Asa Watten, John Bistline, and Geoffrey Blanford posted a paper to arXiv on June 18 titled Have Data Centers Raised Your Electric Bill? Causal Evidence from the United States. Their answer is no. Running an instrumental-variables design across state-year observations from 2015 to 2024, they estimate that a doubling of a state's data center capacity caused average residential retail electricity prices to fall about 3.5 percent. Scaled to what actually happened, the average residential customer between 2019 and 2024 lived in a state where data center capacity grew 160 percent, and the paper attributes a 6 percent rate decrease to it.
This is the first causal estimate anyone has put on the question that is driving the moratorium wave. Every ratepayer revolt we have covered, every county commission vote, and the $64 billion in projects blocked or delayed rest on a premise this paper attacks directly. Read it before somebody's government affairs team hands you a one-slide version.
Watten and Blanford work for the Electric Power Research Institute. Bistline is at Watershed and spent years at EPRI before that. EPRI is funded by the electric utility industry, and a finding that large new load pushes rates down is useful to every utility currently asking a commission for permission to build. Footnote 2 of the paper notes that a version of the argument already appeared in an EPRI publication from 2025, which makes this the empirical backing for a position the institute had taken. The headline load figures, 4.5 percent of US electricity in 2024 and a projected 9 to 17 percent by 2030, also come from EPRI.
None of that makes the work wrong. Utility-funded researchers publishing a methodologically explicit paper with the code path visible and the weak-instrument diagnostics reported is a substantially higher standard than the advocacy on either side of this fight has been meeting. It does mean the affiliation belongs in the first paragraph of any piece citing it, including this one, and it means the caveats deserve as much attention as the point estimate.
The economics are not exotic. Retail electricity tariffs recover average costs across a system carrying enormous fixed capital in generation, transmission, and distribution. Distribution is a natural monopoly. When demand rises within existing capacity limits, grid operators dispatch more expensive generators and wholesale prices climb, which is the effect everyone expects. Running against it is a second effect: those fixed costs now spread across more kilowatt-hours. The paper argues the second effect dominated over this period, and finds economies of scale showing up separately in transmission, distribution, and generation costs.
A robustness check adds controls for new generation capacity by technology, covering wind, solar, natural gas, and storage. The point estimate barely moves. Watten and his coauthors read that as evidence the price-reducing effect runs partly through improved utilization of transmission and distribution capital rather than through changes in the generation mix. Dropping California, which saw prices rise nearly 40 percent on wildfire hardening costs, and dropping 2020 both leave the estimate intact.
The instrument is the total length of the 1947 Eisenhower Interstate Highway Plan within each state. The logic runs that terrestrial fiber follows interstate corridors, fiber drives data center siting, and a highway plan drawn up in 1947 cannot plausibly have been chosen in response to 2020s electricity prices. To satisfy the exclusion restriction the authors condition on state population and GDP, blocking the channels through which highways shaped economic development, and treat the fiber pathway as the remaining relevance channel.
Three things about that should give a careful reader pause. The first-stage F statistic is 7.8, under the conventional weak-instrument threshold of 10, which the authors handle honestly by reporting Anderson-Rubin confidence intervals rather than standard ones. That interval runs from −12.1 percent to −1.4 percent, an order of magnitude wide, so the sign is better established than the size. Second, because the instrument is time-invariant it collinearly wipes out state fixed effects, so the second stage drops them for regional effects. The paper says plainly that this forces heavier reliance on the controls. Third, a first-differences specification returns an estimate of precisely zero for both OLS and two-stage least squares. Watten and his coauthors interpret that as evidence the price effect takes years to materialize, which is a reasonable reading. A skeptic gets to read it as the result failing under the most demanding specification in the paper.
The sign is better established than the magnitude. Anyone quoting "six percent lower" as settled is quoting the midpoint of an interval that runs from negligible to enormous.
This is the part the industry will skip. The authors name three conditions that could reverse their finding, and all three are present in 2026. Gas turbine orders are backlogged with wait times reported as long as seven years and prices climbing. Transformer and electrical equipment supply chains remain constrained. Tariffs on Chinese photovoltaics raised solar costs, and federal approvals for wind and solar have stalled in litigation. Every one of those makes incremental capacity more expensive than the incumbent supply it adds to, which inverts the mechanism the paper identifies.
