The aristocracy of expertise and the politics of the electricity bill

Ten days after the American midterms, the most interesting result for people who build software is not who won the Senate. It is that data centers were on the ballot. On 3 November, 40 of 41 local measures on data centers in ten states went in the direction of restrictions. Democrats took both chambers and flipped Senate seats in Texas, Ohio and Iowa, three states with a lot of new compute. Sherrod Brown in Ohio summarised a whole campaign season: "I do not work for datacenters. I work for you."

Reading the programmes behind those speeches, I see two things mixed together. One is a set of local grievances, some real and some much less so. The other is a national agenda to slow AI as such, carried by people whose professional position depends on being the ones who know. I will call this group the aristocracy of expertise. The idea is old, Pareto already wrote that history is a graveyard of aristocracies.

What is actually on the table

In March Bernie Sanders introduced the AI Data Center Moratorium Act, with a House companion from Alexandria Ocasio-Cortez. It stops construction and upgrades of large AI data centers until Congress passes laws that require federal pre-approval of AI products as "safe and effective", policies against job displacement, a guarantee that utility bills will not rise and a local veto for communities. Each condition is a separate political fight, so the moratorium is in practice open-ended.

In September Sanders and Greg Casar went further with the Ban Artificial Superintelligence Act, which also creates a federal Department of Artificial Intelligence to regulate advanced models. After the election Representative Yassamin Ansari wrote that "a pause on advanced AI development" and a ban on superintelligence should be at the top of the Democratic agenda on day one. The moratorium bill is unlikely to advance for now, but the direction is clear. The question moves from how to build to who is allowed to build.

Who the experts are

The expertise I mean is not technical. AI researchers are the most optimistic group in the country. A Pew survey from 2025 found that 56% of AI experts expect a positive impact on the United States over twenty years, against 17% of the general public. The resistance comes from the licensed and credentialed professions, the people whose income depends on a monopoly on a specific kind of judgement, and from the institutions that speak for them.

The pattern is visible in the details. Illinois banned AI therapy in 2025, with the explicit goal of protecting patients and also, in the words of the state regulator, the jobs of thousands of licensed providers. These are rational positions for the people who hold them. They are less obviously in the interest of the patient who waits three months for an appointment.

Pew also measured something new this year. For the first time a larger share of Democrats than Republicans say they are more concerned than excited about AI, 56% against 49%, and among liberal Democrats the share went from 45% in 2023 to 63%. This is the most educated and most professional slice of the party, and it is the one whose work generative models touch first.

From lobby to market

Professional bodies do not hide what they want, and they do not need to. This week the American Medical Association adopted a policy saying AI can augment physicians but cannot hold a license or a professional title, and it will push for laws that put clinical AI under physician oversight. In 2024 SAG-AFTRA sponsored two California laws on digital replicas, signed by Newsom at the union's headquarters. In New York a proposed bill would stop chatbots from giving any answer that would count as unauthorized practice of a profession if a person gave it.

Each of these rules decides who is allowed to sell what, and that is how a lobby position becomes a market structure. A diagnostic tool that must sit under a physician is a different product, with a different price and different buyers, from one that a patient can use directly. I do not think the associations steer the politicians. Their interests simply align with politicians who need to ride a discontent, and the two reinforce each other.

The data center as a vehicle

A moratorium on superintelligence does not win a county election. A promise that your bill will not go up does. The anti-AI agenda has found in data centers a physical object that people can see, oppose and vote on, and the language that comes with it is borrowed from cost-of-living populism. "Billionaire Big Tech oligarchs" and "the AI billionaire class" come from press releases and campaign speeches, not from academic papers.

I do not think there is a coordinating centre behind this, and I do not need one to explain it. Interests that point in the same direction produce the same behaviour without a meeting. A profession that feels threatened, a press that is both plaintiff and licensor of AI companies, and politicians who need a concrete enemy will each pick up the grievance that works best. The grievance is the vehicle, not the cause.

