Meta doubles down on inference
Meta has expanded a long‑term partnership with CoreWeave in a deal framed as a $21 billion commitment through 2032 to secure inference capacity rather than just training rigs. That bet signals the industry shifting to recurring costs of serving models at scale, not one‑off training investments. (markets.financialcontent.com)
Meta is locking up artificial intelligence computing power years in advance, not for one giant training run but for the daily job of answering billions of prompts after the model is already built. CoreWeave said on April 9 that Meta expanded its agreement to about $21 billion through December 2032, and the company said the capacity is meant to scale inference workloads. (investors.coreweave.com) That word, inference, is the part people usually miss. Training is like teaching a chef a new recipe once, while inference is every plate that chef serves after the restaurant opens. (investors.coreweave.com) Meta is not replacing its own data centers with rented machines. CoreWeave chief executive Michael Intrator told CNBC the two companies are doing both, and Meta called the contract part of a “portfolio-based approach” to infrastructure. (cnbc.com) The timing shows how fast the bill is growing. CNBC reported that this new spending runs from 2027 to 2032 and sits on top of an earlier $14.2 billion arrangement, which means Meta now has roughly $35 billion committed with CoreWeave alone. (cnbc.com) CoreWeave is basically a specialist landlord for artificial intelligence chips. Its data centers are packed with Nvidia graphics processing units, and the new Meta deal includes some of the first deployments of Nvidia’s Vera Rubin platform across multiple locations. (cnbc.com) (investors.coreweave.com) Those multiple locations matter because inference breaks if capacity shows up late or goes down at the wrong moment. CoreWeave said the distributed setup is meant to improve performance, resilience, and scalability for Meta’s live artificial intelligence operations. (investors.coreweave.com) Meta can afford to make bets this large because its advertising machine is still enormous. In its January 28, 2026 results, Meta reported $200.97 billion in 2025 revenue, $72.22 billion in 2025 capital spending, and a 2026 capital spending plan of $115 billion to $135 billion. (investor.atmeta.com) CoreWeave needs contracts like this because the business is growing at breakneck speed and burning huge amounts of cash to build capacity. CoreWeave said it reached $5.13 billion in 2025 revenue, carried $66.8 billion in revenue backlog at year-end, and still posted a $1.17 billion net loss for the year. (investors.coreweave.com) The deeper shift is that artificial intelligence costs are starting to look less like a one-time factory build and more like an electric bill that never stops. If Meta expects people to use assistants, search, image tools, and recommendation systems every hour of every day, then serving those responses becomes a recurring infrastructure expense, and this contract prices that future in through December 2032. (investors.coreweave.com) (cnbc.com)