The second caveat is sharper for anyone selling into the buildout. The forward-looking result depends on demand growth being durable. If announced load materially exceeds what gets consumed, a real possibility given interconnection queue dynamics, utilities overbuild and the fixed costs of that overbuild spread across fewer kilowatt-hours than planned. That reverses the sign. Kansas, Michigan, and Delaware have already moved toward minimum bills and long-term contracts from large loads for exactly this reason.
Put plainly, the paper establishes what happened from 2015 to 2024 under conditions of slack capacity and falling unit costs. It does not establish what happens from 2026 forward under turbine shortages, tariff-inflated solar, and 73,000 megawatts of off-grid gas being built around the interconnection process entirely. The authors say so themselves in the second sentence of the abstract.
Two reasons, and the second one is not in the paper. The first is commercial. Ratepayer anger is the strongest fuel in the coordinated opposition that is stalling projects, and stalled projects are cancelled cooling orders. A credible causal estimate pointing the other way changes what a utility can say in a rate case and what a developer can say at a county hearing. Expect this citation inside a month. Expect it stripped of its confidence interval.
The second is a thermal point the paper gestures at and does not develop. The mechanism rewards high utilization, which means the rate benefit is a function of load factor. Data center load is flat and steady in a way that heat pumps and EV charging are not, and the authors explicitly set aside load shape as beyond their scope while noting electrification profiles differ. Our inference, stated as inference: any move that makes data center load lumpier works against the economics this paper describes. That puts curtailment deals and demand-response programs like the University of Utah's flexibility arrangement in an awkward position, because grid flexibility and maximum asset utilization are pulling in opposite directions. Thermal storage and hardened cooling capacity are what let a facility hold a flat draw through a heat event instead of shedding load. That has never been priced as a ratepayer benefit. It arguably is one.
This is the best evidence in the argument and it is not good enough to end it. The finding is directionally credible, honestly reported, stable when the obvious confounders are dropped, and produced by people whose employer benefits from the answer. The confidence interval is wide, the instrument is weak by convention, and the historical window it covers has materially different supply conditions than the one the buildout is now entering.
Our read is that the effect was real through 2024 and will not survive 2027 intact, because the caveats section is describing the present tense. The industry should cite this paper carefully and with the interval attached, because the version that gets quoted without caveats will be dismantled in public by the first competent economist the opposition retains, and the credibility lost in that exchange will cost more than the citation was ever worth. Cite the range. Concede the supply constraints. Then go win the argument on the part that is actually defensible.
A June 2026 study by Asa Watten, John Bistline, and Geoffrey Blanford estimates the opposite for the 2015 to 2024 period. Using an instrumental-variables design, they find a doubling of state data center capacity caused average residential retail prices to fall about 3.5 percent, with a weak-instrument-robust confidence interval of −12.1 to −1.4 percent. The authors caution that future supply constraints could reverse the effect.
Because retail electricity rates recover average costs rather than marginal costs. Distribution networks are natural monopolies with high fixed costs. When durable new load raises utilization of existing generation, transmission, and distribution assets, those fixed costs spread across more kilowatt-hours, pushing average cost down. The paper finds this outweighed the dispatch effect that pushes wholesale prices up.
The total length of the 1947 Eisenhower Interstate Highway Plan within each state. The relevance channel is that terrestrial fiber-optic infrastructure typically runs along interstate corridors, which drives data center siting. The first-stage F statistic is 7.8 in the preferred specification, below the conventional threshold of 10, so the authors report Anderson-Rubin confidence intervals robust to weak instruments.
The authors name three: supply constraints such as gas turbine backlogs and transformer shortages, the durability of demand growth if anticipated load never materializes, and fuel cost effects. Two more sit in the design. The instrument is time-invariant, so state fixed effects are dropped from the second stage in favor of regional effects, and a first-differences specification returns an estimate of precisely zero.
Source: Watten, A., Bistline, J., and Blanford, G., "Have Data Centers Raised Your Electric Bill? Causal Evidence from the United States," arXiv:2606.19777v1 [physics.soc-ph], submitted June 18, 2026. Point estimates and diagnostics from Table 2 and the Discussion section; robustness checks and spillover results from the Supplemental Appendix. Author affiliations as stated in the paper: Watten and Blanford at EPRI, Bistline at Watershed. Data center capacity data sourced by the authors from S&P Global Market Intelligence; the 1947 highway plan geometry from the National Transportation Atlas. Cryptocurrency mining and hosting are excluded from the capacity measure. The load-shape argument in the "Why the Cooling Industry Should Care" section is The Cooling Report's inference and is not a finding of the paper, which explicitly sets load shape aside as out of scope.