Meta’s $21B CoreWeave Bet
Meta agreed to spend an additional $21 billion with CoreWeave to secure rented GPU capacity, underscoring that hyperscalers are buying external compute at scale. (reuters.com) The expansion brings the relationship toward $35 billion in total and is expected to support early deployments of Nvidia’s next‑gen systems, highlighting that external GPU farms are a critical bottleneck for big models. (cnbc.com)
Meta just agreed to rent so many artificial intelligence chips from CoreWeave that the bill now approaches $35 billion, even though Meta is already spending tens of billions building its own data centers. The new piece alone is about $21 billion and runs through December 2032. (coreweave.com) CoreWeave is not a household-name app company. It is a landlord for graphics processing units, the specialized Nvidia chips that train and run large artificial intelligence models, and it packs those chips into data centers that customers can rent instead of waiting years to build their own. (cnbc.com) That is the surprise in this deal: Meta is one of the world’s richest infrastructure builders, yet it is still buying outside capacity at huge scale. Reuters reported that Meta is deepening the partnership as it races to support more complex artificial intelligence workloads after a weaker model release last year. (reuters.com) The timing tells you what Meta is really buying: speed. CoreWeave said the capacity will be deployed across multiple locations and will include some of the first deployments of Nvidia’s Vera Rubin platform, Nvidia’s next generation of artificial intelligence systems. (investors.coreweave.com) Meta is not replacing its own buildout with rentals. On its latest outlook, the company said it expects 2026 capital spending of $115 billion to $135 billion, up from about $72.2 billion in 2025, with the increase tied to artificial intelligence infrastructure and data centers. (finance.yahoo.com) So the rented chips solve a different problem. If building your own data center is like constructing a steel mill, renting from CoreWeave is like booking production lines that already exist, with power, cooling, networking, and Nvidia hardware already wired together. (cnbc.com) That helps explain why CoreWeave has become so valuable so quickly. The company was founded in 2017 and completed its Nasdaq listing in March 2025, selling investors a pure bet on one bottleneck: there are not enough ready-to-use graphics processing units for everyone who wants to train and serve big models. (investors.coreweave.com) (cnbc.com) The contract also shows that the shortage is no longer just about training new models in secret labs. CoreWeave said Meta will use its cloud platform to scale inference workloads, which is the expensive step where a trained model answers billions of user requests in products people actually touch. (investors.coreweave.com) CoreWeave is spending heavily to keep up with that demand. On the same day as the Meta expansion, CNBC reported that CoreWeave said it would raise $3 billion in fresh debt, a sign that winning giant chip-rental contracts also requires giant financing to buy servers, lease sites, and power the machines. (cnbc.com) Put together, the deal says something simple about the artificial intelligence race in 2026: even the companies trying to build their own empires still need to rent land. Meta’s check buys time, CoreWeave’s backlog buys credibility, and Nvidia’s next chip generation is already being spoken for before most people can even buy the current one. (coreweave.com) (cnbc.com)
