Towns Have Blocked $130 Billion of AI Data Centers. The Bottleneck Was Never the Chips.
For three years the AI infrastructure story was a story about GPUs. Who had them, who was waiting for them, how many a hyperscaler had on order. Allocation was the moat, and the companies that secured supply early were treated as if they had secured the future.
That framing is now out of date. The constraint that decides where AI gets built in 2026 is not silicon. It is the electrical interconnect, and increasingly, the county zoning board.
By one running tally, local resistance had blocked roughly $130 billion of AI data center projects by July of this year. Not delayed. Blocked. Those are capital commitments that had chips, land, and financing lined up and still did not get built, because the community on the other side of the fence decided the trade was not worth it.
Why the town hall turned
The pushback is not irrational, and founders who dismiss it as NIMBYism are misreading the room.
PJM Interconnection, the grid operator serving about 67 million Americans across thirteen states, runs an auction that sets what utilities pay to guarantee future capacity. That price bottomed at $28.92 per megawatt-day in late 2022. It hit $269.92 in mid-2024 and peaked at $333.44 in the December 2025 auction. Those costs flow into retail bills.
Nationally, Goldman Sachs reports electricity prices rose 6.9% in 2025, more than double headline inflation, and expects them to keep rising through the end of the decade, with data centers accounting for roughly 40% of demand growth. Bloomberg's analysis of areas near data center clusters found wholesale costs up as much as 267% over five years.
The evidence is genuinely contested. An EPRI working paper found data centers actually lowered retail rates through at least 2024 by spreading fixed grid costs across more load. A study commissioned by the Data Center Coalition found no clear evidence they are the primary national driver of residential increases. Both can be true: the effect is real in constrained regions like PJM and overstated in regions with spare capacity.
But voters do not experience national averages. They experience their own bill, a new substation on the road they drive, and a facility that employs forty people after construction. As NPR put it this week, there is still no agreement on which grid upgrades are "for" a data center and which are "for" the public, so in practice everyone shares the cost. That is a political problem, and political problems get solved at the ballot box in November.
The shortfall is structural
Goldman projects a U.S. data center power shortfall of about 9.3 gigawatts this year, widening to 45 gigawatts by 2028. For scale, a gigawatt is roughly a nuclear reactor. You cannot procure your way out of a 45-gigawatt gap with a purchase order. Transmission lines take seven to ten years to permit and build. Gas turbines are back-ordered. Small modular reactors are a 2030s story.
This connects directly to the capex argument we made in $725 Billion of AI Capex This Year. Capital is not the scarce input anymore. Power is. And unlike capital, power cannot be raised in a quarter.
What this means if you are not a hyperscaler
Most of the founders we work with will never site a data center. They should still care, because the power constraint shows up in their business in three places.
Inference prices stop falling as fast. The assumption baked into most AI product plans is that the cost per token falls roughly an order of magnitude every year or two. Some of that comes from better models and chips, which continue. But if the marginal megawatt gets more expensive and harder to find, the floor under inference pricing rises. Plan for costs that decline, but more slowly than the last three years taught you to expect.
Capacity becomes regional and allocated. We are already seeing providers prioritize large committed customers when capacity is tight. If your product depends on low-latency inference in a specific region, or on bursty capacity at peak, assume you will be the first customer throttled. Build for graceful degradation and keep a second provider wired in.
Efficiency becomes a feature again. For a while, the dominant strategy was to throw the largest model at every problem and let falling prices bail you out. That stops working when prices plateau. The teams that route easy work to small models, cache aggressively, and measure tokens per outcome will have a margin advantage that compounds.
The opportunity hiding in the constraint
Every hard constraint creates a market around it, and this one is unusually large.
- Grid software. Interconnection queues are managed with spreadsheets and PDFs. Tooling that shortens the study process by even a few months is worth a great deal to anyone waiting in line.
- Flexible load. A data center that can curtail during peak hours is dramatically easier to permit. Orchestration that makes AI workloads interruptible, moving training and batch jobs to off-peak windows, turns a liability into an asset for the grid operator.
- Community economics. The projects that get approved will be the ones that bring something to the town: tax structure, local jobs, paid-for grid upgrades, heat reuse. There is room for companies that structure and verify those commitments.
- On-site generation and storage. Behind-the-meter power is moving from niche to default for new builds, and the financing and operations layer around it is immature.
The honest version
The AI buildout is not going to stop. Demand is real, and the money is already committed. But the era where the only question was "can you get chips?" is over. The next era will be decided by people who understand utilities, permitting, and local politics, which is to say the least glamorous parts of the stack.
The companies that win it will not look like AI companies. They will look like infrastructure companies that happen to serve AI. That has historically been a very good business to be in.
