There are two photographs from the same industry, taken about six years apart.
In the first, a county judge and an economic development director hold oversized scissors in front of a beige building. The press release says “technology campus.” The local paper runs it above the fold, because a county that lands one of these has beaten four other counties that wanted it.
In the second, a high school gymnasium is full at nine o’clock on a Tuesday night. There’s a line at the podium. Someone has made signs.
Same industry. Frequently the same company. What happened?
The answer everyone reaches for first is scale, and scale is real. A data center in 2018 was often 20 or 30 megawatts on forty acres in an industrial park. A modern AI campus can draw a gigawatt — the output of a nuclear reactor — across two thousand acres. That is not the same land use. Calling both of them “a data center” is like calling a corner store and a container port both “commerce.”
But scale alone doesn’t explain the reversal, because communities routinely welcome very large things. They compete for auto plants and refineries and distribution hubs that are bigger, dirtier and louder. Something else changed, and it’s worth being precise about what, because the diagnosis determines whether any of it is fixable.
Here’s my list, from six years of standing in those rooms.
1. The cost stopped being invisible and started appearing on a bill
For nearly two decades, U.S. electricity demand was essentially flat. Efficiency canceled out growth. Utilities forgot how to build, regulators forgot how to approve, and — critically — nobody’s rates moved much.
That ended. Capacity prices in the largest grid in the country went from roughly $29 per megawatt-day to $329 in two auction cycles, with that market’s own independent monitor attributing the spike “almost entirely” to data center load. Gas turbines are up around 195% since 2019. Transformers are up 45 to 95%, with lead times past three years.
Whatever you believe about causation — and the retail-rate evidence is genuinely contested, with serious studies pointing both ways — the political fact is simpler. Residential prices in data center hubs rose 12 to 16%, people noticed, and they had a theory about why.
Nothing radicalizes a community faster than a monthly bill with a number on it. Water and noise are arguments. A utility bill is evidence.
2. The bargain got worse while the pitch stayed the same
Early data centers were sold as part of a technology-jobs story, and in the mid-2010s that story was at least adjacent to true — these were smaller facilities, often near existing tech employment, in metros that captured real ecosystem effects.
The ratios have since collapsed. Industry figures run 20 to 30 permanent staff per 100 megawatts, and roughly 80% of the sector’s jobs and labor income come from construction — real work, well paid, gone in eighteen months. The best-identified study, comparing counties that got facilities against counties where one was announced then cancelled, found a typical county gaining 100 to 200 permanent jobs with wages unchanged. Metro counties capture measurable gains; non-metro counties show essentially none.
That last finding matters, because the industry moved to non-metro counties. The economic story got weaker exactly as it started being told to the audience it fit least.
Meanwhile the pitch didn’t change. Communities kept hearing “800 jobs” for a facility with 50 permanent positions and 750 construction workers. People can do that arithmetic, and when they do it themselves rather than being told, they don’t just discount the number. They discount the messenger.
3. The tax exemptions outlived their justification
Most state sales-tax exemptions for data center equipment were adopted in the 2000s and 2010s to attract a young, mobile industry that might not otherwise have come.
They are still running. They now apply to the wealthiest companies in the history of capitalism, and in many states they’re statutory and automatic on certification — meaning no local body votes on them, and no local body can decline them.
Then GASB 77 arrived and required school districts to disclose their abatement losses in annual financial reports. Suddenly the number wasn’t an abstraction. It was a line item in your district’s audited financials, printed next to the year they cut the bus route.
An incentive designed to attract a fledgling industry, still paid to a trillion-dollar one, disclosed in a document your PTA can download, is not a policy dispute. It’s a fairness dispute — and fairness disputes don’t settle at a negotiated price.
4. Siting moved from industrial parks to somebody’s view
The first generation clustered where the fiber and the substations already were, which usually meant industrial land with industrial neighbors. Nobody organizes to protect the character of a place already full of warehouses.
