Powering AIThe U.S. Data Center Buildout Report

Why the AI buildout is now a power problem

For most of the last decade the binding constraint on large-scale computing was silicon. It isn't any more. What now determines where — and how fast — AI capacity gets built is electricity: how much a region has, how quickly new supply can be connected, and whether the transmission exists to move it.

How much power the buildout actually needs

The forecasts are converging — upward. LBNL's 2025 Update (June 2026) found data centers used 4.7% of U.S. electricity in 2024 and projects about 12% by 2030 (a 9.5–15.3% range), requiring roughly 148 GW of grid connections. EPRI's 2026 analysis lands in the same place: 9–17% by 2030, about 60% above its own 2024 estimate. The leading 2026 studies now span roughly 380–840 TWh a year in 2030, about a twofold spread. We still publish the range rather than pick a winner; every figure here traces to a public primary source and is documented in the methodology.

The bottleneck is the queue, not the power plants

The U.S. is not short of proposed generation. It is short of connected generation. The median wait to plug a new project into the grid has more than doubled, to about 5 years (61 months for projects built in 2025). And most queued capacity is never built at all: of everything that requested interconnection between 2000 and 2020, just 13% is operating, while about 75% was withdrawn. The queue behaves as an attrition filter, not a build pipeline. Compute scales in months; interconnection scales in years. Read the full breakdown →

Demand lands on specific grids

Supply is spread across 2,459 utility-scale plants, but demand concentrates. Northern Virginia alone operates about 4,500 MW — more data-center capacity than any market on Earth. In Texas, roughly 474 GW sits in the ERCOT large-load queue, around 90% of it data centers, and under 1% of it is actually drawing power. Those are the grids where the collision happens first. See all 16 regional deep-dives →

What operators do while they wait

Rather than wait out the queue, developers increasingly route around it — behind-the-meter gas, nuclear power purchase agreements and restarts, and small modular reactors. The queue's composition is shifting to match: total volume fell about 10% in 2025 as withdrawals outpaced new requests, even as natural gas requests surged 86% to 253 GW. How the workarounds actually work → And there is a second constraint behind the first — cooling water, much of it demanded in already high-water-stress regions.

New to the terminology? The glossary defines interconnection queue, behind-the-meter, PUE, WUE, curtailment and more — in plain English, with sources.