Data‑centre limits bite AI

The bottleneck in artificial intelligence is starting to look less like chips and more like plumbing and power lines. On April 10, the International Energy Agency said electricity demand from data centers worldwide is set to more than double to 945 terawatt-hours by 2030, roughly equal to Japan’s current power use. (iea.org, spglobal.com) A data center is just a warehouse full of computers, and every one of those computers turns electricity into heat. The harder companies push artificial intelligence systems, the more those buildings need three basic things at once: megawatts of power, industrial cooling, and land with permits already in hand. (energy.gov, iea.org) The power problem starts before a single server arrives. A July 2024 United States Department of Energy report said utilities were already seeing connection requests for hyperscale facilities of 300 to 1,000 megawatts, with lead times of 1 to 3 years stretching local grids. (energy.gov) The cooling problem got worse when artificial intelligence servers became denser. Nvidia said cooling alone has historically accounted for up to 40% of a data center’s electricity use, and its newest rack-scale systems are being designed around liquid cooling because air cooling struggles at these heat levels. (blogs.nvidia.com) Water is part of that fight, because one common way to cool hot buildings is to evaporate water like a giant swamp cooler. Google says it reports annual water use for each data center site, and Microsoft said in December 2024 that its new design for artificial intelligence workloads uses zero water for cooling by switching to chip-level cooling. (datacenters.google, microsoft.com) Now the buildout is running into the real world. Bloomberg reported in February 2026 that United States data center capacity under construction fell to 5.99 gigawatts at the end of 2025 from 6.35 gigawatts a year earlier, the first drop since 2020, with permitting, zoning, and power procurement slowing projects. (bloomberg.com, datacenterdynamics.com) That is the strange part of this moment: demand is still surging while construction is slipping. CBRE said 74.3% of capacity under construction in North America was already preleased in the first half of 2025, which means much of the space is spoken for before the building is finished. (cbre.com) The queue behind those buildings is even bigger than the sites going up now. JLL said in March 2026 that more than 35 gigawatts of data center capacity was under construction in North America and 92% of the development pipeline was already precommitted, while 64% of capacity under construction had shifted into frontier markets outside the oldest hubs. (jll.com) That shift is happening because the classic data center markets are running short on easy power. The Electric Power Research Institute said last month that data centers could consume 9% to 17% of all United States electricity by 2030, and Virginia alone could see data centers rise from about 25% of state electricity use today to 39% to 57% by 2030. (epri.com) Once power gets scarce, everything else gets expensive. SemiAnalysis says its data center model is now focused on “critical information technology power capacity,” which is a blunt way of saying the limiting factor is no longer just how many graphics processors companies can buy, but how many powered and cooled buildings they can actually plug them into. (semianalysis.com) So the artificial intelligence race is turning into an infrastructure race. The winners may be the companies that can secure transformers, switchgear, water plans, permits, and grid hookups years ahead of time, because a model cannot train inside a press release. (bloomberg.com, energy.gov)