Who Pays for AI Power? How Frontier Labs Got the Electricity Question Wrong
Who Pays for AI Power? How Frontier Labs Got the Electricity Question Wrong
Electricity in the United States is not a product. It is regulated public infrastructure. That distinction has been settled law and public policy for nearly a century. Frontier AI labs, in their race to build the largest models on Earth, either did not understand that distinction or chose to bet against it. Either way, the bet is now losing.
The scale of what happened is not abstract. Regulators approved $11.6 billion in rate increases during 2025, driven almost entirely by AI power demand. In the service region of PJM Interconnection, the country's largest grid operator, capacity prices exploded from $28.92 per megawatt-day in the 2024-2025 period to $329.17 per megawatt-day for 2026-2027. That increase did not land on the balance sheets of OpenAI or Google DeepMind. It landed on monthly utility bills.
The public noticed. And once voters frame a question as "a private company is consuming our public infrastructure for profit," the political math does not reverse.
The Hyperscaler Playbook, Run at the Wrong Speed
The frontier AI industry did not invent the data center buildout. It borrowed the playbook from cloud hyperscalers — Amazon, Google, and Microsoft — who had been building out infrastructure quietly for 15 years without triggering a comparable public revolt.
The reason that playbook worked for hyperscalers is simple: pace. A new facility in northern Virginia or central Iowa every couple of years gave communities time to absorb the grid impact. Utility planning cycles run on 5-to-10-year timelines. That rhythm was compatible with slow, geographically concentrated expansion.
Frontier AI compressed that 15-year timeline into roughly two years. The grid was not built for that rate of change. The assumption that the public benefit argument — the idea that everyone gains from AI, so everyone can reasonably share the infrastructure cost — would hold long enough was borrowed from a playbook that ran at a completely different speed.
Lawrence Berkeley National Laboratory data tells the story plainly. Power demand from U.S. data centers doubled between 2018 and 2024. It is projected to triple by 2028. BloombergNEF estimates data centers will consume more than 100 gigawatts by 2035. The grid was not designed for this. Neither was the public's tolerance.
Electricity is infrastructure, not a product. Frontier labs treated it like server hardware — a cost to minimize and offload. That was the category error that is now being corrected.
The Public Already Figured It Out
The opposition is not a niche concern. It is broad, deep, and bipartisan in a way that is genuinely unusual in contemporary American politics.
A Gallup poll conducted in March 2026 found that 71 percent of American adults oppose building an AI data center near their home. A separate Heatmap Pro survey in May 2026 confirmed that figure, with 55 percent saying they were strongly opposed. An NBC News Decision Desk survey found 69 percent opposition. A New York Times and Siena University poll of likely midterm voters put overall opposition at 61 percent.
The Gallup finding that put the issue in sharpest relief: Americans now appear more willing to live near a nuclear power plant than an AI data center. Fifty-three percent oppose a local nuclear plant. Seventy-one percent oppose a local data center.
The opposition crosses partisan lines. Gallup found between 63 and 75 percent opposition among both Republicans and Democrats. The New York Times poll found 75 percent of Democrats opposed and 49 percent of Republicans opposed — a rare case where a majority of both parties, including the president's own party, disagrees with the direction of a major industry the president has publicly championed.
In the first quarter of 2026 alone, community and legislative actions blocked or delayed more than 75 data center initiatives representing over $130 billion in planned investment. More than 300 state-level bills targeting data center regulation were introduced in the first half of 2026. New York's governor signed an executive order in July 2026 pausing new hyperscale data center permits for a year.
We Have Seen This Before
The pattern is not new. It is one of the most reliably recurring stories in American economic history: a private industry receives access to public infrastructure, frames that access as a public benefit, and quietly socializes the costs until the public stops accepting the arrangement.
The railroad land grant era is the closest analog. Between the 1850s and 1870s, Congress granted railroads approximately 170 million acres of public land and right-of-way access, on the argument that everyone benefits from rail connectivity. The public absorbed enormous indirect costs. Within 20 years, the political backlash produced the Interstate Commerce Act of 1887 — the first major federal regulation of a private industry in American history. The argument that finally shifted public opinion was straightforward: we gave you everything, and you are still charging us.
The Rural Electrification Act of 1936 cemented something else into American civic consciousness: electricity itself is infrastructure, not a commodity. Private utilities had refused to serve rural areas because it was not profitable. The government stepped in. The lesson that generation absorbed — that power is a public good — is exactly the frame voters are now applying to AI data centers.
Once a resource gets categorized as public infrastructure in the American political mind, the companies consuming it for private profit face a very specific kind of reckoning. It does not come quickly. But it comes.
Once voters think of power as a public good being consumed by private profit, the political math changes permanently. Frontier labs are learning that the hard way.
The Pledge and Its Limits
In February 2026, during his State of the Union address, President Trump announced a "ratepayer protection pledge," telling major technology companies they have the obligation to provide for their own power needs. By early March, Google, Microsoft, and Amazon had signed on. The companies agreed to build or buy new sources of power generation for their data centers and cover the expense of infrastructure upgrades.
The pledge has a structural problem. Experts noted immediately that the president cannot unilaterally change how data centers source their energy. These arrangements are governed by contracts with utilities. Any change in how data centers pay for energy has to go through state utility regulatory processes. Those processes move in years, not months.
The gap between the announcement and the regulatory reality is precisely where the next wave of public anger lives. Bills will continue to rise while the pledge works its way through state commissions. The companies that moved earliest toward behind-the-meter power generation — xAI, which trucked mobile generators into Memphis to power its Colossus supercomputer cluster, and Microsoft, which has pursued nuclear power agreements — will be better positioned than those who waited for a federal mandate that was never enforceable to begin with.
The pledge is not the end of this story. It is the opening signal that the voluntary era is closing.
Power Was Only the First Domino
The electricity fight is where the public noticed first. It will not be the last front.
Water consumption is next. Data centers require enormous amounts of water for cooling. That conflict is already beginning in drought-stressed regions where water rights carry the same political weight that electricity rates carry in the Midwest and Mid-Atlantic. After water will come local zoning battles, tax abatement clawbacks as communities recalculate whether promised jobs materialized, and environmental reviews of backup power infrastructure.
The underlying problem in each case is the same one. Frontier AI labs did not just miscalculate one line item. They misread the nature of the resources they were consuming. Electricity is infrastructure, not a product. Water is a public resource, not a utility input. Land use is a community decision, not a permitting formality.
The correction that started with the ratepayer protection pledge is not the correction itself. It is the signal that a larger and longer correction has begun. Voluntary pledges are what industries offer when they sense that mandatory regulation is coming and they want to write the terms. The terms will be written regardless. The only question is who writes them.
Aaron Rose is the founder and publisher of Tech Reader Magazine, covering system architecture, cloud platforms, and AI policy.