Power bottleneck emerges for hyperscalers, risking GPU capacity expansion

The artificial intelligence buildout is running into a simpler problem than chips: many new data centers still cannot get enough electricity on time. (ft.com) Amazon said it expects about $200 billion of capital spending in 2026, while Alphabet projected $175 billion to $185 billion and Meta projected $115 billion to $135 billion. (cnbc.com 1) (cnbc.com 2) (datacenterdynamics.com) Bloomberg reported in February that Amazon, Alphabet, Meta and Microsoft together were on track for about $650 billion of capital expenditures in 2026, largely tied to new data centers and computing gear. (bloomberg.com) A data center is a warehouse of servers, networking gear and cooling equipment. For AI, those buildings are packed with graphics processing units, or GPUs, which draw far more power per rack than older cloud workloads. (ft.com 1) (ft.com 2) That changes the bottleneck. Financial Times reported that U.S. data centers represent about 51 gigawatts of combined capacity today, and S&P Global Energy estimates another 44 gigawatts will be needed by 2028. (ft.com) Only about 25 gigawatts of power capacity coming online in the next three years is likely to be available for those new data centers, leaving a 19 gigawatt gap, according to the same S&P estimate. (ft.com) The delays are not just about generation. Semafor reported in February, citing Sightline Climate, that power constraints and shortages of transformers, switchgear and batteries could delay 30% to 50% of data-center projects in 2026. (semafor.com) That shortage is showing up in where projects get built. Datacenter Frontier reported in March that Northern Virginia, the biggest U.S. data-center market, faces grid constraints, while Texas and Georgia are gaining share because power and land are easier to secure. (datacenterfrontier.com) The federal government has started responding as if electricity, not semiconductors, is the near-term choke point. On March 4, the White House said major AI companies and hyperscalers signed a “Ratepayer Protection Pledge” to build, bring or buy the power needed for their data centers and cover related grid upgrades. (whitehouse.gov) (federalregister.gov) The market response has been to lock up powered sites wherever they can. Applied Digital said on April 23 that it signed a $7.5 billion long-term lease with an unnamed U.S. hyperscaler at its Delta Forge 1 campus. (reuters.com) The race for GPUs is not over, but the practical question has changed. A company can order chips, yet still wait years for a substation, transmission upgrade or transformer before those chips can be turned on. (ft.com) (semafor.com) That is why the next phase of the AI boom looks less like a software rollout and more like a utility buildout, with hyperscalers chasing megawatts as aggressively as they chase Nvidia allocations. (ft.com) (bloomberg.com)