Selection, Allocation, and Timing: Three Obstacles to Measuring the Effect of Large-Load Growth on Household Electricity Prices

Published estimates of how hyperscale computing load affects household electricity prices disagree about sign as well as magnitude. This article argues that much of the disagreement reflects three identification problems that existing designs often absorb rather than model directly. Large loads choose low-cost locations, making exposure endogenous. Average revenue per kilowatt-hour cannot distinguish a change in rates from a change in the composition of customers within a class. In jurisdictions that allow construction costs into rates before a plant enters service, bills may change years before the associated load appears in sales data. I examine each problem using public data. With state-level price and sales data for 2023 and 2024 for the 50 states and the District of Columbia, I find direct evidence of siting selection: the 2023 residential price level explains 22 percent of the cross-state variance in subsequent non-residential sales growth. That sorting appears to follow a state’s general cost level rather than the depth of the discount already received by large customers, a more tractable pattern for identification. Non-residential sales growth has no detectable relation to residential price growth, although the design lacks power to detect elasticities below roughly 0.9, so the null is not substantively informative. The residential-to-industrial price ratio rises with load growth, but decomposition shows that seven-tenths of the movement comes from falling non-residential prices rather than rising residential prices; the same pattern appears relative to the commercial class, and half of the association disappears after adding a four-category regional control. Translating the reduced-form slope into an implied price for incremental load shows that a pure composition account would require these kilowatt-hours to be served at about one-third of the class average, a demanding but possible figure whose confidence interval cannot adjudicate the issue. The evidence favors a regime-dependent interpretation of the literature and indicates that future work should treat capital deployment, rather than realized energy sales, as the central treatment variable.