Virginia has more large data centres than any other US state. The preprint names PJM, the regional grid operator covering much of the mid-Atlantic, as Virginia's dominant balancing authority. A national share of electricity can look small on paper while a regional share is not. A new study of 403 US hyperscale data centres puts the central-scenario figure at about 1.8 per cent of US electricity. The number that matters for stability is the one the national average hides: how much of that load sits in a few places, on a few lines.

What the study counts

The preprint compiles facility-level data on 403 US hyperscale data centres that operated from May 2024 to April 2025. It was posted to arXiv on 3 June 2026 as arXiv:2606.05420, under the title "Assessing the Carbon Emissions and Energy Consumption of U.S. Hyperscale Data Centers."

Its consumption figure is a range, because it depends on assumptions about how much of each facility's capacity was in use. Under the central scenario, the authors estimate demand of about 82 terawatt-hours, or about 1.8 per cent of US electricity consumption. Across all facility-load scenarios, the range is 68 to 99 terawatt-hours.

The authors also report that 53.9 per cent of the electricity attributed to these facilities came from fossil-fuel generation, with 20.9 per cent from nuclear and 25.3 per cent from renewable sources.

Our reading: the national share is not the number that matters for stability. A grid is run region by region, and a share that looks small nationally can be large in one balancing area.

Where the load sits

Virginia has 142 of the facilities, more than any other state. Oregon has 56 and Ohio has 38. The paper also finds that facilities in Virginia had the highest aggregate electricity consumption of any state, followed by Oregon, Ohio and Iowa. In the central scenario, Virginia's facilities account for about 21.0 TWh, Oregon about 11.0 TWh, Ohio about 9.0 TWh and Iowa about 7.0 TWh.

The preprint states the concentration directly: "Together, these four states accounted for more than 50% of total HDC electricity consumption in our analytical sample" (arXiv:2606.05420, HTML version).

The authors say many of the facilities are geographically clustered. Because of that, they test their model by holding out whole balancing authorities or states, since a random split of closely packed facilities would tell them less about how the model performs in a new region. The clustering is a finding about the data, and it is also a warning about the grid.

Our reading: a grid does not spread its risk evenly. A failure on a major line, a heat wave that takes generators offline, or a fuel shortage can hit one region at once. The more large loads sit in that region, the more one event has to disrupt.

Why the arithmetic changes at campus scale

The study measures facilities that were already operating between May 2024 and April 2025. Its figures describe that period. They are not a forecast of what the grid will carry next year.

Our reading: the planning question is less whether a region can generate enough power on average, and more whether the transmission and distribution network can carry a new block of demand. Data centres run around the clock. A grid that treats them like ordinary customers may find them less forgiving of a bad hour. The preprint does not test this. The argument here is an inference from the clustering the authors report, not a result of the study.

The fuel question

The fossil share is the second part of the picture. The paper also gives a carbon intensity for the electricity attributed to these facilities of about 545 grams of CO2 per kilowatt-hour, which it says is about 48 per cent above the contemporaneous US national grid average.

Our reading: if demand grows in a clustered region, the generation built to serve it could lock in fossil capacity for years. This essay does not forecast the future generation mix. That would need a separate model, and the preprint does not supply one.

Implications: who carries the risk

Three groups carry most of the exposure.

Our reading: ratepayers in the clustered regions may pay for new transmission and generation built to serve a few large loads, even though the costs are spread across everyone on the network.

Our reading: grid operators face single-point stress. A balancing authority that must absorb several large loads in a short period is exposed in a way that a dispersed load is not. The clustering finding makes that exposure visible, although the preprint does not model failure.

Our reading: host communities carry local costs, such as land use and infrastructure, while the benefits of computing are spread across a wider network and across the firms that own it.

Speculation: if a regional grid were tested by a multi-day outage during a heat wave, the first question would not be how much power the region produced in a year. It would be which loads could be shed, and who decided.

Conclusion: the grid remembers where it was built

Compute is becoming a large and concentrated load on the grid, and the concentration is the part the national numbers hide. The preprint gives a careful count and a useful map. It does not tell us how a region will cope when several large campuses arrive together, or when the next hard hour comes. That question belongs to the planners who approve the next campus, before the next transmission line is built.

--- Sources: Guidi G., Dominici F., Squartini T., Sprinkle C., Gilmour J., Butler K., Bell E., Delaney S., Bargagli-Stoffi F. J., "Assessing the Carbon Emissions and Energy Consumption of U.S. Hyperscale Data Centers," arXiv:2606.05420, submitted 3 June 2026 (

https://arxiv.org/abs/2606.05420

). Affiliations, state counts, PJM and the fuel shares are taken from the paper's HTML version (

https://arxiv.org/html/2606.05420

). This is an arXiv preprint; this essay does not treat it as peer-reviewed.