For two years the limiting factor on America’s AI infrastructure was supposed to be silicon — who could secure Nvidia allocations, who could finance the next hyperscale campus. That story is going out of date. The binding constraint is increasingly a zoning board, a county moratorium, and a utility interconnection queue, and the politics behind all three are turning sharply against speed.
Start with the numbers that frame the fight. In a Reuters/Ipsos poll of 4,531 adults fielded in early June, just 33% of Americans said the rapid construction of AI data centers is mainly a good thing; 64% disagreed. Ask the same people about a facility near their own home and the asymmetry widens: 57% said they would oppose a data center in their community, and only 14% said they would be comfortable with one nearby. The opposition is bipartisan. Republicans are somewhat more supportive than Democrats, but neither party’s voters are net positive, and neither party has built a trusted message on AI infrastructure.
That is the asymmetry that should worry developers more than any chip shortage. National abstractions poll badly; specific local proposals poll worse. And the American permitting system is engineered to convert local opposition into delay.
Where the “no” gets tested
The clearest illustration of both the power and the limits of local opposition is Saline Township, Michigan. There, a planning commission and then the township board voted — 4 to 1 — to reject rezoning 575 acres of farmland for a Related Digital campus tied to the OpenAI–Oracle Stargate program, a project pegged at roughly $16 billion across 1,000 acres. The vote was unambiguous, and residents turned out with yard signs and public comment to register it.
It did not hold. Related Digital sued the township for “exclusionary zoning,” arguing Michigan law did not allow the community to bar a legitimate land use. Within weeks the township settled, approved the project under a court agreement, and secured roughly $14 million in community benefits — farmland-preservation money and environmental conditions — in exchange for letting construction proceed. Its attorneys had warned that fighting risked an expensive courtroom loss on top of losing the site anyway.
Saline is the cautionary version of a local victory: the vote was real, the political signal was loud, and the project still got built. But it also reveals the new cost structure. Developers now budget for litigation, community-benefit payouts, and months of delay as standard line items. The friction is no longer hypothetical, and it lands before a single server is racked.
Where the “no” sticks
Where local opposition is proving more durable is the moratorium — the blunt instrument that buys time without forcing a fight over a single site. Fulton County, Indiana, enacted a one-year ban in March. DeKalb County, Georgia, pushed new developments off through the middle of the year, and a state bill would bar any new data-center permit in Georgia until March 2027. Indianapolis passed a non-binding pause resolution. By Stateline’s count, more than a dozen states are weighing temporary bans, and dozens of local governments have already adopted some form of pause.
The aggregate is striking. By one industry tally, at least 75 data-center projects worth roughly $130 billion were blocked or delayed in the first quarter of 2026 alone. None of that shows up in a chip-supply model. All of it shows up in a build schedule.
Crucially, the issue is now electoral. Heading into the 2026 midterms, Democrats have found that power-price risk — the worry that data-center demand will push household electricity bills higher — is a message that travels in both rural and suburban districts. That reframes data centers from the economic-development win governors once pitched into a cost-of-living liability. Once a project becomes a campaign issue, the permitting calendar stops being a formality.
The grid is the second veto
Even projects that clear the politics hit a second gate: the wires. U.S. grid interconnection queues now stretch for years, with backlogs measured in thousands of gigawatts and median waits to reach commercial operation approaching half a decade. In the markets where AI demand is most concentrated — Northern Virginia, Phoenix, Dallas — interconnection timelines now run several years, not several months.
And the long-lead bottleneck is no longer compute. It is the unglamorous electrical hardware: transformers, switchgear, high-voltage equipment, and battery cells, much of it on order books that stretch past the data-center build cycle itself. A developer can win a rezoning, settle with the neighbors, and still wait years for a substation upgrade that does not yet exist. The scarce resource is firm power on a timeline, not floor space.
The new map of the buildout
Put the three layers together and the geography of AI infrastructure starts to make sense. Developers are increasingly steering toward rural sites precisely to bypass city-council votes and land-use scrutiny — and discovering that rural counties have learned to say no too. They are budgeting for lawsuits and community-benefit agreements where they cannot avoid populated jurisdictions. And they are competing for deliverable electricity rather than square footage.
The headline figure — one in three Americans comfortable with the national pace — is not just a sentiment reading. It is a forecast of friction. The capital is available; the chips, for now, are available. What is scarce is local consent and on-time power, and both are decided far below the level where AI strategy is usually debated. For an industry that measured progress in model parameters and capex announcements, the binding constraint has moved to the county commission and the interconnection queue. That is where the buildout actually slows.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.