AI stymied by power
Access to electricity and grid capacity—not chips—is emerging as the main bottleneck for the next wave of AI data‑centre builds, slowing planned expansion even where compute demand is high. Energy Intelligence and reporting from PBS show developers can’t energise many projects fast enough, provoking local backlash over water, noise and rising wholesale prices near proposed sites. (energyintel.com) (pbs.org)
The bottleneck for the next wave of artificial intelligence data centers is starting to look less like computer chips and more like whether a utility can deliver hundreds of megawatts of power to one patch of land on time. Energy Intelligence reported on April 9 that grid constraints, high connection costs, and slow new generation buildouts are now cutting into demand forecasts for new United States projects. (energyintel.com) A data center is basically a warehouse full of computers, and the newest artificial intelligence ones draw power more like a steel mill than a normal office park. The International Energy Agency said in April 2025 that electricity demand from data centers worldwide could more than double to 945 terawatt-hours by 2030. (energyintel.com) The problem is not just making enough electricity over a year. It is getting enough power to the exact substation, transmission line, and transformer a project needs, in the exact month the developer wants to switch it on. (epri.com) Utilities are getting flooded with requests they do not fully trust yet, because one proposed campus can ask for a block of power so large it reshapes a regional forecast. The Electric Power Research Institute said its 2024 utility survey found shared problems in processing data-center service requests and folding them into load planning. (epri.com) That is why projects can look approved on paper and still sit for years waiting for an energized connection. Energy Intelligence said developers are running into the physical limits of the grid and the long timelines for adding both new generation and new wires. (energyintel.com) The scale is already visible in the numbers. Lawrence Berkeley National Laboratory said United States data centers used about 4.4 percent of national electricity in 2023, and its report projects that share could rise to between 6.7 percent and 12 percent by 2028. (lbl.gov) Some states are much more exposed than the national average because the industry clusters in a few places with tax breaks, fiber links, and available land. The Electric Power Research Institute said Virginia already gets more than 20 percent of its electricity demand from data centers, and in its medium scenario seven more states could cross that threshold by 2030. (epri.com) That concentration is turning a business story into a local political fight. An Associated Press report published by PBS on April 9 said lawmakers in states with fast data-center growth are weighing new rules as residents complain about water use, diesel backup generators, constant fan noise, and pressure on clean-energy targets. (pbs.org) Electricity prices are part of the backlash too, because big new loads can force utilities to buy more power, build more infrastructure, and recover those costs from customers. PBS reported in September 2025 that rising artificial-intelligence electricity demand is straining grids and feeding higher bills in some regions. (pbs.org) So the artificial intelligence buildout is starting to run into the oldest constraint in industry: you can order more servers faster than you can build a power plant, a transmission line, or a substation. In 2026, the race is shifting from who can buy the most chips to who can secure real electricity at a real site on a real date. (energyintel.com)
