Meta and Anthropic discuss $10 billion AI compute lease

Meta and Anthropic are in early talks about a computing agreement that could be worth up to $10 billion over two years. The New York Times first reported the discussions on July 17, citing three people familiar with them. Reuters and CNN subsequently confirmed that talks are taking place. Anthropic reportedly proposed the arrangement in June, but no contract has been signed. Both companies declined to comment, while a source told CNN that published numbers remain speculative.

The possible arrangement would be unusual because Meta and Anthropic compete in artificial intelligence. Meta develops its own Muse models and consumer assistant, while Anthropic builds Claude. Under the proposal, however, Meta would act as an infrastructure supplier to a rival, leasing part of its expanding stock of chips and data-center capacity. Reports say either company could retain options to leave during the two-year period, underlining how preliminary the plan remains.

For Meta, a deal could create a new return on its enormous AI investment. The company expects capital expenditure of between $125 billion and $145 billion this year, much of it for servers, networks, power and data centers. Mark Zuckerberg said in May that selling spare compute was an option if Meta found it had built more than it needed. Supplying Anthropic would move that idea from contingency plan toward a cloud business competing with Amazon, Microsoft and Google.

For Anthropic, the talks show that access to computing power is still a strategic constraint. The company already relies on several large infrastructure partners, including Amazon, Google, Microsoft and SpaceXAI. Adding Meta could diversify supply and provide more capacity for training models and serving Claude, although the reports do not identify particular facilities, chips or workloads.

Nothing changes for Claude or Meta AI users today. A final agreement could eventually support greater availability and faster product expansion, but those benefits are not guaranteed. The immediate significance is for the market: companies that compete on models may increasingly cooperate on the scarce physical infrastructure beneath them.

For makers and businesses, the story is a reminder that model quality and API prices are only the visible layer of AI. Reliability also depends on long-term access to chips, electricity and data centers. If Meta becomes a major compute supplier, customers may gain another infrastructure route, while facing a more complicated web of relationships between model developers and cloud providers.