The number that should worry this industry is not 50%. It is 28%.
A POLITICO/PF Global survey circulated in the final weeks before the midterms reported that opposition to a data center within roughly three miles of one's home had climbed from 28% to 50% — a 22-point jump on a single question. I have spent enough of my career reading audit reports to distrust a move that large on a single instrument. A shift of that magnitude usually means the question changed, the panel changed, or the framing changed. Treat the figure as directional; treat the methodology as unverified.
But here is what actually concerns me: the mechanism the poll points at does not require the poll to be accurate. The hard numbers underneath it — interconnection queues, transformer lead times, capacity auction clearing prices — are independently verifiable. And they all say the same thing.
For fifteen years, the binding constraint on AI compute was the fab. CoWoS packaging allocation, HBM supply, export controls — that was the queue everyone watched. That constraint has not disappeared, but it is no longer the first one you hit.
The first one you hit is the grid, and the grid runs on a clock nobody in this industry controls. A transmission line — siting, environmental review, right-of-way acquisition, construction — takes five to ten years in most US jurisdictions. A hyperscale data center takes one to two. That three-to-five-fold mismatch is the physical root of everything that follows. It is not that there is not enough electricity. It is that the electricity cannot arrive in time to serve a building that will be finished before the substation is permitted.
The second bottleneck sits in equipment. Large power transformers and high-voltage switchgear now carry lead times in the two-to-four-year range, and capacity cannot be conjured with capital, because building the plant itself takes another two to three years. Then there is price. PJM's capacity market has cleared at levels roughly nine to ten times its historical low, near the offer cap. That number reaches residential bills. When a voter sees the increase on a statement, no white paper is needed to establish causation. The bill is the most efficient political mobilization instrument ever built.
So the constraint has migrated: from chips to power, and from power to permit. That migration changes the nature of the problem, not merely its cost.
Consider what a data center project now is. It used to be site selection — pick the cheapest power, the best fiber, the friendliest tax regime. It is now a gauntlet: county planning commission, state public utility commission, state legislature, federal permitting. Each is a node with veto power, and each can be tripped by a single motivated neighbor. Tracing the hidden vulnerabilities in the code is what I do professionally; the difference here is that the code is a zoning ordinance, and the exploit is a public comment period.
The consequence is a deliverability premium. Capacity that is energized, permitted, and under a signed power purchase agreement is a different asset class from capacity that has capital and GPUs but no interconnection date. A two-tier market is forming in real time: A-grade capacity with a queue position, and B-grade capacity with a business plan. The B-tier's valuation premium will compress, because its risk is no longer technological.
One element the coverage misses entirely: water. In Arizona, West Texas, and Utah, the first objection in an arid basin is usually not electricity — it is the cooling loop. Electricity dominates the narrative because electricity has a price tag and water does not. That asymmetry means the water fight arrives later and with far less warning.
There is also an architecture consequence I think about daily, because my own work sits on the demand side of it. When siting becomes the bottleneck, the physical form of compute splits. Training stays centralized — it needs the fabric, the bandwidth, the colocation of tens of thousands of accelerators. Inference and proof generation do not. A ZK prover cluster is far more portable than a training cluster: it tolerates latency, it schedules, and below roughly 10MW it sits beneath the nuisance radius that triggers the objection in the first place. Political friction is quietly pushing capital toward edge inference and distributed proving — not because it is technically optimal, but because it is permittable. Quietly securing the layers beneath the hype has always meant accepting that deployability beats elegance.
Then there is the identity shift nobody has priced. Facing queue delays, hyperscalers are going behind the meter — gas turbines, fuel cells, on-site generation. That solves the interconnection problem and creates a larger one: the technology sector becomes a holder of generating assets and a regulated emissions source. It walks into the climate litigation arena it has spent two decades avoiding. The transition from electricity consumer to emitter carries political consequences far beyond a zoning hearing.
Nuclear brownfield co-location is the opposite trade. Existing sites carry decades of accumulated social license and environmental review, and that scarcity is systematically undervalued. Redefining what ownership means in the digital age applies here too — the scarce asset is not the GPU, it is the interconnection agreement.
Here is where I part company with the consensus reading.
The industry is treating this poll as collective bad news. It is not. It is a redistribution. Every incumbent holding an approved permit, an energized substation, and a signed PPA gains scarcity rent precisely because the marginal new entrant's cost just went up. Opposition is a moat for whoever cleared the gate first. The coverage describes a sector-wide problem; it actually describes an internal transfer of value that nobody has marked to market.
Second, the poll measures the wrong thing. Opposition on a "within three miles" question and support on a "local tax revenue and jobs" question can come from the same respondent in the same sitting. Question design is itself politics, and conflict framing wins the headline. Meanwhile the genuinely silent group is not the opponents — it is the deindustrialized counties actively bidding, offering abatements and fast-track review. They are invisible in a survey that only asks about proximity.
And third: a poll does not deny a permit. A planning commission does. Nobody has published the conversion rate from objection to actual denial, and until someone does, that 22-point figure is a sentiment indicator, not a project pipeline forecast. Building trust through rigorous, unseen diligence means refusing to confuse the two.
Watch state legislatures, not Congress — that is where the binding votes are. Watch the next capacity auction clearing price. Watch whether transformer lead times begin to converge; if they do not within a year, the deliverability premium becomes structural rather than cyclical. And watch the water question, which has not yet been asked.
The uncomfortable possibility for everyone building in this sector: the ceiling on AI compute may end up set not by lithography, but by a county supervisor up for reelection.
