AI's power problem
- AI compute demand is moving from a tech issue into an electricity and infrastructure problem for grids and planners. - An IEA review summarised by NS Energy warns about interconnection delays and speculative "phantom load" in data-centre forecasts. - That uncertainty is delaying data-centre builds, raising costs and pushing firms toward distributed power and new site strategies (nsenergybusiness.com).
Artificial intelligence is running into a power problem: the grid, not the chip supply, is becoming the hard limit on new data-centre growth. (iea.org) A data centre is a warehouse of servers, networking gear and cooling systems, and the servers alone account for about 60% of its electricity use in modern facilities. The International Energy Agency estimated data centres used about 415 terawatt-hours of electricity in 2024, or roughly 1.5% of global consumption. (iea.org) The agency said on April 10, 2025 that global electricity demand from data centres is set to more than double by 2030. In its base case, power generation serving data centres rises from 460 terawatt-hours in 2024 to more than 1,000 terawatt-hours in 2030. (iea.org) The bottleneck is the connection to the grid. The International Energy Agency said about 20% of planned data-centre projects could face delays if grid risks are not addressed, and it said interconnection waits can stretch to a decade in some regions. (iea.org) That is turning an internet business into a utility planning problem. The International Energy Agency said a typical AI-focused data centre uses as much electricity as 100,000 households, while the largest facilities under construction today can use 20 times that amount. (iea.org) Utilities are also struggling to tell which proposed projects are real. Grid Strategies said in a November 2025 review that data centres account for about 55% of forecast U.S. peak-load growth over the next five years, but utility forecasts likely overstate data-centre demand by roughly 25 gigawatts. (gridstrategiesllc.com) That gap comes from what power planners call “phantom load” — projects that ask multiple utilities for capacity, arrive with aggressive timelines, or never get built at the size first proposed. Grid Strategies said utilities’ filings point to about 90 gigawatts of added data-centre load by 2030, while market analysts suggest growth is more likely closer to 65 gigawatts. (gridstrategiesllc.com) The supply response is getting messier too. The International Energy Agency said renewables meet nearly half of additional data-centre electricity demand through 2030, but natural gas and coal together still supply more than 40% of the increase. (iea.org) Grid build-outs are already behind demand growth. The International Energy Agency said more than 2,500 gigawatts of renewable, large-load and storage projects are stalled in grid queues worldwide, leaving power systems with rising congestion and long waits for new connections. (iea.org) That is why developers are changing where and how they build. The International Energy Agency said some operators are investing in co-located renewables and onsite generation, as the race to add AI capacity starts to depend on finding power that can actually show up on time. (iea.org)
