Harwood

On Tuesday more than forty local ballot measures asked American voters about data centers. Every one went against the industry except one, in a North Dakota town of 940 people. Harwood isn’t the exception. It’s the rule, showing from the other side.

Three hundred and twenty-five to one hundred and five.

That’s the vote in Harwood, North Dakota, ten miles outside Fargo, on a charter amendment that would have required extra review and public hearings for large industrial projects. Harwood said no, and in doing so became the only place in the country on Tuesday that voted the data center industry’s way.

Everywhere else went the other way.

what happened

CNBC counted 41 data center measures in ten states. Business Insider counted 45 propositions. On either count the result is the same: voters sided against the data centers in all but one. Ohio had 21 of them on its own. The others ran from California to Maine. Depending on the town, voters banned construction, halted every planned facility, or put new limits on what could be built. CNBC put the share of voters on the anti-data-center side at 77 percent. In Joliet, Illinois, Capital B News reported close to 80 percent voting to restrict them.

It wasn’t just ballot measures. In NBC News exit polling, 56 percent of voters called data centers and AI a crucial or major factor in their vote. In Texas, Gina Hinojosa campaigned on a data center moratorium and unseated Greg Abbott. Capital B estimates the tech industry outspent the citizen campaigns about a hundred to one, and Elon Musk alone put in at least $474 million. It didn’t help. Musk’s response afterward was that data centers in space were “looking pretty appealing right now.”

The industry heard it. Meta’s president, Dina Powell McCormick, told the Wall Street Journal’s tech conference on Wednesday that starting new US build-outs today “will be very challenging,” and that every hyperscaler is now asking, “Do we need to build outside the United States?” Her diagnosis was blunter than most: “The visible sort of manifestation of AI is a data center, and it’s something that people can say, ‘I don’t want this in my community.’”

the one yes

So what was different about Harwood?

Applied Digital is already building a $3 billion AI data center there, on 160 acres, with Oracle as the planned tenant. According to Business Insider it’s expected to bring 200 jobs to a town of 940. The mayor, Blake Hankey, argued against the amendment on the grounds that its language was imprecise and its requirements “would have been financially burdensome to a city our size.”

Not everyone agreed. One supporter of the amendment told Business Insider it took 21 days from announcement to construction, and that many residents didn’t know what was being built. Another said the money had decided it. They may be right. But the money is the point.

Mancur Olson explained this in 1965, in The Logic of Collective Action. Small groups with a lot at stake per head organize. Large groups with a little at stake per head don’t, because no single member gains enough to bother. That asymmetry, concentrated interests beating diffuse ones, explains more policy than ideology does. It’s why a subsidy worth millions to twenty firms survives when it costs everyone else a few cents.

The data center industry has assumed for years that it was the concentrated interest. It had the money, the lawyers and the state economic development offices, and the opposition looked scattered. Tuesday showed it had the arithmetic backwards.

A data center’s costs are concentrated, and its benefits are diffuse. The utility bill, the water draw, the noise and the diesel backup generators all land on one county. The benefit, faster answers and cheaper inference, gets spread across a few hundred million users who will never vote in Joliet. In Olson’s terms the neighbors are the small group with a large stake per head, and the users are the large group that never shows up. The neighbors organized, and they won 40 times out of 41. “Almost every time,” Jamil Davis of Black Voters Matter told Capital B, “utility bills, data centers [are] the key things that people bring up.”

Harwood is the same arithmetic run with the benefit moved into town. Two hundred jobs among 940 people is a concentrated benefit, and the people receiving it live in the same place as the people paying the costs. Olson predicts that town votes yes. It did, three to one.

what this predicts

If that’s right, the industry’s options are narrower than its spending suggests.

It can go where the benefit is already local: towns like Harwood, places where the plant is the employer. It can leave, abroad or, if Musk is serious, to orbit. Or it can do the harder thing and move the benefit to where the cost is, so the county that hosts the load also gets the cheap power, the tax base, or the jobs, in writing and in amounts people can feel. A press release about national competitiveness isn’t a local benefit. It’s a diffuse one, which is exactly the problem.

What it can’t do is outspend the problem. A hundred to one wasn’t enough.

the county you live in

Here’s where it comes home, because you’re running the same arithmetic at work.

When you use AI to get through more work, the benefit is diffuse. Your output goes up a bit, your team’s throughput goes up a bit, and the company books the gain. The cost isn’t diffuse. It lands on whoever has to check the result. In an October survey of 2,052 US workers who use AI weekly, Founder Reports found that 57 percent of managers had had to fix a colleague’s work that relied too heavily on AI. Sixty percent of respondents said overseeing and fixing AI output takes at least as much effort as training a new hire. Half named quality control of AI output as the most important emerging skill.

Your reviewer is your county. They get the noise and the water bill. You get the jobs.

And reviewers organize the way Ohio did. It starts with a rule that big generated diffs get split up, then a required checklist, then a team norm that nobody says out loud but that everyone has noticed: your pull requests take longer to get approved than they used to. That isn’t hostility to AI. It’s 77 percent voting no on a cost they bear and a benefit they don’t share.

So be Harwood. Put some of the benefit where the cost lands:

• Review your own output first. Hand over less to check, and say what you checked.
• Make it small. A change someone can read in ten minutes gets a yes. A change the size of seventeen football fields gets a hearing.
• Give the time back. If the tool saved you a day, spend an hour of it on someone else’s review queue.

None of this is about using less AI. It’s about not being the plant that the whole county votes against.

The benefit has to live where the cost does. Harwood got the jobs, and it said yes.

Sources: CNBC: voters rebuff data centers in every ballot measure except one ↗ · Business Insider via Yahoo: the North Dakota town that voted yes ↗ · Capital B News: Joliet, spending gap, Jamil Davis ↗ · Gizmodo: Powell McCormick at the WSJ Tech conference ↗ · Yahoo Finance: “I don’t want this in my community” ↗ · KESQ/Stacker: Founder Reports survey on overseeing AI work ↗ · Olson, The Logic of Collective Action (1965), Wikipedia ↗