent after the report. The share moves suggest investors quickly read Meta as a possible rival to neocloud providers selling GPU capacity to AI labs.

D.A. Davidson analyst Gil Luria made a narrower call. Meta’s entry would probably pressure smaller capacity sellers rather than the largest cloud groups, he told Reuters. “The impact of adding Meta’s capacity to the market is more likely to be on neoclouds than the big hyperscalers,” Luria said.

For Anthropic, outside infrastructure could ease a bottleneck without waiting for more of its own capacity. For Meta, a deal of this size would test whether heavy AI spending can produce external revenue, not just support its advertising systems and Llama models. There is still plenty to settle: delivery terms, pricing and how much spare capacity Meta can really commit. Signed or not, the episode points to the same constraint in frontier AI. Demand is not the choke point. Compute is.