AI's next bottleneck may be electricity, not chips.
US data centers consumed about 4.4% of the nation’s electricity in 2023, and that share could reach 6.7% to 12% by 2028. For cloud providers, colocation operators and enterprise AI buyers, the challenge is no longer just securing accelerators or cloud capacity. It is whether utilities can deliver firm power to the regions, substations and transmission corridors where AI campuses are planned.
AI power demand is growing faster than local grids can adapt
The DOE-backed report projected US data center electricity use could climb from about 176 terawatt-hours in 2023 to between 325 and 580 terawatt-hours by 2028, nearly tripling the sector’s 2023 share at the high end
That national figure can obscure the more urgent local problem. The Secretary of Energy Advisory Board said hyperscale connection requests of 300 to 1,000 megawatts or more, often with lead times of one to three years, are stretching local grids. Transmission infrastructure typically takes longer to permit, finance and build.
In transmission-constrained areas, the gap can be narrow but consequential. One SEAB example describes a utility that can serve a data center for about 350 days a year, while the remaining 15 days require backup generation, storage or another reliability solution.
For AI infrastructure planners, that creates grid-access risk. Regions with constrained transmission may not be able to serve new hyperscale facilities on the timelines that AI deployment roadmaps assume. Recent AI compute deals show that power, cooling, networking and data center space now have to arrive together.
Flexible AI workloads can help, but firm power is still the gap
Load flexibility is one near-term option. SEAB said some AI inference work can be routed across regions based on local grid load and the amount of renewable power on the grid, with limited user impact when latency is not critical. This could shift some workloads away from stressed grids during peak demand.
Flexibility can help manage grid stress, but it does not eliminate the need for firm power. SEAB said stakeholders generally saw new natural gas capacity, along with solar, wind and batteries, as the main options available today to maintain reliability. That complicates clean-energy planning as the AI infrastructure race expands from chips into data centers, networking and power.
Nuclear power is becoming part of the longer-term hedge. The Associated Press reported that Constellation Energy aims to bring the 835-megawatt reactor at Three Mile Island back online in 2027 under a plan tied to a 20-year power supply deal with Microsoft. Google, Kairos Power and the Tennessee Valley Authority said Hermes 2 in Oak Ridge, Tennessee, is scheduled to begin operations in 2030 and deliver up to 50 megawatts to the TVA grid. But neither project erases the near-term grid gap facing AI campuses seeking power now.
Regulators are also weighing who should pay for grid upgrades tied to new AI campuses. SEAB recommended tariffs that would have data centers share or cover those costs, and it warned that some companies were considering sites outside the United States if domestic power timelines remained unworkable.
For enterprise AI teams, power availability now belongs in infrastructure planning alongside model performance, accelerator supply and workload placement.
Also read: As AI compute demand rises, China’s subsea data center project shows how operators are experimenting with cooling, siting and renewable power.