Analysis · LBNL · EIA · EPRI · IEA · DOE

The grid bottleneck: can the U.S. power the AI data-center buildout through 2030?

There is no shortage of power plants waiting to be built. There is a shortage of ways to connect them. That distinction is the whole story.

A Gridlas analysis · built from public data · figures 2023–2026, primary sources cited throughout

The debate over whether the U.S. can power AI usually gets framed as a generation problem — do we have enough electricity? The data says the real constraint sits one layer down, in the interconnection queue: the multi-year line every new power plant and battery must wait in before it can plug into the grid. Roughly 2,290 GW of capacity is queued today — far more than enough — but most of it will never get built, and what does will arrive years after the demand. Here is what the primary sources actually show.

1. Demand is real — but the forecasts don't agree

Every credible forecaster agrees U.S. electricity demand is climbing after two flat decades, and that data centers are the driver. The EIA calls it the strongest four-year demand growth since 2000 — the first four straight years of growth since 2007 — projecting +1% in 2026 and +3% in 2027. Data centers used about 176 TWh (4.4% of U.S. electricity) in 2023, already more than double their 2018 level.

Where forecasters diverge is how high, how fast — and the spread is enormous. That disagreement, not any single number, is the honest headline.

0% 5% 10% 15% today ~4–5% LBNL / DOE by 2028 LBNL / DOE (by 2028): 6.7–12.0% of U.S. electricity 6.7–12% EPRI — 2024 estimate by 2030 EPRI — 2024 estimate (by 2030): 4.6–9.1% of U.S. electricity 4.6–9.1% EPRI — 2026 estimate by 2030 EPRI — 2026 estimate (by 2030): 9.0–17.0% of U.S. electricity 9–17%
Data centers as a share of U.S. electricity — competing forecasts. Note the horizon differs (LBNL to 2028; EPRI to 2030) and EPRI revised its own 2030 range up ~60% in under two years. Sources: LBNL/DOE (2024); EPRI "Powering Intelligence" (2024 & 2026).
How to read this — These are not apples-to-apples. LBNL's 6.7–12% figure is a 2028 projection (bounded by chip-shipment forecasts), not 2030, and uses equipment-shipment data; EPRI's is a 2030 projection built from commercial project pipelines. Across all models, published 2030 U.S. estimates span roughly 200 to over 1,050 TWh/year — a 5× spread. Treat every point estimate as scenario-dependent.
6.7–12%
of U.S. electricity by 2028 (LBNL/DOE) = 74–132 GW
9–17%
by 2030 (EPRI 2026) — up ~60% from its 2024 estimate
of all U.S. demand growth to 2030 is data centers (IEA)

2. The bottleneck is the queue, not the megawatts

Here is the counterintuitive core. As of end-2024, about 2,290 GW of generation and storage — some 10,300 projects — was seeking to connect to the U.S. grid, the queues covering ~98% of national capacity. That is multiples of what the buildout needs. The problem is throughput: the median time from interconnection request to switched-on has more than doubled, from under two years for projects built in 2000–07 to about five years for those built in 2023.

0 yr 2 yr 4 yr 6 yr 2000–07: <2 yrs <2 yrs built 2000–07 2015: ~3 yrs ~3 yrs built 2015 2023: ~5 yrs ~5 yrs built 2023
Median duration from interconnection request to commercial operation, by build cohort (regions with available data). Source: LBNL "Queued Up" (2024 & 2025 editions), Berkeley Lab.

And the wait is only half of it. Most queued capacity is never built at all. Of everything that requested interconnection between 2000 and 2019, only 13% had been built by end-2024 — while 77% was withdrawn. The queue is an attrition filter, not a build pipeline.

Withdrawn: 77% 77% Built: 13% 13% Still in queue: 10% 10% Withdrawn Built Still in queue
Fate of generation + storage capacity that requested U.S. interconnection, 2000–2019 (status as of end-2024). Source: LBNL "Queued Up" 2025 Edition.
There is no capacity shortage in the queue. There is a throughput shortage in the process. Compute scales in months; interconnection scales in years.

3. The queue is quietly changing shape

For the first time in a decade, the queue actually shrank — total volume fell about 12% in 2024 as withdrawals outpaced new requests, with solar (−12%), storage (−13%) and wind (−26%) all down. One resource ran the other way: natural gas surged 72% to 136 GW. The queue is still ~95% clean by volume, but gas is the only major category growing — an early signal that developers are chasing firm, always-on power they can count on connecting.

4. What's actually stepping in — and what's slowing it

On the supply side, the IEA's global outlook (to 2035) has renewables covering roughly half of data-center demand growth, with natural gas (notably in the U.S., ~+175 TWh) and nuclear each adding about the same, and the first small modular reactors arriving around 2030. But every one of those has a lead-time problem:

The fixes that need no new steel are the ones that work fastest: grid-enhancing technologies and advanced-conductor reconductoring can lift the capacity of lines already in the ground (per the U.S. DOE and GridLab), and making large new loads flexible — willing to curtail briefly at peaks — can free meaningful headroom on the existing grid. Those are the levers with the shortest path to a connected megawatt.

5. Where the load lands decides who feels it

National averages hide the real risk, because the load concentrates. About half of U.S. data centers under development sit in pre-existing large clusters (IEA), stacking new demand onto grids already carrying it. Virginia is the extreme case — the only state where data centers already exceed 20% of electricity use, a share that could reach 39–57% by 2030 (EPRI) — and on EPRI's medium scenario, seven more states could cross 20% by decade's end. That geography is exactly what the regional maps track, market by market.

The question isn't whether the U.S. has enough power. It's whether the right wires, in the right places, get built before the GPUs show up.

The full picture, mapped. The Gridlas report puts these forecasts, the queue data, and five regional deep-dives onto the grid — with high-res maps and the underlying dataset (CSV/GeoJSON), built entirely from public data.

Get the report →
On the numbers — This piece is built only from primary and authoritative sources, and every headline figure was cross-checked before publishing. Where forecasts disagree we show the range rather than pick a winner. Two caveats worth repeating: the LBNL 6.7–12% figure is a 2028 projection, not 2030; and the IEA generation-mix and "20% at risk" figures are global, to 2035, not a U.S.-through-2030 breakdown.
Sources: See also the full methodology & sources and the regional breakdowns.
Prefer the map view? The visual analysis plots this over the U.S. power grid, and the regional pages break it down market by market.

Gridlas · independent & unaffiliated · built from public data.