Hyperscalers now control capacity
Hyperscalers are tightening their grip on physical capacity: posts report they now control nearly half of global data-centre capacity and could hit about two-thirds by 2031, while model labs like Anthropic are publicly pressing for large U.S. data-centre builds (reported needs above 50GW) to remain competitive. That concentration means buyers increasingly ask not just about chip speed but about guaranteed capacity, power and procurement visibility. (x.com) (x.com)
The fight in artificial intelligence is starting to look less like a chip race and more like a land-and-power race. New market data says hyperscale cloud companies now control 48 percent of global data-centre capacity, up from 41 percent a year earlier. (datacenterdynamics.com) That 48 percent is not just rented space. Synergy Research says about 60 percent of hyperscaler capacity now sits in data centres those companies built and own themselves, which gives them tighter control over expansion timing, power contracts, and equipment layouts. (datacenterdynamics.com) The old model is shrinking fast. Enterprise companies running their own server rooms accounted for 56 percent of worldwide capacity in 2018, and that share is now down to 32 percent, while non-hyperscale colocation providers sit at 20 percent. (datacenterdynamics.com) Synergy’s forecast goes further: hyperscalers could control more than two-thirds of all global data-centre capacity by 2031. That means the companies selling cloud computing are also becoming the companies with first claim on the buildings, substations, and construction pipelines behind it. (datacenterdynamics.com) This is why artificial intelligence labs are talking about electricity in the same breath as models. Anthropic said on February 11, 2026 that the United States artificial intelligence sector will need at least 50 gigawatts of capacity over the next several years, and that training one frontier model will soon require gigawatts of power. (anthropic.com) Anthropic put money behind that claim in November 2025 with a $50 billion plan to build custom data centres with Fluidstack in Texas and New York. Three days ago, it said a new Google Cloud and Broadcom partnership would place the vast majority of additional compute in the United States. (anthropic.com 1) (anthropic.com 2) The scale is easiest to see in power numbers. The Financial Times reported that all American data centres together represent about 51 gigawatts today, while S&P Global Energy estimated that new data centres could need another 44 gigawatts by 2028. (ft.com) The grid is not keeping up evenly. The same Financial Times report said only about 25 gigawatts of power capacity expected in the next three years is likely to be available for those new data centres, leaving a 19 gigawatt gap. (ft.com) When that kind of shortage shows up, “capacity” stops meaning rack space and starts meaning who can actually get steel, transformers, turbines, utility approvals, and long-term power. A smaller model company can buy the same chips on paper and still lose if a hyperscaler already locked up the site and the electricity. (ft.com) (datacenterdynamics.com) That changes how customers buy too. Anthropic’s own recent expansion with Google Cloud said the deal would bring well over one gigawatt of capacity online in 2026, which is the kind of promise buyers now want alongside price and performance. (anthropic.com) So the new pecking order is becoming physical. The winners are the companies that can guarantee a building, a power feed, and a delivery date, not just the ones with the fastest chips on a slide deck. (datacenterdynamics.com) (anthropic.com)
