Anthropic is in early talks to lease $10 billion of Meta computing power
The New York Times reports a two-year arrangement with monthly payments and walk-away rights for both sides. It would seed Meta's cloud ambitions and extend Anthropic's compute hedge.
- Anthropic is in very preliminary talks to lease roughly $10 billion of computing capacity from Meta over two years, the New York Times reported citing three sources, with CNBC separately confirming.
- The structure under discussion has Anthropic paying monthly, with either party able to walk away before the term ends; talks began in June and both companies declined comment.
- It would follow Anthropic's $45 billion, three-year capacity deal at SpaceX's Colossus 1, and would seed a nascent compute-rental business for Meta.
The biggest deals in AI are increasingly for electricity and silicon, not equity. Anthropic is in very preliminary talks to lease roughly $10 billion of computing capacity from Meta over two years, the New York Times reported Friday citing three confidential sources, with CNBC separately confirming the discussions, which continued to draw coverage through the weekend. The structure under discussion has Anthropic making monthly payments, with either side able to walk away before the term ends. Talks began in June, both companies declined comment, and no agreement is guaranteed.
For Anthropic, the logic is diversification at scale. The company signed a $45 billion, three-year agreement, about $1.25 billion a month, for capacity at SpaceX's Colossus 1 data center just weeks ago, and a Meta arrangement would spread its training and inference load across yet another provider's buildout, the same capacity scramble that has startups locking billion-dollar compute deals well before revenue justifies them.
Meta, landlord
For Meta the talks would validate a business it has only hinted at. CEO Mark Zuckerberg said in May that companies had approached Meta about buying computing capacity at premium rates, and the company's 2026 capital expenditure is projected around $145 billion, up from $72 billion in 2025, overwhelmingly for AI infrastructure, per Quartz. Renting spare gigawatts to a rival lab turns capex into revenue and makes the overbuild itself the product.
The desk's caution
Preliminary means preliminary. But the direction of travel is consistent with everything on this beat lately, from Moonshot racing its revenue curve to a listing to memory markets whipsawing on AI capacity news: the industry's scarcest asset is scheduled compute, and the companies that own it are becoming everyone else's counterparty.