Compute and power are strategic bottlenecks

Meta just added another $21 billion of artificial intelligence cloud spending with CoreWeave, on top of an earlier $14.2 billion arrangement, and the new contract runs through 2032. That is not how companies buy a cheap, abundant resource; that is how they lock up scarce capacity years in advance. (coreweave.com) (cnbc.com) CoreWeave said the extra capacity is for Meta’s inference workloads, which means the part of artificial intelligence that answers users after a model is already trained. Training is like building the engine once, while inference is like running a taxi fleet every minute of every day. (coreweave.com) That shift changes the bottleneck. A training run can be scheduled like a movie premiere, but inference behaves more like rush-hour traffic, where millions of requests arrive at once and the cloud provider has to keep enough chips live to avoid delays. (coreweave.com) The hardware at the center of these deals is still mostly Nvidia gear. CoreWeave said Meta’s new buildout will include some of the first deployments of Nvidia’s Vera Rubin platform, and Nvidia said in March that Vera Rubin had entered full production for large “artificial intelligence factories.” (coreweave.com) (nvidianews.nvidia.com) Nvidia’s grip shows up in the numbers. The company reported $62.3 billion in quarterly data center revenue for the quarter ending January 25, 2026, and $215.9 billion in full-year revenue, which means the chip supplier is still taking a huge share of the money flowing into artificial intelligence infrastructure. (nvidianews.nvidia.com) The other scarce input is electricity. Meta broke ground in February on a 1 gigawatt data center campus in Lebanon, Indiana, and a gigawatt is the scale of a utility asset, not a normal corporate office expansion. (about.fb.com) Meta is also trying not to depend on one chip vendor. In February it announced a long-term Nvidia partnership for training and inference, and in the same month it announced an Advanced Micro Devices deal for up to 6 gigawatts of Instinct graphics processing unit capacity as part of what it called a portfolio-based infrastructure approach. (about.fb.com 1) (about.fb.com 2) CoreWeave’s financing shows how far this has gone. The company closed an $8.5 billion graphics processing unit-backed financing facility on March 31, using chips and contracted demand as collateral, which is closer to project finance for power plants than old-style cloud renting. (investors.coreweave.com) Once chips and power are both constrained, software teams cannot act like compute is infinite. Nvidia’s January Rubin launch said the platform was designed to cut inference token cost and reduce the number of graphics processing units needed for some model training jobs, which is why companies are suddenly obsessed with routing simple requests to smaller models and saving the biggest systems for the hardest prompts. (investor.nvidia.com) That is why tiered inference and caching have moved from engineering tricks to budget controls. If a company can answer repeated questions from storage instead of recomputing them on premium chips, it saves both money and scarce capacity in the same way an airline saves gates by turning planes faster. (nvidianews.nvidia.com) (coreweave.com) The headline is not just that artificial intelligence is expensive. The headline is that in April 2026 the limiting factors are now specific physical things — Nvidia systems, signed cloud contracts, and grid-scale power — and the companies moving first are buying all three years before they need them. (coreweave.com) (nvidianews.nvidia.com) (about.fb.com)