Meta's $21B CoreWeave Deal
Meta agreed to a fresh $21 billion cloud‑compute pact with CoreWeave to lock in large-scale AI capacity through 2032, effectively outsourcing parts of its infrastructure build to a specialist supplier. The deal underscores how scarce, GPU‑heavy compute is creating strong revenue visibility for niche cloud providers while increasing customer‑concentration risk for those vendors. (reuters.com)
Meta just agreed to spend another $21 billion with CoreWeave for artificial intelligence computing, and the striking part is that Meta is still outsourcing a huge slice of capacity even while planning up to $135 billion of capital spending in 2026. (cnbc.com) (reuters.com) The new commitment runs from 2027 through 2032, and it comes on top of an earlier $14.2 billion arrangement between the same two companies that was disclosed in September and runs through 2031. (cnbc.com) (bloomberg.com) CoreWeave is not a social network or a software app company. It is a specialist cloud provider that fills data centers with Nvidia graphics processing units, the chips used to train and run large artificial intelligence models. (cnbc.com) (coreweave.com) Meta said this expanded deal is for “inference” work, which is the part where an already-trained model answers questions, writes text, ranks ads, or powers features inside Facebook, Instagram, and WhatsApp. Training is like teaching a student for months; inference is the student taking millions of tests every day. (coreweave.com) That distinction explains why Meta is buying so much outside capacity. Inference needs huge numbers of chips spread across many locations because the work arrives nonstop from billions of app interactions, and delays of even a second can ruin the product. (coreweave.com) (reuters.com) CoreWeave said the capacity will be deployed across multiple sites through December 2032, and some of those deployments will use Nvidia’s Vera Rubin platform, which is Nvidia’s next-generation system for heavy artificial intelligence workloads. (coreweave.com) For CoreWeave, this kind of contract turns a young cloud company into something closer to a utility with years of booked demand. The company said in February that its 2025 revenue hit $5.131 billion and its revenue backlog reached $66.8 billion. (investors.coreweave.com) But the same deal that gives CoreWeave visibility also sharpens one of its biggest risks: dependence on a few giant customers. Reuters reported that Microsoft made up 67% of CoreWeave’s 2025 revenue, so every new mega-contract helps growth while also tying the company’s future to a small number of buyers with enormous bargaining power. (reuters.com) (investors.coreweave.com) For Meta, the message is that building its own data centers is not enough on its own. CNBC reported that Meta is also spending $10 billion on a Texas data center, but Chief Executive Officer Mike Intrator said customers like Meta keep buying from CoreWeave because “there’s just too much risk not to,” meaning the cost of being short on chips is now higher than the cost of renting extra ones. (cnbc.com) That is what this deal really captures: in 2026, the scarce thing in artificial intelligence is not ideas, and it is not even talent alone. It is reliable access to enough Nvidia-powered machines, in enough buildings, for long enough that a company like Meta can promise new artificial intelligence features years before its own construction catches up. (cnbc.com) (coreweave.com)
