AI on the Ballot

Summary

The setup. Democrats won both chambers of Congress in a landslide. Thompson grants that Iran and the economy drove much of it, but what stood out to him, watching Brewers games in Wisconsin all fall, was that the most common campaign-ad topic was data centers, and both parties ran against them, each accusing the other of being pro-data center. In Texas, Democrat Gina Hinojosa won the governorship on a pledge to impose a data-center moratorium in a state holding more than a quarter of planned US capacity through 2030; her Republican opponent, Greg Abbott, had already frozen new data-center permits in September. Nearly every local anti-data-center ballot measure won (he notes North Dakota as the one exception so far).

The data center veto. He re-quotes his own earlier argument: data centers built for the general internet never drew this opposition, so the backlash can’t be separated from AI, a technology pitched as taking everyone’s jobs in the best case and ending humanity in the worst. His contrast with globalization is the key idea: when factories moved to China, affected towns had no say. Data centers require local permission, so people who had no voice in globalization suddenly have one over AI, and they are using it.

The thesis. The constraints on AI keep moving: first compute (it ramped), now memory (it will ramp), then power (fixable). The last constraint is political, and it is the hardest, because it also blocks solving the others. Many AI theories assume intelligence becomes a commodity; Thompson thinks scarcity may last much longer than expected.

Ten takes on what a durably impaired ability to build would mean (he stresses he’s not declaring it permanent):

1. The AI trade already priced some of this in, alongside tightening credit markets.
2. Neoclouds are squeezed: already-permitted sites become scarce assets, but neoclouds need new permits and have less political pull than hyperscalers.
3. Hyperscalers win share, and political opposition gives them cover to moderate spending.
4. Nvidia is a mixed bag: fewer data centers and weaker neoclouds (which it backstops) hurt short term, but when power and sites are the limit, tokens per megawatt is what matters, which favors Nvidia’s efficiency lead over hyperscaler ASICs.
5. Advantage China, which has abundant power and can brute-force with less efficient chips (Huawei’s CloudMatrix 384).
6. Data centers in space look more plausible: if opposition is bipartisan and durable, the “stick” for orbital compute is real, which he had flagged in his SpaceX S-1 analysis.
7. Frontier labs benefit: the public is effectively forcing the pace of the R&D arms race to slow, efficient models matter more, and Democrats seem more likely to pass laws that favor frontier labs over open models. Counterpoint: costlier compute strengthens right-sized open models, and labs stay dependent on hyperscalers.
8. Apple gains: on-device inference matters more, a slower AI transition helps the incumbent of the last paradigm, and a slower buildout could ease the memory crunch.
9. Startups suffer: less cheap compute, slower capability gains, and more time for the labs to lock up end users.
10. The human factor: users lose the “Web 2.0” scenario of near-free compute, but if intelligence commoditizes more slowly, society gets more time to adapt.

The close. Citizens often want the ends while opposing the means, and the demand usually wins eventually. The difference now is that tech must do the convincing and stop assuming it gets the benefit of the doubt.

Notable quotes

“People who didn’t have a say in globalization are suddenly finding they do have a say about AI.”

“The final constraint is political, and that may be the most imposing constraint of all.”

“Maybe this is a break we needed; what is increasingly clear is it’s the break Americans wanted.”

“AI wasn’t technically on the ballot, but it was implicitly, and will be forevermore.”

How it connects

• Ben Thompson · Stratechery — author and publication.
• Data Centers · Compute · Neocloud · Edge Computing — the infrastructure layer the argument is about.
• Nvidia · Jensen Huang · Huawei · Apple · Anthropic · SpaceX — the companies in the ten takes.
• Techlash — the political backlash against tech, now aimed at AI’s physical footprint.
• Same-day midterm reads: $500 Million Bought Elon Nothing (Charlie Warzel on Musk’s spending and the data-center votes), The Election (Marginal Revolution) (Tyler Cowen), The Midterm Postmortem (Curtis Yarvin).