Meta in Talks to Lease $10B of Compute to Anthropic

A potential compute-leasing agreement between Meta and Anthropic could reshape how the AI startup sources the infrastructure it needs to train and serve its Claude models, while offering Meta a new revenue stream well outside its advertising core.

Key points

• Meta and Anthropic are in early talks over a deal that could see Meta lease computing power to Anthropic for up to $10 billion over two years, according to reporting citing the New York Times.
• Anthropic reportedly initiated the conversation, proposing the deal in June; Meta is now weighing the terms.
• Both sides would retain an option to exit the agreement early, suggesting the final structure remains fluid.
• The deal would give Meta a path to monetize its massive GPU buildout beyond its own AI product needs.
• China’s Moonshot AI released Kimi K3, a model being positioned as a rival to Claude and ChatGPT, adding competitive pressure from abroad even as Anthropic manages its infrastructure situation at home.

What the deal structure signals

At $10 billion across two years, this would rank among the largest compute-leasing arrangements in the AI industry to date. The scale matters because it implies Anthropic needs access to GPU capacity that goes beyond what its existing cloud partnerships with AWS and Google can efficiently or economically provide, at least for certain workloads.

Anthropic proposing the deal, rather than Meta pitching it, is worth underscoring. The startup is effectively shopping for capacity on its own terms, which suggests its growth in inference demand is outpacing its current supply arrangements. For a company on a widely reported path toward a potential IPO, locking in multi-year compute access at known costs would also help stabilize a key line item in any prospectus.

Why Meta would say yes

For Meta, the logic is different. The company has committed tens of billions of dollars to AI infrastructure, partly to power its own products but also in a deliberate bet that surplus capacity becomes an asset. Leasing compute to a well-capitalized external buyer like Anthropic converts sunk capital expenditure into recurring revenue, diversifying away from the advertising business that still drives the overwhelming majority of Meta’s income.

The early-exit clauses on both sides, as reported, reflect how uncertain long-range AI planning remains. Neither company wants to be locked into a decade-long obligation when model architectures, chip generations, and market conditions are all moving quickly. A two-year horizon with optionality is the pragmatic middle ground.

The competitive backdrop

The Kimi K3 news landing the same week is a reminder of the broader pressure Anthropic operates under. Chinese labs have continued to close benchmark gaps with leading Western models despite U.S. export controls on advanced chips. If Kimi K3 genuinely rivals Claude on key tasks, the competitive argument for Anthropic securing more compute at scale becomes stronger, not weaker. More capacity means faster iteration, larger training runs, and the ability to push model capability forward without rationing GPU hours.

It also adds a wrinkle for investors watching Anthropic’s pre-IPO trajectory. Gross margins in AI are heavily influenced by infrastructure costs. A deal that gives Anthropic predictable, large-scale compute access could be a positive signal for unit economics, provided the pricing is competitive. The talks are early, and no agreement has been finalized, so the actual economics remain unknown for now.

Both companies declined to confirm the discussions publicly, and the terms described in current reporting should be treated as preliminary. The situation is developing.