It helps to separate the grievances, because they are not equal. The local ones are the easiest to mobilise and the weakest on the merits:

• Noise and land use are siting problems. Many large campuses are built far from towns, in industrial or rural zones, and setbacks and zoning are well known tools. When a county has no zoning, that is a planning gap, not a property of AI.
• Water depends on the cooling design and the local basin, not on distance. Closed-loop and dry cooling remove most evaporation, at the cost of more electricity. Microsoft, for example, has designed its new AI sites since 2024 around closed-loop cooling that uses no water for evaporation after the initial fill.
• Grid capacity is the one cost that is diffuse and real. In the PJM market, which covers much of the eastern United States, data centers account for roughly 38% of capacity costs in the auction for 2028/29, and residential prices in PJM areas have risen faster than elsewhere. Capacity costs are spread across the whole grid, so a family pays even if the nearest data center is two hundred kilometres away.

The third point is the interesting one, because it has a known fix. If large loads pay for the capacity they require, the cost shift goes away. Ohio already did a version of this in 2025, when regulators required new data centers above 25 MW in AEP's territory to pay for at least 85% of the capacity they reserve, with long contracts and exit fees. Connections resumed under the new tariff. A moratorium on AI development is a much bigger instrument for a problem that a tariff can largely solve, which tells me the target is not only the bill.

Where the thesis gets uncomfortable

This reading has holes, and some of them are large. The first is the voters. A University of Massachusetts poll in August found 65% opposed to a data center nearby, and the number barely moves with education, between 63% and 71% in every group. Rural counties voted the same way, and so did many Republicans. The people at the ballot box are not an aristocracy of anything. If the idea holds, it holds one level up, among the people who write the bills, the policy papers and the editorials that tell voters what a data center means.

The second problem is that the right says the same things. Steve Bannon calls AI the most dangerous technology in the history of mankind, Josh Hawley has asked for a national data center moratorium, and in July 2025 the Senate killed the ten-year preemption of state AI laws 99 to 1. In Texas it was a Republican governor who froze new grid connections first. The mood is bipartisan. What the Democrats did differently is give it a programme, with bills, a new department and a timeline.

The press is also less monolithic than the thesis needs. Studies do find a negative framing, but most of it looks like the usual negativity bias of news rather than an editorial line. And the New York Times sues OpenAI while it licenses its archive to Amazon for training, which looks more like a price negotiation than a crusade.

Then there is the part I cannot dismiss. The Stanford Digital Economy Lab found that employment of workers aged 22 to 25 in the most AI-exposed jobs is about 19% below the trend of less exposed peers, mostly through fewer hires. The authors are careful to call it descriptive, not causal. A 24-year-old does not need a causal model to notice that the first step of the ladder is getting shorter, and whoever offers protection to these people is offering something real.

Who ends up paying

So the version I believe is narrower than the one circulating online. Nobody runs the backlash, and most people vote against a data center for ordinary reasons. What I see is a set of groups whose authority comes from scarce, certified knowledge, discovering that the household bill is a better argument than their own status.

What bothers me most is where the cost lands. The speeches name Musk, Zuckerberg, Altman and Bezos, because a billionaire is easy to vote against. A moratorium, pre-release licensing or fifty different state rulebooks do not care about names. The hyperscalers have the lawyers, the long-term power contracts and enough cash to wait. A startup that needs 30 MW next year, or a small lab publishing open weights, has none of that. Rules written against the oligarchs end up building them a moat.

The cheap fixes are known. Tariffs for large loads, published water numbers and proper siting would remove most of the legitimate grievance without deciding who is allowed to train models. I suspect the new Congress is in no hurry, because a grievance that gets solved stops being useful.

The view from Europe

From Spain the American debate has a familiar shape, with one difference. In Europe the aristocracy of expertise does not need to win referendums, because it already sits inside the regulator. Most of the AI Act has applied since August 2026, with obligations layered by risk class. This summer the EU adopted the Digital Omnibus on AI, which moves the high-risk rules to December 2027 and August 2028, officially because the standards needed to comply were not ready. My reading is that Brussels also noticed who carries the weight of compliance, and it is not the American hyperscalers.

The grid constraint is real here too. Ireland now treats Greater Dublin as fully constrained for large data centers, and new sites must bring their own dispatchable generation or storage and a plan to cover 80% of their demand with new Irish renewables. The uncomfortable part is the second-order effect. If the United States slows itself for political reasons, Europe is not in a position to take the space, because it lacks the compute, the cheap energy and the capital. We could fall behind twice, or import the same populism without ever having built the industry it claims to fight.