Cornell study: data centers shift grid costs onto rural, low-income homes

Data centers are pushing up US power costs, and households are paying a lopsided share, a Cornell study published November 9 in Nature Energy finds. In regulated tariffs, residential customers are allocated a median of 47% of costs "despite contributing far less to demand growth", and cumulative wholesale price increases attributable to data centers exceed 10% in the most exposed hubs by 2025, Cheng Feng and Fengqi You report. The poorest and most rural households take the heaviest hit.

The paper traces two channels. One runs through wholesale markets, where the authors estimate price effects from price records, anchoring each zone's exposure to its data center footprint before the large language model surge (end-2019 for most zones, end-2021 for ERCOT). The other runs through retail tariffs, read from transmission plans and regulatory cost-allocation dockets. Cornell's summary says the team drew on 4,000 existing and proposed data centers and studied four major US grids from 2020 to 2025.

Where prices moved

Silicon Valley tops the list. In the paper's supplementary tables, the PG&E Santa Clara and Bay Area zone's projected 2025 increase equals 19.93% of its average price over the sample period, followed by ERCOT North at 16.84%, Boardman, Oregon at 15.62% and Dominion's zone in PJM at 15.17%, the last on 7.9 GW of added data center capacity. The intervals are wide: Dominion's 95% confidence interval runs from 3.33% to 27.01%, and ERCOT North's from 0.91% to 32.78%. The New York, New England and Southwest Power Pool markets show no statistically significant effect.

Transmission is where the allocation rules do the work. Cornell's summary says data centers paid $3.12 billion of an $8.25 billion bill in 2025 for infrastructure linked to their growth, or 38%, and that transmission investment rose from $2 billion in 2020 to $34 billion in 2025, with data centers driving much of that rise. The allocation is often based on past consumption, which a data center still being built does not have. "But before they build, they're not using energy," You said in the summary.

Virginia shows the shape of it. Dominion Energy Virginia's residential class carries 59.75% of the company's Virginia jurisdictional transmission revenue requirement, against 8.25% for GS-4, the large commercial and industrial class that contains data centers, per the Dominion rate filing the supplement reproduces. Dominion's zone also has the largest data center capacity addition in the study's zone price table.

Who carries it

The burden falls unevenly. Per the paper, cost sharing leaves rural communities with burden increases 76% larger than urban ones, and the lowest-income households face extra energy burdens of up to 28 percentage points. That top figure is the extreme case: the supplement puts the maximum for households at 0–30% of area median income at 27.7 points, with an unweighted 99th-percentile increase of 12.6. Cornell's summary describes the rural figure as the share of rural families spending more than 6% of their incomes on energy rising 76% more than among urban residents.

The costs also cross state lines. Residents of New Jersey, Pennsylvania and Ohio help pay for grid upgrades linked to data center growth in Virginia, the researchers said. "You do not have to live next to a data center to help pay for its power infrastructure," You said.

Bigger incentives do not buy proportionally bigger returns. Most states clear break-even: across the 28 states with incentive data, the supplement puts the aggregate at 21.2 dollars of labor income per dollar of incentive, and 25 of the 27 states with a defined ratio exceed one. But the paper finds local economic returns "almost unrelated to spending", with a spending elasticity of 0.16. In the supplement's estimate across 23 states, a tenfold increase in incentive spending is associated with about 1.4 times the labor income, and adding each state's 2020 employment in the data processing and hosting industry drives the spending effect to 0.015. Labor income follows the workforce a state already had, not the size of the check it wrote. The states absorbing the largest burden increases, the paper says, capture the smallest returns.

The fixes on the table

You and Feng recommend, per Cornell's summary, that data centers cut power use during 22 of each region's peak demand hours a year, which they found would relieve grid strain and minimize the need to expand in all four regions. States could also charge data centers directly for the new demand they add, or direct funds to low-income households.

Regulators are already testing the direct-charge and curtailment ideas. Florida's commission denied Duke Energy Florida's large-load tariff on November 3 and told it to file a large-load rate schedule with adequate cost support, and Pennsylvania's consumer advocate is seeking a separate data center rate class at PPL. At FERC, PJM's large-load curtailment plan was rejected on October 9, and PJM plans its large-load filings on or about November 16, as tracked on our interconnection queues page.

Anyone wanting to check the work can get most of the way there: the aggregated data, docket records and code are public on Zenodo. The facility-level inventory, licensed from S&P Global, cannot be redistributed.