Power is the AI bottleneck
The expansion of AI data centres is being limited more by electricity supply and grid capacity than by chips or models, raising complaints from local communities and policy questions about emissions and transmission. Residents near planned facilities in North Carolina are warning of higher household bills, while Brookings and industry analysts argue power planning must now sit at the centre of AI regulation and infrastructure strategy. (wunc.org) (brookings.edu) (apnnews.com).
A data center used to be a real-estate story. In 2026, it is an electricity story, because the hard part is no longer finding land or even chips but finding enough power lines, transformers, and generation to keep the servers on. (iea.org) That shift is showing up in North Carolina, where residents told public radio station WUNC they fear new artificial intelligence facilities will push household bills even higher as Duke Energy asks regulators for another rate increase. Duke says only about one-third of incoming projects are data centers, but those projects account for 80% of projected demand from new economic growth in the state. (wunc.org) The basic problem is simple: an artificial intelligence model runs inside racks of servers, and those servers turn electricity into math and heat. The more chips you pack into one building, the more power you need every second, not just over a year. (brookings.edu) That is why the bottleneck has moved from graphics processing units to megawatts. Neel Somani, who has worked in both power markets and artificial intelligence, argues the key constraint is now grid capacity, because buying chips is useless if the local utility cannot deliver the electricity. (apnnews.com) Utilities cannot solve that overnight. A new transmission line, substation, or gas plant takes years of permits, financing, and construction, so a cloud company can order servers faster than a region can build the wires to feed them. (iea.org) Brookings says this turns artificial intelligence regulation into an energy-planning problem as much as a software problem. Its April 10, 2026 discussion focused on rising electricity and water use from data centers and on the need for international governance that includes infrastructure, emissions, and resource constraints. (brookings.edu) The politics get sharper when utilities spread upgrade costs across everyone on the system. In North Carolina, ratepayers and local officials are asking whether families should help fund grid expansion for some of the richest technology companies in the world. (wunc.org) (ncuc.gov) The climate math is awkward too. The International Energy Agency said in 2025 that countries chasing artificial intelligence growth would need faster investment in generation and grids, and that renewables and natural gas were likely to supply much of the added demand in key markets. (iea.org) That means the next race in artificial intelligence is partly geographic. Companies will favor places with spare power, faster interconnection queues, and regulators willing to approve big loads, which can pull investment toward certain regions and leave others fighting over reliability and price. (apnnews.com) (brookings.edu) So the new question is not just who has the best model. It is who can secure 24-hour electricity at industrial scale without blowing up local bills, local politics, or the grid itself. (wunc.org) (iea.org)
