Power Is the Bottleneck
Electricity — not just chips — is emerging as the main constraint on where and how fast AI data centres can be built, as local communities and utilities confront demand spikes and the cost of new transmission and generation. Analysts and local reporting say required substations and grid upgrades can push costs onto ordinary ratepayers and change site-selection economics, effectively turning AI capacity into heavy industry with political as well as engineering trade‑offs. Vendors are responding with efficiency measures like liquid cooling, but observers warn efficiency only delays a hard infrastructure limit that will advantage the best‑capitalised firms. (wunc.org) (brookings.edu) (infotechlead.com)
A data center used to be sold like a warehouse full of computers. In 2026, it is being treated more like a steel mill: Duke Energy says data centers are only about a third of incoming big-business projects in North Carolina, but they make up 80% of projected demand from new economic growth. (wunc.org) That shift starts with one simple fact: artificial intelligence runs on electricity twice over, first to train models and then to answer prompts millions of times a day. Brookings says hyperscalers are now openly seeking more generation and more transmission as they race to build larger sites for training and inference. (brookings.edu) The national numbers are no longer small enough to hide inside normal grid growth. The United States Department of Energy said in December 2024 that data centers used about 4.4% of all United States electricity in 2023 and could reach roughly 6.7% to 12% by 2028. (energy.gov) In raw power terms, Lawrence Berkeley National Laboratory put United States data-center use at 176 terawatt-hours in 2023, up from 58 terawatt-hours in 2014. The same report estimated 325 to 580 terawatt-hours by 2028, which is why utilities are talking about substations, transmission lines, and new generation instead of just server racks. (energy.gov) (eta-publications.lbl.gov) That is where the bottleneck moves from Silicon Valley to the local grid. WUNC reports that the Electric Power Research Institute says a single data center can require as much energy in a year as an entire power plant produces, which turns site selection into a search for spare electrical capacity, not just cheap land. (wunc.org) North Carolina shows how fast this becomes political. WUNC reports there are about 100 data centers in the state already using up to 3% of statewide electricity demand, and that demand is expected to more than double by 2030 as larger artificial-intelligence facilities arrive. (wunc.org) Once a utility has to build new wires and substations for a handful of giant customers, the argument becomes who pays. North Carolina regulators spent part of late 2025 in a “large load technical conference” focused on how to serve data centers without shifting too much risk onto ordinary households and small businesses. (wunc.org) Residents are reacting before many of the biggest projects are even online. WUNC reported in January 2026 that Duke Energy Progress was seeking an average residential rate increase of 18.5%, with implementation starting in 2027 if regulators approve it. (wunc.org) The industry answer is to squeeze more computing out of each watt. InfotechLead reported on April 10 that vendors are rolling out long-duration batteries, liquid-cooling systems, and denser artificial-intelligence infrastructure so data centers can cut waste and ride through grid stress more smoothly. (infotechlead.com 1) (infotechlead.com 2) But efficiency changes the slope, not the ceiling. Brookings says future constraints from power availability and grid capacity may place hard limits on artificial-intelligence compute, which means the winners will not just be the firms with the best chips, but the firms that can secure electricity, permits, transmission access, and enough capital to wait for all four. (brookings.edu)
