Power is the AI bottleneck
AI growth is bumping up against the electricity grid: analysts say that after years of GPU shortages, the real constraint for scaling large models is where the power will come from and who pays for it. That shift turns AI deployment into an energy and policy problem as much as a technical one, with implications for data‑centre siting and cost modelling. (apnnews.com)
For two years, the story in artificial intelligence was a chip shortage. In 2025, the International Energy Agency said the next limit is electricity: global power use from data centres is projected to rise from about 460 terawatt-hours in 2024 to more than 1,000 terawatt-hours in 2030. (iea.org) A data centre is a warehouse full of computers, and those computers do two jobs for artificial intelligence: training models and answering user requests. The International Energy Agency says both jobs happen mainly inside data centres packed with servers, networking gear, and cooling equipment. (iea.org) The power draw is not just from the chips doing math. Every rack also needs fans, pumps, and chillers, because advanced processors turn electricity into heat the way a space heater does. (iea.org) That is why the constraint shifted from “Can you buy enough graphics processing units?” to “Can you get enough megawatts at one address?” The International Energy Agency says affordable, reliable electricity supply will be a crucial determinant of which countries and regions capture artificial intelligence growth. (iea.org) The United States is where this shows up first. The International Energy Agency said in April 2025 that American data centres are on course to account for almost half of US electricity-demand growth between 2024 and 2030. (iea.org) By 2030, the agency expects the United States to use more electricity for processing data than for making aluminium, steel, cement, and chemicals combined. That comparison turns artificial intelligence from a software story into an industrial-load story. (iea.org) The bottleneck is not only generation. In February 2026, the International Energy Agency said grid capacity itself is becoming a critical bottleneck, with record connection queues slowing new demand, new power plants, and storage projects in many regions. (iea.org) That creates a timing mismatch. The International Energy Agency wrote in October 2024 that data centres can be built quickly, while transmission lines, substations, and new generation often take years to permit and connect. (iea.org) So the question for an artificial intelligence company is no longer just rent, land, and chips. It is whether a utility can deliver a large block of power fast enough, and whether the developer pays for upgrades to wires, transformers, and backup equipment. (iea.org) The energy mix matters too. The International Energy Agency expects renewables and natural gas to lead new electricity supply for data centres in key markets, because those sources are widely available and often cheaper to add than alternatives. (iea.org) That changes where new data centres get built. Places with spare grid capacity, faster interconnection, and easier access to generation become more attractive than famous tech hubs with crowded networks and long waits for power. (iea.org) It also changes the math inside artificial intelligence itself. If electricity and grid upgrades become the scarce inputs, then model deployment is no longer priced like pure computing; it starts to look like a utility-backed infrastructure project with power contracts attached. (iea.org)
