TL;DR
- A study published in Nature Energy on 9 November 2026 splits the 2025 cost of grid expansion tied to US data center growth: 8,25 billion dollars in total, 3,12 billion paid by the data centers, 5,13 billion left with everyone else.
- Transmission investment in the projects examined went from 2 billion dollars in 2020 to 34 billion in 2025.
- Three days earlier a Wharton and Harvard working paper put the data center contribution to residential price growth at roughly 4 percent, about 1,70 dollars per MWh. Both figures hold up. They answer different questions.
- Whichever one governs policy is decided by a cost allocation rule, and that rule is being rewritten right now by a body no voter elects.
Assumptions
- The Cornell figures come from the paper's own institutional summary and cover four US grid regions, around 4.000 existing and proposed data centers, 2020 through 2025. I have not read the full text behind the paywall.
- The Wharton and Harvard estimate (Rana, Netessine, Kaps, SSRN, October 2026) clears a 10 percent significance threshold and not a 5 percent one. The authors say so themselves.
- This is US-specific. European cost allocation works differently, so the mechanism travels and the numbers do not.
The sentence worth stealing
Fengqi You, who wrote the Nature Energy paper with Cheng Feng, summarised the result like this:
"You do not have to live next to a data center to help pay for its power infrastructure. The grid connects communities across states, and cost-sharing rules connect them financially."
That is a description of plumbing, not a complaint. Residents of New Jersey, Pennsylvania and Ohio help fund upgrades driven by growth in northern Virginia. New York and New England show no clear data center effect, because little got built there. The money moves along the wires and along the tariff, and the tariff is the longer of the two.
Why the two studies disagree
The Wharton and Harvard paper, The Cost of Intelligence, looked at 385 large data centers across 42 utilities and found a residential effect near 1,70 dollars per MWh, about 1,4 percent of the average 2020 residential price. On industrial customers the same method found 5,40 dollars per MWh, close to 16 percent of the observed increase. The Washington Post reported it on 6 November under a headline saying bills may not be rising that much.
Take 5,13 billion dollars and spread it across every household in four grid regions, and the monthly statement rounds it to nothing. That is not a refutation of the Cornell number. It is the same number, held at the other end.
An average built across tens of millions of households is the one instrument guaranteed not to detect a cost that was spread across tens of millions of households.
The Cornell paper measures the distribution instead, and the distribution is not flat. Cumulative wholesale increases attributed to data centers passed 10 percent in the most exposed hubs by 2025, reaching 15 percent in Virginia and 20 percent in the southern Bay Area. The share of rural families spending more than 6 percent of income on energy rose 76 percent more than among urban residents. And the compensating benefit is thinner than the brochure: state incentive returns ranged from about 180 dollars of worker income per dollar spent in California down to 2 dollars in Virginia.
The precedent nobody cites
American industry has solved this exact problem before, in the literal sense of making a local measurement look good.
Through the 1970s, coal plants met local air quality limits by building taller chimneys. The US had two smokestacks above 500 feet in 1970 and more than 180 by 1985, according to the Government Accountability Office. Local readings improved. Sulfur dioxide travelled further, which is how acid rain reached New England and Canada.
Congress eventually had to write the obvious into statute. Section 123, added to the Clean Air Act by the 1977 amendments, bars states from crediting the dispersion effect of a stack built above its good engineering practice height. You can build the chimney. You just cannot count the dilution as compliance.
There is no Section 123 for ratepayers. In late June 2026 FERC issued show cause orders to PJM and five other grid operators, finding that large load interconnection rules appear unjust in part because other consumers may be covering data center transmission costs. Consumer advocates in Delaware, Illinois, Maryland and Ohio replied that the order does not touch the allocation of the network upgrades themselves. On 9 October 2026 FERC rejected the core of PJM's large load plan, keeping the definition and the registry.
So the year produced a diagnosis, a rejection, and no allocation rule.
The question he did not ask
You asks who pays and who benefits. The sharper question is which rule would have to change, and who has standing to change it.
The county board that approves a campus controls the tax base and the zoning. It has no authority over the transmission revenue requirement that its approval triggers in four states. The body that does have that authority is federal, technical, and electorally untouchable. Household anger is currently being expressed in state races, where the lever is not installed.
Why this matters for your business
If you are modelling energy cost for anything compute-heavy between now and 2030, you are modelling on a tariff that is under active revision in Washington. The number to watch in your own utility's filings is not the average monthly bill. It is the residential class share of the transmission revenue requirement, because that is the line that moves when the rule finally gets written, and reallocation in rate cases tends to arrive backdated.
Dilution worked for fifteen years. Then somebody measured downwind.