The US AI data-center buildout now spans 1,547 mapped facilities across 46 states, with 83 frontier sites accounting for roughly 12.7 gigawatts of capacity and 530 state and federal bills filed to regulate them. Those figures come from a live snapshot published August 21 that also tracks 521.6 gigawatts of grid demand across seven markets, a number that lands just above the roughly 506 gigawatt annual average implied by total US electricity generation in 2025. The tracker refreshes from public feeds hourly and labels each figure with its source and confidence.
Key facts
- 1,547 facilities mapped across 46 states; 83 frontier data centers totaling about 12.7 gigawatts; 530 data-center bills logged.
- Virginia leads with 342 facilities, then Texas at 140, Oregon at 133, California at 111, and Ohio at 90.
- Snapshot dated August 21, 2026, published at datacenters.builtfor.ai with a documented methodology.
- The five largest named sites: Colossus 2 at 946 megawatts, Anthropic and Amazon's New Carlisle at 910, Microsoft's Fairwater Atlanta at 636, Meta's Prometheus at 562, and OpenAI's Stargate Abilene at 421.
The single most important thing to get right about this report is what the 521.6 gigawatt figure means, because it is being misread. It is live grid demand across the seven market regions the tracker monitors, drawn from federal energy feeds and California's grid operator. It is not data-center consumption, not announced load, not permitted capacity, and not anything anyone plans to build. It is how much electricity those regions are using, full stop. The reason it appears in a data-center report is context: it is the size of the system this buildout is being inserted into.
With that framing, the useful comparison is the frontier slice. The named megawatt figures put five facilities in a class that did not exist three years ago. Colossus 2 at 946 megawatts is roughly the output of a large nuclear reactor unit dedicated to one building. Anthropic and Amazon's New Carlisle site is in the same range. Together the 83 frontier data centers come to about 12.7 gigawatts, which is small against national generation and enormous against any single utility's planning horizon. Grids are not sized for national averages; they are sized for what shows up in one interconnection queue in one county.
That is why the geographic concentration matters more than the totals. Virginia's 342 facilities are not spread across Virginia. They are stacked into a corridor whose transmission and water were provisioned for a very different future, and the political consequences are already visible: Ground Truth has covered Virginia ordering its utility to bill data centers for their own power lines, New York pausing new hyperscale approvals, Nashville voting to seize a data-center site, and voters removing a state senate president over one. The 530 tracked bills are the legislative shadow of that same pressure.
The growth curve is what turns a snapshot into a trajectory. Epoch AI reports that the record for the largest AI data center has doubled roughly every seven months since Colossus 1 came online in August 2024. Doubling times are hard to feel intuitively, so put it concretely: a facility that would have been the largest in the world in mid-2024 would not make the top tier today, and at that rate today's 946 megawatt leader is a mid-sized site by late 2027. Epoch separately finds that spending on AI data centers, hardware, and networking reached 1.4 percent of US GDP in the first quarter of 2026, doubled from 0.7 percent. Whatever else this is, it is now a macroeconomic quantity.
Grounding that against real generation: the Energy Information Administration reports US electricity generation of 4.43 thousand terawatt-hours in 2025, a record. Divide by the hours in a year and the system averages about 506 gigawatts. That average is the honest denominator for anyone trying to reason about how much headroom exists, and it is the number that makes the frontier buildout's 12.7 gigawatts legible: roughly two and a half percent of average national output, concentrated into 83 buildings, most of them clustered in five states.
The honest caveat is about the source itself. This is an aggregator, not a regulator, and it is transparent about that: it pulls public primary data, tags confidence, and refreshes hourly. Facility counts depend on what counts as a facility, and megawatt figures for private buildings are frequently announcements rather than meter readings. One number circulating alongside the report needs correcting outright: a claim of $4.1 trillion in AI data-center spending by 2028, attributed to UBS, does not match the accessible UBS note, which says $3 to $4 trillion annually by 2030 and names access to power as the binding constraint. The years and the framing are different, and the difference matters for anyone doing arithmetic on it.
Originally published on Ground Truth, where every claim is checked against the primary source.
