AI compute is consolidating
Access to frontier AI compute is increasingly concentrated in a handful of hyperscalers, changing who can run the biggest models. Meta expanded its CoreWeave contract by $21 billion — bringing that relationship to $35 billion through 2032 — a move that locks up large-scale capacity for its AI work. At the same time analysts note Google now controls the largest share of global AI compute, which shifts infrastructure decisions from pure architecture to platform access and pricing power. (thenextweb.com ) (networkworld.com)
Meta just agreed to spend another $21 billion with CoreWeave for artificial intelligence cloud capacity, pushing its total commitment to about $35 billion through December 2032. The deal adds to a $14.2 billion agreement signed in September 2025, so one customer is now reserving years of high-end chip time in advance. (cnbc.com) (sec.gov) CoreWeave is not a consumer app or a chip designer. It is a cloud company that rents out huge clusters of Nvidia graphics processors, which are the specialized chips most companies use to train large artificial intelligence models. (thenextweb.com) (sec.gov) That changes what “buying compute” looks like. Instead of shopping for servers one quarter at a time, companies like Meta are locking up capacity years ahead, the way an airline might reserve gates at a crowded airport before everyone else does. (cnbc.com) (bloomberg.com) The timing matters because the biggest models now need enormous training runs, and the scarce part is often not the software but access to enough chips, power, and data center space at the same time. CoreWeave said the expanded Meta deal includes early deployments of Nvidia’s next-generation Vera Rubin systems, which are not even broadly available yet. (thenextweb.com) (thestreet.com) At the same time, researchers at Epoch AI estimate that more than 60% of global artificial intelligence compute is now owned by United States hyperscalers, and Google alone holds about one quarter of worldwide cumulative capacity as of the fourth quarter of 2025. That makes Google the single largest owner of artificial intelligence compute by their count. (epoch.ai) (networkworld.com) Google got there differently from CoreWeave. While CoreWeave mainly rents Nvidia chips, Google built its own Tensor Processing Units, which are custom chips designed specifically for artificial intelligence workloads and used across Google’s internal systems and cloud business. (networkworld.com) (cnbc.com) That means the biggest artificial intelligence companies are no longer all competing on the same hardware market. Some rely on Nvidia supply through clouds like CoreWeave, while Google can steer customers onto its own chips and its own cloud, which gives it more control over pricing and availability. (networkworld.com) (epoch.ai) The result is a smaller map than it looks from the outside. Meta is wealthy enough to build its own infrastructure and still reserve tens of billions of dollars of outside capacity, while smaller labs and startups have to compete for whatever high-end compute is left on the market. (cnbc.com) (networkworld.com) So the bottleneck is shifting from “who has the best model design” to “who can actually get the machines.” In 2026, frontier artificial intelligence is starting to look less like open cloud computing and more like a few giant landlords controlling the most desirable land. (epoch.ai) (thenextweb.com)
