A 3-2 vote in Coweta County, Georgia, approved a hyperscale campus of more than 4.9 million square feet running behind Laura Beth's house. Her Facebook group opposing it, "Stop Project Sail," now has more than 8,000 members. That is the shape of the constraint nobody underwrote.
Every other bottleneck in the AI buildout is a supply problem. HBM shortages clear when Micron and SK Hynix add capacity. CoWoS packaging clears when TSMC builds another line. Advanced-node scarcity clears in quarters, because money and time eventually produce the wafer.
A county board vote is not a supply problem. There is no second source for a permit.
A county board vote is not a supply problem. There is no second source for a permit. Consent is a veto, and vetoes don't scale.
You cannot price around a Nashville mayor. Freddie O'Connell is preparing to wield a blocking measure the city council passed last week to kill a $700 million data center next to the Nashville Zoo. No capex line item overrides that.
This is the asymmetry that repricing risk missed. Silicon and memory are constraints that respond to spending. Consent is a veto, and vetoes don't scale.
For most of the buildout, planners could treat community opposition as a soft externality—a public-relations cost, offset by local-investment sweeteners and rural siting. That framing is now obsolete, because the resistance has a number attached.
Roughly 40% of contested data center projects were canceled over the past five years, per Heatmap Pro. Today, a project challenged by local activists has about a 50% chance of dying.
Read that again as a capital allocation fact: money committed to a contested site is now buying a coin flip. A contested data center has roughly the same odds of cancellation as a contested wind or solar farm—the technologies whose deployment curves were bent by exactly this kind of local resistance.
The trend line is the alarming part. More than 50 data centers have been canceled this year after local pushback, more than twice as many as in all of 2025, with eight killed in July alone. The rate of projects facing cancellation is rising over time.
Underneath that number sits a wall of law. More than 500 counties and municipalities now actively restrict or block new data centers, counting only the most severe constraints—outright permit bans, impossible noise limits, prohibitive setbacks. The overwhelming majority were enacted this year; nearly 190 since June 1 alone.
Peter Freed, a founding partner at Near Horizon Group and Meta's former director of energy strategy, told Heatmap the pace of moratoriums is "accelerating."
And the public mood that feeds it has inverted with startling speed. Opposition to a nearby data center ran at 42% last summer, per Embold Research; it now sits at 71%. Heatmap's polling clocks a 49-point swing against supporting a data center in one's area.
A Gallup poll reportedly puts opposition to AI data centers generally at 71%, while a June Reuters poll found only 14% of Americans are comfortable with a data center built nearby. Whatever the instrument, the answer is the same, and it is not partisan—Texas and New York, Republican counties and Democratic ones, all appear on the ban list.
Here is where the siting layer stops looking like a soft constraint and starts looking like a discount rate.
Elsewhere in this issue you'll read about $2 trillion in commitments and multi-decade offtake leases—the financing scale, the counterparty mechanics, the duration bets. Set all of that aside. The point that matters here is narrower and sharper: those long-dated leases are being signed against a siting probability that has measurably worsened.
Duration risk and consent risk are two faces of the same bet. A fifteen-year offtake agreement assumes the megawatts get sited on schedule. The siting layer now says there's a coin-flip chance a given contested campus never gets built at all.
That mismatch is the story. The capital markets are underwriting a decade-plus of certainty on top of a permitting layer that can't guarantee the next twelve months.
The certainty problem gets worse when you look at how unevenly the veto is distributed. This is not a national permitting regime; it is a patchwork of jurisdictions, and the win rate swings wildly by geography.
In Michigan, over 70% of contested data center proposals were eventually canceled. In Indiana, 56%. In Texas—despite Governor Abbott's own approval moratorium—only 17% of contested projects have died.
A dollar of capex is not the same dollar in Coweta County as it is in Pecos. The same project, same sweeteners, same noise mitigation, faces a four-fold difference in cancellation odds depending on which county board hears it. Siting risk is now a function of zip code, and there is no hedge for zip code.
It is not a coincidence that Michigan runs hottest. It is a midterm battleground with a Senate seat and all 13 House seats up this fall, and local officials are listening to constituents accordingly. Politics, not physics, is setting the win rate.
Which brings us to the trap developers are quietly walking into. Many of the hundreds of ban ordinances are written as temporary—set to expire in the coming year. That reads like relief: wait out the election cycle, and the moratoriums lapse.
But the wind and solar precedent says otherwise. Heatmap notes that many restrictions on wind and solar were also initially described as temporary. Officials renewed them year after year, to avoid re-litigating a controversial issue.
A "one-year pause" is an option the local government holds, and it costs them nothing to keep exercising it. Broomfield, Colorado passed an 18-month pause; Prince George's County, Maryland a two-year hyperscale pause; Reno already extended its moratorium into 2027. The direction of renewal is not toward expiration.
So the base case for a contested site is not "delayed twelve months." It is "possibly dead, possibly deferred indefinitely, definitely a coin flip." That is a fundamentally different input than the schedules the offtake was signed against.
The hyperscalers are not passive here—but their tools are precisely the ones that don't work on this constraint. They have White House access, evangelism platforms, rural-siting strategies, behind-the-grid power, and millions in local-investment offers. Amazon, Microsoft, and Google each still pulled large projects in Arizona, Wisconsin, and Indiana after sustained pushback.
Meta CFO Susan Li called the infrastructure environment "dynamic and uncertain" on a recent earnings call. That is corporate for: we can no longer forecast where we're allowed to build.
The megawatts are real. The demand is real—North American vacancy sits at 0.9%, most capacity through 2027 is already sold. None of that is in dispute.
What's in dispute is whether the physical footprint can land where the capital assumes it will. A 37-building campus near the Manassas National Battlefield was canceled this month after sustained protest. The 4.9-million-square-foot Coweta project passed 3-2—one vote from the other outcome.
Chips get second-sourced. Memory gets financed. A town council's noise ordinance does neither, and—given that officials have renewed wind and solar restrictions year after year—it may not expire on the schedule your lease assumes. That is the chokepoint, and it is the one line item no war chest can buy its way past.