The Biggest Constraint on AI Is a Zoning Board
In the first three months of 2026, 75 major data centre projects worth about $130 billion were delayed or cancelled, partly because of organised local opposition. On dollars that single quarter came close to all of 2025, and on project count it passed it. The constraint on AI capacity that everyone models is chips and capital. The one now holding up build-out at that scale is a land-use vote.
This is the least glamorous risk in the entire AI supply chain and the one with the least sophisticated modelling behind it. Compute forecasts are built from fab capacity, power purchase agreements and capex schedules. Almost none of them carry a variable for whether a county board will grant the permit, which is now the step where the projects are dying.
The number a capacity planner should be looking at
The figure comes from Data Center Watch and was reported by TIME in July 2026: 75 major projects, roughly $130 billion, delayed or cancelled in Q1 2026 alone. Data Center Watch put the whole of 2025 at $156 billion across 48 projects, so one quarter of 2026 nearly matched a full year on value and exceeded it on count. Brookings cites the same source in its analysis of the data centre backlash.
The rate of change is the part that matters. Going from 48 affected projects across a year to 75 in a quarter is not a background condition, it is a trend, and it is arriving during the period when most announced AI capacity is supposed to come online. TIME also reports, citing The Information, that more than 300 cities, towns and counties now have bans or moratoriums on hyperscale data centre construction.
The vetoes are moving up from town to state level
A municipal moratorium is an inconvenience that a developer routes around by picking the next county. A state-level one is a different class of problem, and 2026 is the year they started appearing.
| Jurisdiction | Action | Date |
|---|---|---|
| New York | First statewide moratorium by executive order, pausing permits for facilities using 50 MW or more | 14 July 2026 |
| Texas | Governor ordered an audit of every data centre in the interconnection queue; ERCOT paused its large-load queue notifications | 3 August 2026 |
| Maine | Legislature passed an 18-month moratorium on facilities above 20 MW; the governor vetoed it and the override failed 72 to 65 | 24 April 2026 |
| Monterey Park, California | Ban approved by 88% of voters | June 2026 |
| Seattle | One-year moratorium | June 2026 |
| Holyoke, Massachusetts | Ban on facilities exceeding 12 MW | 2026 |
The trend predates the 2026 numbers, which is part of why it looks durable rather than seasonal. Data Center Watch counted 6 data centre project cancellations in 2024 and 25 in 2025, before the quarter above. What changed in 2026 is the scale of the dollars involved and the level of government at which the decision now gets taken.
This is not a fringe position, which is what makes it durable
The easy dismissal is that this is a vocal minority. The polling says otherwise, and it says so consistently. A Reuters/Ipsos poll in June 2026 found 33% of Americans approve of the current pace of data centre construction and 14% are comfortable with one being built near them. Gallup, fielding in March 2026, found 71% opposed to a new AI data centre in their own neighbourhood, 48% of them strongly. On 18 July 2026 there were 142 protests across 42 states.
The grievances are mundane and locally verifiable, which is exactly why they work politically: noise, water consumption, electricity prices, air quality, land, and a job count that rarely justifies the footprint. None of that requires anyone to hold an opinion about artificial intelligence at all.
Tom Wheeler frames it at Brookings as a fight over “power and whether it will be exercised by democratic institutions or those that control the infrastructure of AI”. The tell is that the concentration argument is now being made from both ends of the US political spectrum, by Josh Hawley on the right and Elizabeth Warren on the left. Opposition that both parties can claim does not get lobbied away in one cycle.
“We are watching a handful of companies assemble a concentration of capital, information, and political power without precedent in the American experience.”
Sen. Josh Hawley (R-Mo.), quoted by Tom Wheeler at BrookingsThat bipartisan quality is what separates this from a normal siting dispute, and it is why the industry’s political spending is unlikely to solve it. The same money that fought state AI legislation, which we covered in the AI super PAC fight, is not well shaped for a thousand separate county hearings.
What this means if you are building on that compute
Most teams reading this do not build data centres. They rent capacity inside them, which makes siting risk a supply question rather than a political one.
- Announced capacity is not contracted capacity. A roadmap that assumes a region comes online on schedule is now carrying a permitting assumption that nobody wrote down.
- Regional concentration is the risk. Capacity clusters where power is cheap, and that is precisely where the grid-cost politics are sharpest. Texas pausing approvals pending a grid audit is a supply event, not a news item.
- The jurisdictions with settled rules gain relative value. This is the unglamorous case for planning around EU data sovereignty, where the constraint is a known regulation rather than an open question about a county vote.
- Self-hosting changes the shape of the exposure, not its existence. The self-hosted versus cloud API decision usually turns on cost and control. Siting politics adds a third axis, because hardware you own sits in a building somebody had to get permission to run.
For smaller teams the practical read is simpler and mostly reassuring. The workloads in a realistic SMB AI budget and in a private AI pilot run on hardware that already exists. The exposure sits with anyone whose plan depends on capacity that has not been built yet.
Why it matters
The AI industry spent 2026 arguing about model capability, safety thresholds and regulation, all of which are contested at national level and none of which stopped a single build. Meanwhile $130 billion of planned capacity stalled in one quarter at the least sophisticated layer of the stack, decided by people voting on what gets built next to their houses.
One honest caveat: the Data Center Watch figure attributes these delays only partly to organised opposition, and some of those projects would have slipped on financing or interconnection queues regardless. Untangling the two is genuinely hard, and anyone claiming a clean causal number is overstating what the data supports. The direction is not in doubt, and the direction is what a plan has to survive.
Compute has always been treated as an engineering input that money eventually solves. It is turning into a political input that money does not reliably solve, resolved locally, slowly, and by people with no stake in the roadmap.
Related reading
- AI labs bought both sides of the regulation fight on why political spending has not settled the argument.
- EU data sovereignty for AI in 2026 on building where the rules are already written.
- Self-hosted AI versus cloud APIs on where your inference actually runs.
- The open-weight window is closing on the other constraint nobody priced in.
- Securing self-hosted AI infrastructure on what you take on when you bring it in-house.
- A private AI pilot, week by week on starting with the hardware that already exists.