For most of the 2000s and 2010s, US electricity demand barely moved. Efficiency gains in lighting, appliances and industrial processes offset growth in computing, and utilities planned for flat load. That assumption broke in 2023, when Microsoft, Google and Meta began disclosing that their data center energy use was climbing at double-digit annual rates driven by AI. By 2026, the gap between what operators want to build and what the grid can deliver has become the single most cited constraint on AI infrastructure expansion.

The Department of Energy's Lawrence Berkeley National Laboratory published analysis in 2024 estimating that data centers consumed between 4% and 6% of US electricity, and projected that share could reach 8% or more by 2030 under high-growth scenarios. The Electric Power Research Institute issued comparable figures, placing 2023 data center consumption at roughly 4% and noting that AI workloads — which draw power continuously during training and inference — are far more energy-dense than the web-serving workloads they replace.

Why it matters

The constraint matters because it translates directly into cost and schedule risk for every company building AI products. If a model developer cannot secure power for a new training cluster, the cluster does not get built on time, and the roadmap slips. Microsoft, Google and Amazon have all disclosed in earnings calls and SEC filings that power availability is now a gating factor on data center capacity expansion, alongside GPU supply.

It also matters beyond the technology sector. The same transmission and generation capacity that AI data centers want is needed for electrification of transport and building heating, and for retiring coal and gas plants. Regulators in PJM, the grid operator covering the mid-Atlantic and parts of the Midwest, have warned that reserve margins are tightening as older generation retires faster than new generation and transmission can replace it.

How it works

The bottleneck has three layers. First, generation: new natural gas plants take three to five years to build and permit, and utility-scale solar plus battery storage takes two to three years, so neither can respond to a sudden demand spike inside a single budget cycle. Nuclear is the longest lead item of all, with the only recent US reactor completions at Vogtle in Georgia running more than a decade.

Second, transmission: the US grid was built to move power within utility territories, not to move large volumes between regions. Building a high-voltage line across state lines requires permits from multiple agencies and routinely takes seven to ten years. The Department of Energy's 2024 National Transmission Needs Study identified dozens of corridors where congestion already costs consumers billions annually.

Third, components: large power transformers, the hardware that steps voltage down at substations, are produced by a small number of global manufacturers. Lead times exceeded 100 weeks in 2024, according to the Department of Energy's supply chain review, meaning a utility that orders a transformer today may not install it until 2028 or later. The same bottleneck applies to high-voltage cable and switchgear.

Evidence

PJM Interconnection, the grid operator serving the region that includes Northern Virginia's data center corridor, reported in its 2024 load forecast that summer peak demand could rise by nearly 40% by 2039, driven substantially by data centers, and that the pace of new generation interconnection requests had overwhelmed its queue. ERCOT, the Texas grid operator, issued reliability warnings during summer 2024 heat and has flagged rising industrial load as a planning factor.

Microsoft, Google and Meta have each signed power purchase agreements for nuclear capacity — Microsoft with Constellation Energy to restart the Three Mile Island Unit 1 reactor, Google with Kairos Power for small modular reactors, and Meta with Constellation for the Clinton nuclear plant in Illinois — which is itself evidence that conventional grid procurement cannot meet their timelines. The Department of Energy's Lawrence Berkeley National Laboratory and the Electric Power Research Institute have published the demand projections cited above, and the DOE's Grid Deployment Office maintains the transmission needs study and supply chain reviews that document the component and permitting constraints.

What happens next

The most likely near-term outcome is continued tension between AI infrastructure build-out and grid capacity, with two visible responses. Hyperscalers will increasingly sign behind-the-meter generation deals — on-site gas turbines, solar plus storage, and eventually small modular reactors — to bypass transmission queues entirely, as Microsoft's Three Mile Island agreement illustrates. Regulators and grid operators will face pressure to accelerate interconnection reform and transmission permitting, though both move on multi-year timelines.

For technology buyers, the practical consequence is that inference capacity may become locally constrained: a company that needs GPU capacity in a specific region may find it unavailable not because the chips are absent but because the building cannot draw enough power to run them. Expect more disclosure of power as a risk factor in technology company filings, and expect state-level competition for data center investment to increasingly turn on who can offer electricity on a faster timeline.