Saturation pushed the industry outward — to farmland, to exurban counties, to places with no industrial history, no data center in the code, no prior exposure and no reason to have thought about any of this. The adjacent property owner stopped being another logistics tenant and became a homeowner, or a family that has farmed the section for four generations.
It also brought noise into contact with bedrooms. Cooling fans run 70 to 85 dBA at fifty feet, attenuating about 6 dB per doubling of distance — survivable at industrial setbacks, brutal at rural ones. And the low-frequency, tonal, never-stopping character of the sound is exactly what standard ordinances are worst at capturing, so facilities pass their noise tests and generate complaint files anyway. Which teaches residents that the rules don’t protect them.
5. The acquisition playbook stayed the same while the stakes multiplied twentyfold
Shell LLCs. Project code names — a color, an animal, a mineral. Site-selection consultants standing in for unnamed end users. NDAs presented to local officials before a single public word is spoken.
None of that was invented for AI. It’s standard practice in competitive land assembly, it has real commercial logic, and at 20 megawatts on forty acres nobody much cared.
At two thousand acres next to an elementary school, it reads as a conspiracy — because from the outside, it is functionally indistinguishable from one. When a public records request revealed that the shell company behind one Indiana project was Google, the concealment itself became the organizing narrative. Not the megawatts. Not the water. The fact that they hadn’t said. The company withdrew minutes before the decisive vote.
The tactics didn’t get worse. The stakes got large enough that the tactics became the story.
6. Your county’s first data center fight is not its first data center fight
This is the change I think gets underrated most.
In 2018, a community facing a proposal was starting from zero. No template ordinance, no model conditions, no comparable case, no one to call.
Today there are national networks, shared document libraries, model moratorium language, sample cross-examination questions, and organizers who have done this in four other counties. A community that learns about a project on Monday can have a coalition, a name, a Facebook group and a list of the twelve questions the applicant can’t answer by Friday.
The industry professionalized its land acquisition. The opposition professionalized faster, and it did it in public where everyone could copy it. That asymmetry — one side’s playbook is proprietary, the other side’s is open source — is a bigger deal than most developers have internalized.
7. And it became the one place you get to vote on AI
Briefly, since I’ve written about it before: you can’t picket a model or file an injunction against a training run. A data center has an address, a rezoning application, and a hearing on Tuesday. For a public carrying real anxiety about AI, it is the only physical object the abstraction ever produces — and the most damning arguments available are supplied by the industry’s own leadership, about extinction risk and half of entry-level white-collar work.
You cannot ask a community to host the infrastructure of a technology whose builders warn about it publicly and then be surprised at the reception.
The part I’d push back on
“Almost universally hated” overstates it, and the overstatement is doing work.
Public comment runs roughly four to one against proposals of nearly every kind, everywhere — that’s the baseline, not a signal about data centers. And when a community actually wants one, it can be loud about it: Jay, Maine spent two years working to redevelop a blown-up paper mill as a data center, said yes in writing at the town, county and chamber level, and then had to be rescued from a statewide moratorium by a governor’s veto.
Sometimes a town weighs it and says yes. That yes deserves the same respect as a no, and the national coverage almost never finds it.
What actually reverses this
Look back at the seven. Only the first two are about the technology. The other five — the exemptions, the siting, the concealment, the asymmetry in preparation, the proxy war — are about conduct and process, and every one of them is a choice.
Which is the genuinely hopeful reading. If communities hated data centers because data centers are inherently intolerable, there’d be nothing to do. They mostly don’t. What they’ve learned to hate is being surprised, being rushed, being handed a jobs number that doesn’t survive division, and being told an exemption their school district is paying for is not up for discussion.
That’s a much easier problem. It’s also, so far, a much less popular one to solve — because solving it requires showing up early, telling the truth about power and water, putting the benefits in writing with teeth, and siting where you’re actually wanted.
The scissors and the pitchforks are both photographs of the same choice, made differently.