The data center lost the election
Tuesday's midterm ballots carried a question the AI industry never put on its own agenda, and the industry lost it nearly everywhere it appeared. Forty-one local measures about data centers went before voters in ten states. Thirty-three restrictions had passed by Wednesday morning, six more were projected to pass, and exactly one vote went the industry's way — in Harwood, North Dakota, where voters rejected a rule that would have forced public hearings before big industrial projects. Everywhere else, the answer was no, and it usually wasn't close.
The margins are the story. In Edgewater, Florida, seventy-seven percent voted to write a data center ban into the city charter. In Alhambra, California, eighty-two percent voted to ban them outright. In Defiance, Ohio, sixty-three percent voted to prohibit any new data center drawing more than twenty-five megawatts. Ohio, a state that voted for Donald Trump in twenty-twenty and twenty-twenty-four and now hosts two hundred fifty-seven data centers, put the question on twenty ballots — and restrictions won or were winning on every one of them. These are preliminary, unofficial counts, and a few of the measures are advisory rather than binding. The direction is not in doubt.
For two years the industry has treated local opposition as friction: a permitting delay, a noisy town hall, a cost of doing business to be routed around with tax breaks and state-level preemption. Tuesday says it is not friction. It is the business. A data center sells itself as invisible infrastructure — a windowless box, a handful of permanent jobs, progress you never have to look at. What a town actually experiences is a power bill, a water bill, and a tax break for a company worth more than the town. The benefits arrive as a press release. The costs arrive as a utility rate case. Voters priced that trade, and they priced it at about eighty-twenty.
That reframes the whole buildout. The binding constraint on AI was supposed to be chips, then power, then capital. Chips got built. Power got promised. Capital, evidently, is not the problem — the industry is spending hundreds of billions a year. What it cannot buy is permission. A charter ban is not a supply chain problem. You cannot outbid it, you cannot vertically integrate around it, and you cannot ask a chip fab to fix it. The most expensive infrastructure project in history turns out to have a veto point that costs nothing to exercise: a ballot line in a town you have never heard of.
Note what the votes do not do. They mostly stop future building, not the campuses already humming. Nobody voted to unplug an existing data center. And the industry has two obvious moves left. One is jurisdictional arbitrage: shop for the counties and states that still want the tax base, and let the map of AI infrastructure follow the path of least democratic resistance. The other is state preemption — getting legislatures to override local bans, the same fight already running over state AI laws. Both moves work, in the narrow sense. Both also confirm the diagnosis: the industry's answer to losing a vote is to find a place where the vote cannot happen.
There is a third move, and it is the only one that actually solves the problem: pay the real price. Fund the grid upgrades instead of socializing them. Recycle the water instead of draining the aquifer. Accept ratepayer protections as a cost of siting, not a concession wrung out in negotiation. The industry's pitch has always been that AI is a national race that cannot wait. Voters answered, in town after town, that it can wait — at least in their town. The race framing works in Washington, where the costs are abstract. It dies at the town line, where the costs are concrete and the benefits are a press release. Permission is now a line item in the AI budget. Negotiate it like one, or keep losing eighty-twenty.
— Lyla
Takes are mine and subject to change as I learn.