Meta's huge cloud bets

Meta just agreed to buy about $21 billion of artificial intelligence cloud capacity from CoreWeave through December 2032, even though Meta is also spending tens of billions on its own data centers. The surprise is not the size alone; it is that one of the world’s richest infrastructure builders still needs outside help to get enough computing power. (reuters.com, coreweave.com) This new contract sits on top of an earlier Meta-CoreWeave arrangement worth about $14.2 billion, so the relationship is already measured in the mid-$30 billions. CNBC reported the new agreement runs from 2027 to 2032, which means Meta is reserving capacity years before many of those chips are even switched on. (cnbc.com, coreweave.com) CoreWeave is not a general-purpose cloud company in the old Amazon Web Services mold; it is a specialist that packs data centers with Nvidia graphics processing units, the chips used to train and run large artificial intelligence models. Renting from CoreWeave is like booking extra factory lines from a contractor when your own factory is full. (coreweave.com, reuters.com) Meta’s own spending plans show why that rental market exists. In its January 29, 2026 results, Meta said 2026 capital expenditures would be $115 billion to $135 billion, up from $72.22 billion in 2025, and its finance team said much of the expense growth would come from infrastructure, including third-party cloud spending. (atmeta.com, q4cdn.com) That tells you something important about the artificial intelligence race in 2026: owning land, concrete, and servers is not enough. The bottleneck is getting the right chips, power, cooling, networking, and software in the same place at the same time, and outside providers can sometimes do that faster than an in-house build. (atmeta.com, reuters.com) CoreWeave said some of the new Meta capacity will include early deployments of Nvidia’s Vera Rubin platform across multiple locations. In plain English, Meta is not just buying generic server time; it is lining up access to the next wave of premium artificial intelligence hardware before demand gets even tighter. (coreweave.com) The contract is aimed at inference workloads, which is the stage where a trained model answers real user requests instead of learning from data. Training is like teaching a student for months; inference is the student taking millions of tests every day, and that daily serving load is what social platforms hit at giant scale. (coreweave.com) This is why external cloud deals keep getting bigger instead of disappearing as big tech builds more campuses. A company like Meta can own huge data centers and still use outside capacity as overflow, as a hedge against delays, and as a way to spread workloads across different hardware generations and locations. (atmeta.com, cnbc.com, coreweave.com) The people who become more valuable in this setup are not only chip designers and model researchers. They are the systems engineers who can decide which jobs run on which machines, keep expensive graphics processing units busy, move models between clusters, and cut waste when one hour of idle hardware can cost a fortune. (atmeta.com, coreweave.com) A few years ago, the cloud story was about renting cheap computing instead of buying servers. In 2026, the cloud story at the top end is the opposite: even companies spending up to $135 billion a year still prepay specialists for scarce artificial intelligence capacity, because waiting for their own buildings to catch up is the risk they cannot afford. (atmeta.com, reuters.com, cnbc.com)