AI build‑out hits the grid
Data‑centre expansion for AI is bumping into power limits as developers struggle to secure substations, transmission and permitting — turning megawatts into a gating factor for growth. Industry reports and the U.S. energy outlook say rising data‑centre demand will reshape power needs through 2050 and force projects to contend with electricity access and local opposition. (energyintel.com, greentechlead.com)
A data center can be finished in 18 months and still sit half-useless if the power hookup is years away, because the real bottleneck is no longer concrete or servers but the substation and transmission line that feed the building. That is why developers now talk about megawatts the way builders talk about square feet: a campus without assured electricity is like a factory without a road. The Federal Energy Regulatory Commission opened a proceeding on February 20, 2025, because large Artificial Intelligence data centers trying to connect near power plants were raising questions about reliability and who pays the costs. The demand jump is not small. S&P Global Market Intelligence 451 Research said U.S. utility power used by data centers would reach 82.3 gigawatts in 2026, up 28% from 2025 and more than double the level from three years earlier. By 2030, that same forecast puts U.S. data-center demand at about 167 gigawatts, which is roughly the scale of adding a whole new giant industrial sector to the grid in one planning cycle. Virginia alone is expected to need 16.6 gigawatts for data centers in 2026 and more than 33 gigawatts by 2030. The U.S. Energy Information Administration’s Annual Energy Outlook 2026, released on April 8, projects server electricity use could climb to 818 billion kilowatt-hours in 2050 in its high-demand case. That is more than 16 times the 2020 level for servers alone. Once that much new load shows up, the problem moves upstream. Utilities need transformers, substations, and high-voltage transmission lines, and each of those pieces takes permits, land, specialized equipment, and years of queue time before a single watt moves. The queue is already crowded on the supply side. Lawrence Berkeley National Laboratory’s 2025 edition of Queued Up says it tracks requests to connect new power plants to the bulk transmission grid, which means every new data-center load is arriving in a system where generation and transmission projects are already lining up for access. That is why companies have started chasing workarounds like putting data centers next to existing power plants. Federal regulators stepped in because those co-location deals can shift grid costs and reliability risks onto everyone else if the rules are vague. The fight is no longer only between developers and utilities. S&P Global reported at least 188 activist groups opposing projects nationwide, and Data Center Watch counted 20 projects delayed or canceled in a three-month stretch in 2025, representing about $100 billion of investment. Residents are not usually arguing about cloud computing in the abstract. They are arguing about diesel backup generators, transmission corridors, water use, noise, land rezoning, and whether household ratepayers end up subsidizing campuses built for a handful of technology companies. So the next phase of the Artificial Intelligence boom looks less like a software story and more like a utility story. The winners will be the developers that can secure real power with signed interconnection rights, local permits, and equipment delivery dates, not just land and financing.