An EPRI paper found data centres cut US retail power prices from 2015 to 2024, against forecasts of rises

The single most repeated claim in the backlash against artificial intelligence is that data centres are putting up your power bill. The only long historical dataset anyone has assembled says the opposite happened.

The paper is titled Have Data Centers Raised Your Electric Bill? Causal Evidence from the United States, by Asa Watten, John Bistline and Geoffrey Blanford of the Electric Power Research Institute. It was posted to arXiv on 18 June 2026. Using an instrumental variables approach, the authors estimate that data centres caused average retail electricity rates in the United States to fall modestly between 2015 and 2024. The American Public Power Association, reporting the study on 23 June 2026, puts the central figure at a 3.5 per cent fall in residential prices for a doubling of data centre capacity, with a larger effect at state level. Fortune, on 26 July 2026, reports the state level figure at about 6 per cent.

The mechanism the authors give is not exotic. Electricity systems carry very large fixed costs in generation, transmission and distribution. When demand grows durably, those costs are spread across more kilowatt hours, so average cost per unit falls. Durable demand also pulls forward new investment, and new plant is cheaper per unit of output than the older assets it displaces. The paper documents economies of scale in transmission, distribution and generation, and across retail customer classes. Two case observations sit alongside the regression. California's retail prices rose about 40 per cent between 2019 and 2024, driven mainly by wildfire related costs rather than by data centres. Virginia, where data centres take more than a fifth of state demand, saw below average rate increases.

The authors do not claim this continues. The paper's abstract cautions that future supply constraints could reverse the effect, and RTO Insider's report of 30 June 2026 carries the subheading that the authors caution 2015 to 2024 impacts may not extrapolate to the upcoming buildout. That caveat is doing a great deal of work, because every forward looking number points the other way.

Fortune's reporting carries Goldman Sachs projecting a 6 per cent rise in electricity costs across 2026 to 2027 and a further 3 per cent by 2028, and PJM, the regional grid operator, forecasting a 6.3 billion dollar increase in consumer costs over three years, attributed mainly to data centre demand. A separate 2026 model from North Carolina State University, Carnegie Mellon University and others, reported by Forbes on 3 August 2026, finds that data centre and cryptocurrency demand growth pushes demand weighted wholesale electricity prices up by 6 to 29 per cent nationally by 2030 relative to a scenario without that growth, and by as much as 57 per cent in the hardest hit regions. Forbes is explicit that those are modelled wholesale prices, not household bills, and that what reaches a household depends on rate design.

So the disagreement is not really about economics. It is about whether the forecast demand shows up, and about who pays for the capacity built in anticipation of it. The Belfer Center at Harvard Kennedy School, in a research brief published on 10 February 2026 by Rachel Mural, Henry Lee, Le Xie and colleagues, sets out the scale: data centres consumed 176 terawatt hours in 2023, about 4.4 per cent of national electricity use, and are projected to reach 325 to 580 terawatt hours by 2028, or 6.7 to 12 per cent. Texas peak summer demand is projected to reach 145 gigawatts by 2031 against 85 gigawatts in 2024, with about 32 gigawatts of that new demand from data centres. The brief flags that Dominion Energy in Virginia has proposed its first base rate increase since 1992, adding roughly 8.51 dollars a month for a typical household in 2026 and about 2.00 dollars more in 2027. Its central worry is structural rather than arithmetic: contract based financing has moved projects away from guaranteed rate base recovery towards special tariffs and power purchase agreements, arrangements it describes as lacking transparency and capable of shifting costs onto other consumers. The brief's questions are left open. Who pays for grid improvements prompted by data centre development, who benefits, and how should the costs be allocated.

The Forbes piece answers that question in the only place it is currently being answered, which is state regulation. Ohio and Virginia have adopted tariffs requiring data centre customers to commit for eight to fourteen years, pay for reserved capacity whether or not they use it, post collateral and cover exit fees. The point of those terms is stranded cost: if a utility builds substations and lines for a facility that is delayed, downsized or cancelled, someone carries the investment, and the tariffs are an attempt to make it the developer rather than the ratepayer.

There is a scenario in which both the historical finding and the forecasts are right and the bills still fall, and Fortune quotes the investor Mark Cuban gesturing at it: if AI becomes cheap enough per unit of compute, a large share of announced capacity never gets used, and he suggests a lot of data centres end up as pickleball courts. That is his prediction, not a finding.

What is not known is whether the instrumental variables design survives scrutiny. The identifying assumption is that the instrument moves data centre siting without separately moving prices, and data centres plainly prefer places where power is already cheap. The paper is a working paper, not yet peer reviewed, and no published critique of its instrument has appeared.