The social media giant is in early talks to lease its data center capacity to the Claude maker in a deal that could reshape the AI infrastructure market
Meta Platforms is in preliminary discussions with Anthropic to lease AI computing capacity from its data centers in a potential arrangement worth up to $10 billion over two years. The proposal, initiated by Anthropic in June 2026, would effectively transform Meta from a consumer of AI infrastructure into a seller of it.
The deal structure reportedly involves monthly payments with both companies retaining early termination options.
Meta’s infrastructure pivot
CEO Mark Zuckerberg has publicly expressed interest in selling excess computing power, positioning Meta alongside the traditional hyperscalers like Amazon Web Services, Microsoft Azure, and Google Cloud.
Meta has raised its 2026 capital expenditure guidance to between $130 billion and $145 billion. For context, that’s roughly double the approximately $72 billion the company spent in 2025. Nearly all of that increase is directed toward AI data center development.
Why Anthropic needs the compute
Anthropic, the maker of the Claude family of AI models, recently locked in a $45 billion, three-year deal with SpaceX for computing resources. By initiating discussions in June 2026, Anthropic appears to be locking in access before the market tightens further.
What this means for the AI infrastructure landscape
The negotiations are still early-stage, and sources indicate the talks could fall through or change significantly before any formal agreement is reached. Neither Meta nor Anthropic has commented publicly on the discussions.
For investors watching Meta’s stock, the potential deal carries mixed signals. On one hand, monetizing data center capacity would help offset the company’s massive capex increases, addressing one of the primary concerns Wall Street has flagged about Meta’s AI spending spree. On the other hand, the deal’s termination clauses mean revenue could evaporate quickly if Anthropic finds cheaper or better alternatives.
