Meta’s $21B CoreWeave Bet

Meta just chose to rent a giant slice of someone else’s artificial intelligence machine room through December 2032 instead of waiting for spare capacity to show up when it needs it. CoreWeave said on April 9 that the expanded agreement is worth about $21 billion and is aimed at Meta’s artificial intelligence inference workloads. (coreweave.com) Inference is the part where a trained model answers your question, writes the caption, or ranks the post in front of you, and it has to happen fast every time a user taps a screen. That is different from training, which is the one-time giant cram session, and Meta said this CoreWeave deal is for the day-to-day serving side of the business. (coreweave.com) CoreWeave is not a consumer internet giant with a social network or phone business. It is a cloud provider built around renting out graphics processing units, the specialized chips that do the heavy lifting for artificial intelligence, and it has turned that niche into contracts measured in the tens of billions. (coreweave.com) (reuters.com) This is not the first check Meta has written to CoreWeave. Reuters reported on April 9 that the new $21 billion agreement deepens a partnership that already included a deal worth up to $14.2 billion announced in September 2025 for capacity running through December 2031, with an option into 2032. (reuters.com) (finance.yahoo.com) The reason companies do this is simple: the spot market is like trying to hail a taxi in a thunderstorm, while a long-term contract is like leasing part of the fleet. If Meta expects billions of searches, recommendations, ad decisions, and assistant replies to hit its systems every day, it cannot gamble on spare chips being available at the exact hour demand spikes. (reuters.com) (coreweave.com) Meta is making the same bet inside its own walls too. In February, the company said it was breaking ground on a 1 gigawatt data center campus in Lebanon, Indiana, a project it described as more than $10 billion and one of its largest infrastructure investments to date. (about.fb.com) A 1 gigawatt campus is utility-scale power, not a normal server building, and Meta has been lining up the rest of the supply chain around it. Since January, the company has announced infrastructure moves involving nuclear power, glass from Corning, chips from Advanced Micro Devices, and systems from NVIDIA, which shows that compute is now constrained by power, buildings, networking, and components all at once. (about.fb.com 1) (about.fb.com 2) That is why CoreWeave has become valuable so quickly. The company said it listed on Nasdaq in March 2025, and by March 2026 it had already closed an $8.5 billion financing facility tied to its graphics processing unit fleet, which is the kind of borrowing structure you build only when customers are willing to sign long contracts against future capacity. (finance.yahoo.com) (investors.coreweave.com) Reuters tied Meta’s urgency to the competitive pressure around its artificial intelligence products after an underwhelming model release last year. When a company feels behind, buying guaranteed capacity through 2032 is less about shaving costs and more about making sure the next model actually has somewhere to run. (reuters.com) The bigger shift is that cloud capacity for artificial intelligence is starting to look less like ordinary computing and more like reserved industrial infrastructure. Oil refiners lock up pipelines, airlines lock up aircraft, and now internet companies are locking up chip-filled data centers years in advance. (coreweave.com) (about.fb.com)