Where AI infrastructure is landing — and the grid bottlenecks standing in the way. Built entirely from public data.
The bottleneck for AI isn't chips — it's electricity, and the years-long wait to connect to the grid. I wanted to see that collision spatially, so I pulled public datasets — EIA for power plants and transmission lines, Lawrence Berkeley National Laboratory's "Queued Up" for interconnection queues, and ERCOT's load reports — and rendered them into maps. Here's what stood out.
For two decades U.S. electricity demand was essentially flat — efficiency offset growth. AI broke that. Data centers used 4.7% of U.S. electricity in 2024, a share LBNL now projects will reach about 12% by 2030 (range 9.5–15.3%). The load is also concentrated: it lands on specific grids, around the clock, faster than utilities have ever had to respond.

Plotting every major cluster, one market dwarfs the rest: Northern Virginia, at roughly 4,040 MW at the end of 2025, was about 3.5× every secondary U.S. market combined — and it reached about 4,500 MW by mid-2026 (CBRE, H1 2026). Dallas–Fort Worth and Atlanta have each crossed a gigawatt; Phoenix and central Ohio are climbing fast.


No market shows the strain like Texas. ERCOT's large-load interconnection queue went from 63 GW at the end of 2024 to ~226 GW by November 2025 — roughly a 4× jump in a year, with about 77% of it data centers aiming to connect by 2030. It then more than doubled again, to ~474 GW by June 2026, about 90% of it data centers (ERCOT, July 2026).

But a queue is a wish list, not a build plan. Of that 226 GW, only about 1.8% was actually operational and drawing power — more than half hadn't even submitted enough information to begin review. By mid-2026 it was worse: under 1% of the ~474 GW queue (about 4.3 GW) was drawing power.

Why so little gets through? Part of it is supply: new power plants and storage need years to connect. Nationally, their median time from interconnection request to commercial operation has more than doubled — from under two years in the 2000s to about 4.5 years for projects reaching operation in 2024 — and about 5 years (61 months) for 2025 builds, per LBNL's 2026 edition. Data centers wait separately in utility large-load queues, where timelines vary by utility; in Virginia, Dominion is assigning connection dates out to 2031 and beyond.

Layering ~2,459 utility-scale power plants (≥100 MW, ~1,089 GW) under the demand picture makes the mismatch clear: generation blankets the country, but data-center load concentrates onto a handful of already-stressed grids. The problem isn't total capacity — it's location and timing.

Zoom back out and the national picture resolves into a short list. Five markets absorb the bulk of the operating load and nearly all of the queue pressure — each for a different reason. I broke each one down separately; the short version:

What ties them together isn't size — it's that each one is asking a single grid to absorb, in a few years, load that used to take a decade to plan for. The full regional breakdown is here.
If the bottleneck is power and time, every fix is really trying to buy one or the other. Four are moving fastest:
Which levers each market actually pulls is where the regional story gets specific: Texas leans on behind-the-meter gas, Virginia and Ohio on utility buildout and flexibility, Phoenix on solar-plus-storage. That's the thread the regional breakdowns and the full report pick up.
Demand is one story; the checkbook is another. Pull the four hyperscalers' capital expenditure straight from their 10-K filings and the scale of the bet is unmistakable: combined capex climbed from ~$125B in 2021 to ~$358B in fiscal 2025 — roughly 2.9× in four years. The companies attribute the surge overwhelmingly to AI infrastructure, and every dollar of it eventually needs somewhere to plug in.
| Fiscal year | Amazon | Alphabet | Meta | Microsoft | Total |
|---|---|---|---|---|---|
| FY2021 | $61.0B | $24.6B | $18.6B | $20.6B | $124.9B |
| FY2022 | $63.6B | $31.5B | $31.4B | $23.9B | $150.5B |
| FY2023 | $52.7B | $32.2B | $27.3B | $28.1B | $140.4B |
| FY2024 | $83.0B | $52.5B | $37.3B | $44.5B | $217.3B |
| FY2025 | $131.8B | $91.5B | $69.7B | $64.5B | $357.5B |
That is the capital force behind every map in this report. The spending is committed and still accelerating — which is exactly why the binding constraint has moved downstream, from "will they build?" to "can the grid take it?" The rest of this analysis is what happens when $358B of ambition meets a power system that scales in years.
Tools: maps and charts rendered with QGIS and Python (matplotlib) from raw public datasets. No proprietary data.
Sources:
Caveat: data-center cluster locations and capacities are approximate, compiled from public market reports. Generation/transmission/queue figures are from the government/lab sources above. Current as of late 2025; headline demand, ERCOT, wait-time and Northern Virginia figures updated October 2026.
How much electricity do U.S. data centers use?
Data centers used about 4.7% of U.S. electricity in 2024, according to Lawrence Berkeley National Laboratory's 2025 Update (June 2026). LBNL projects that share will reach about 12% by 2030 (a 9.5% to 15.3% range), reversing two decades of essentially flat national demand.
Why does it take so long to connect a data center to the power grid?
There are two lines, and both are slow. Data centers wait in utilities' large-load queues, where timelines vary by utility; in Virginia, Dominion is assigning connection dates out to 2031 and beyond. The new power plants needed to serve them wait in the generation interconnection queue, where the national median from request to operation has more than doubled — from under two years in the 2000s to about 5 years (61 months) for projects reaching operation in 2025 — because transmission planning, studies and grid upgrades move far slower than the load wants to arrive.
Which U.S. region has the most data centers?
Northern Virginia is the largest data-center market in the world, at roughly 4,500 MW of operating capacity (CBRE, H1 2026). Atlanta, Dallas–Fort Worth, Phoenix and Chicago follow.
What is the ERCOT interconnection queue?
ERCOT's large-load interconnection queue is the list of big new loads seeking to connect to the Texas grid. It grew from about 63 GW at the end of 2024 to roughly 226 GW by November 2025, then more than doubled again to about 474 GW by June 2026, around 90% of it data centers — but under 1% is actually drawing power, since a queue is a wish list, not a build plan.
Is there enough power for AI data centers?
The problem is location and timing, not total capacity. The U.S. has ample generation nationally, but data-center load concentrates onto a handful of already-stressed grids faster than they can respond. Studies suggest much of the near-term demand could be met by using existing grid capacity more flexibly.
Prefer the free 3-page summary first? The 8 findings and the chart that explains the AI power crunch — straight to your inbox.
You’ll also get Power Watch, our free weekly digest of data-center power news, every Tuesday. No spam. Unsubscribe anytime.