Opinion: Households are subsidizing America's AI data center boom

Three Democratic senators spent nearly a year asking seven of the biggest companies in the world a simple question: who is really paying for the power that artificial intelligence consumes. The answer, delivered this week, is that the companies building the data centers are not the ones footing the bill for the electrical grid those data centers require. Ordinary households are.

That is the blunt conclusion of a Senate investigation released this week by Sens. Elizabeth Warren, Chris Van Hollen and Richard Blumenthal, following letters sent last December to Amazon, Google, Meta, Microsoft, CoreWeave, Digital Realty and Equinix. It is not a partisan talking point invented for the occasion. It is the plain, verifiable shape of how America's electricity markets currently work, and it should trouble anyone who pays a power bill, regardless of what they think about artificial intelligence itself.

What the numbers actually show

The senators' 27-page report found that the seven companies "commit to paying for infrastructure that solely benefits them, but dispute how much they should pay for shared infrastructure" — the substations, transmission lines and generation capacity that serve a data center alongside everyone else on the grid. They resist what regulators call "but-for" cost allocation, the principle that if an upgrade would not be needed but for a specific new customer, that customer should pay for it. All of them use nondisclosure agreements with utilities, landowners and sometimes local officials, limiting what the public can learn about the deals being struck in their name. And despite collecting billions of dollars in state and local tax incentives, none of the seven could or would supply comprehensive data on how many permanent jobs their facilities actually create.

These are not abstractions. PJM Interconnection, the grid operator serving 65 million people from Illinois to the mid-Atlantic, publishes the clearest paper trail. Its independent market monitor's own report found that existing and forecast data center load drove a combined $29.4 billion increase in capacity-market revenue across PJM's last four base auctions, and that data center demand alone made up roughly 9% of all wholesale power costs in the region through July of this year. In Virginia, the state most saturated with data centers, regulators approved a Dominion Energy rate increase that raises the typical residential bill by $11.24 a month, and the State Corporation Commission has since ordered Dominion to draft a new tariff assigning more of the transmission costs specifically to large-load customers, because the existing framework was not doing that job. Nationally, federal forecasters project the average residential price of electricity will climb to about 18.2 cents per kilowatt-hour this year, with the steepest increases concentrated on the East Coast, where the government's own energy outlook anticipates annual growth of 5% to 7% through 2027.

How the costs became invisible

None of this happened by accident; it happened because electricity-market rules were not built for this moment. Utilities recover grid investment through rates charged to everyone on the system, a model that made sense when demand grew slowly and predictably. AI data centers broke that assumption: a single hyperscale campus can demand as much power as a mid-sized city, and can appear, expand or walk away from a site faster than a regulator can finish a rate case. States compete to land these projects with sales-tax exemptions worth billions, while companies negotiate confidentially, utility by utility, so no single regulator ever sees the whole picture.

Tax policy compounds the imbalance. A separate analysis by the Institute on Taxation and Economic Policy found that Meta's effective federal income tax rate fell to roughly 3.5% in 2025, with its tax bill dropping to about $2.8 billion from $9.6 billion the year before, largely on the strength of accelerated depreciation and research credits tied to its AI infrastructure buildout. A company can simultaneously minimize what it owes the federal government for the privilege of building a data center and resist paying the local utility the full cost of plugging that data center into the grid. Both moves are legal. Together, they describe an industry treating the public balance sheet, federal and local alike, as a line item to be minimized rather than an obligation to be met.

Who actually bears the cost

The people absorbing this are not shareholders or executives. They are residential customers in Loudoun County, Virginia, and in the exurbs of Atlanta and Columbus, who open a monthly bill that is higher regardless of whether they have ever used a chatbot. They are renters and fixed-income retirees with the least room in a budget to absorb a $10 or $20 increase. They are residents of towns that waived tax revenue for a data center on the promise of jobs the companies cannot, or will not, document. And they are future ratepayers generally, who will still be paying off decades-long transmission investments if the AI investment cycle cools before the capacity it required is actually needed — a risk regulators are only beginning to confront, since forecast demand, not just built demand, is already driving up auction prices today.

The strongest counterargument, taken seriously

The companies' defense is not frivolous. They argue, correctly, that they already pay substantial sums for dedicated infrastructure, and that large, steady industrial loads can in principle spread a utility's fixed costs over more kilowatt-hours sold, lowering the average rate for everyone else. An industry-funded backcast analysis commissioned by the data center developer CloudHQ argues that incremental data center load growth has modestly lowered, not raised, the average Dominion residential bill. There is a real version of this argument: a round-the-clock customer locked into a long contract is, other things equal, a better customer for a grid than a household whose demand spikes for three hours on a summer evening. Virginia's new large-load rate class, requiring data centers to sign 14-year contracts covering a minimum share of transmission and generation costs, is a serious attempt to capture that logic rather than simply punish the industry.

But that argument proves the case for structural reform, not against it. It only holds when the cost allocation is actually calibrated to cause and effect — when the company signing the long contract is the one paying for the specific upgrade its load requires. The Senate report's core finding is that this calibration mostly is not happening: companies are disputing "but-for" standards precisely where they would be on the hook, while readily accepting infrastructure that benefits only them. The industry cannot credibly claim both that data centers subsidize the grid and that assigning them the cost of the upgrades their own load triggers is an unreasonable overreach. Virginia regulators evidently agreed, which is why they ordered a new tariff rather than accepting the status quo.

"The data center industry is bulldozing local communities… leaving ordinary citizens stuck with the bills."

That line from the senators' report is pointed, but the arithmetic behind it — a $29.4 billion capacity-cost increase documented by a grid operator's own market monitor, a residential rate hike approved by state regulators, a federal tax bill that fell by more than two-thirds in a single year — is not partisan rhetoric. It is what happens when market rules lag a technological boom by several years.

What should happen now

The fix does not require banning data centers or slowing the deployment of artificial intelligence, however tempting either might be to the technology's opponents. It requires closing the gap between who causes a cost and who pays it. The Federal Energy Regulatory Commission should require large-load tariffs nationwide, modeled on the direction Virginia has set, with cost-causation standards and take-or-pay terms strong enough that a canceled project cannot strand its costs on everyone else. Congress should extend the mandatory data center energy reporting it has already secured from the Energy Information Administration to cover cost allocation and job creation, not merely power consumption. And state utility commissions should stop approving nondisclosure agreements that shield the terms of these deals from the ratepayers who will spend decades paying for them.

Artificial intelligence may well be as transformative as its promoters insist, and the country has a legitimate interest in building the infrastructure to support it. But that is precisely the argument for getting the accounting right now, while the rules are still being written, rather than after millions of households have spent years quietly covering a bill that was never really theirs to pay.