Meta ups CoreWeave pact

Meta just agreed to spend about $21 billion more with CoreWeave, a cloud company that rents out giant clusters of Nvidia chips, and the contract runs through December 2032. That pushes Meta’s total CoreWeave commitment to roughly $35 billion, which is an unusually long reservation for computing power in a market where the best machines sell out fast. (coreweave.com) CoreWeave is not a normal cloud landlord like Amazon Web Services or Microsoft Azure. It built its business around graphics processing units, which are the chips used to train and run artificial intelligence models, and it became one of Nvidia’s earliest specialist partners. (coreweave.com, nvidianews.nvidia.com) This deal is mostly about inference, which is the step where an artificial intelligence model answers your question after the training is done. Training is like cramming for the exam, while inference is sitting for the exam millions of times a day for real users. (coreweave.com) Meta is buying dedicated capacity, which means CoreWeave sets aside machines for Meta instead of tossing them into a shared pool for whoever pays that hour. That matters because Meta runs consumer products with billions of users, and a slowdown in chip supply can delay model launches just as surely as a software bug can. (businesswire.com, cnbc.com) The hardware in the contract includes some of the first deployments of Nvidia’s Vera Rubin platform, which is Nvidia’s next big artificial intelligence system after Blackwell. Nvidia said CoreWeave would be among the first companies to offer Rubin, so Meta is effectively pre-ordering seats on a plane that has barely started boarding. (coreweave.com, nvidianews.nvidia.com) Meta could build more of this itself, and it is already spending at a historic pace on its own data centers. On its January 2026 earnings call, Meta said 2026 capital spending would be between $115 billion and $135 billion, up from $72.22 billion in 2025, with much of the increase tied to artificial intelligence infrastructure. (datacenterdynamics.com) So why rent from CoreWeave anyway. Because waiting for your own building to finish is slower than leasing finished space, and in artificial intelligence the company that gets more chips earlier can ship better models earlier and collect more usage before rivals catch up. (cnbc.com, theinformation.com) This is also a huge vote of confidence in CoreWeave’s business model. CoreWeave said the new Meta agreement sits on top of an earlier arrangement, and reporting on the prior deal put that earlier commitment at $14.2 billion, which means one customer now accounts for tens of billions of dollars of future demand. (coreweave.com, cnbc.com) CoreWeave needs that demand because supplying these chip clusters is brutally expensive. After announcing the Meta expansion, CoreWeave also moved to raise billions more in debt, showing how this market works now: lock in long contracts first, then borrow against them to buy the machines. (datacenterdynamics.com, finance.yahoo.com) The bigger shift is that computing power is no longer just an operating expense like electricity or office rent. By signing through 2032 for early Vera Rubin systems, Meta is treating chip access more like a pipeline or a port: a piece of infrastructure you secure years ahead because the companies that miss the reservation may miss the market. (coreweave.com, nvidianews.nvidia.com)