The zoning wall

On 21 May, the Milwaukee Common Council introduced an ordinance that would ban data centres larger than 60,000 square feet anywhere inside city limits, and impose conditional-use review on anything between 20,000 and 60,000.¹ Milwaukee is not a hyperscale market. The ordinance is not, in itself, a binding constraint on US AI infrastructure. What is interesting is that Milwaukee bothered.

A municipality that was not under siege from data-centre proposals has now legislated, prospectively, against a class of facility it does not yet host. That is the move worth paying attention to. Cities that have watched what happened in Loudoun County, in Memphis, in the Atlanta exurbs, are deciding to settle the question before the trucks arrive. Robert Bryce's tracker, which aggregates rejections and moratoria, records that 2026 has already passed the full-year 2025 total for blocked or delayed data-centre projects, with seven months still to run.² Project Censored counts 188 active community opposition groups across 17 states, and reports that 66% of tracked Q2 2026 projects were blocked or delayed at the planning stage.³

These are not isolated NIMBY incidents. They constitute a map.

The federal assumption. The Trump administration's AI Executive Order framework, and the broader policy stack around it, assumes essentially permissive local siting conditions. The model in the policymakers' heads is that capital, given a green light from Washington, will pour concrete wherever the unit economics work. The implicit geography is the one US infrastructure had between roughly 2010 and 2023: a fight here and there, mostly winnable, occasionally lost, never structural. That geography no longer exists. The gap between federal AI ambition and the actual planning-board landscape is now the binding constraint on US buildout, and it is not a transitional friction that goes away once the policy framework matures. It is the policy framework's blind spot.

To see how far the ground has shifted, consider what data-centre opposition looked like five years ago and what it looks like now. In 2020, a contested project meant one or two angry planning meetings, a local newspaper story, and, usually, an approved permit with some landscaping concessions. In 2026, a contested project means a multi-county advocacy coalition, an environmental-impact challenge filed by a state-level group, a utility-board fight over interconnection capacity, a separate fight over water rights, and, increasingly, a pre-emptive ordinance from the next county over that watched the first one happen and decided it did not want the same. The opposition has institutionalised. That is the difference.

Why this happened

The proximate cause is power. A hyperscale AI campus draws somewhere between 100 and 500 megawatts; the largest proposed builds reach into the gigawatts.² A 300 MW facility consumes roughly the electrical output of a mid-sized natural gas plant, and its waste heat is non-trivial. The water draw, depending on cooling design, can run to millions of gallons a day. None of this was true of the colocation and cloud-era data centres that communities had quietly tolerated for fifteen years. The thing being proposed in 2026 is a different category of industrial facility, and communities have noticed.

The distal cause is harder and more political. Data-centre siting decisions in 2024-2026 happened to coincide with the moment AI became culturally legible, and culturally contested, to people who do not work in technology. The facility going up next to the elementary school is no longer an abstract piece of cloud infrastructure; it is, in the imagination of the person living next to it, the building that will replace their daughter's first job. Whether that imagination is accurate is a separate argument. What matters for siting is that opposition now has a cultural reservoir to draw on that opposition to (say) a UPS distribution warehouse does not. A poll cited in Bryce's tracker put US opposition to siting a data centre in one's own community at 65%.² The number is high, and the framing, "in your own community", is the standard NIMBY elicitation, which tends to overstate organised resistance relative to passive preference. But even discounted, a two-thirds opposition baseline is a different operating environment from anything the hyperscalers' siting teams have worked in before.

The capital response

Sophisticated capital has, characteristically, repriced the problem before the policy world has noticed it changed. The most interesting recent move is not a zoning fight at all. It is the NextEra/Dominion acquisition, which is being read by infrastructure analysts as a power-availability play: rather than fight grid-interconnection battles community by community, acquire the generation assets at source.¹⁴ If you own the utility, the question of whether the local zoning board wants a data centre matters less, because the data centre can be sited where the generation already is, on land the utility already controls or can readily acquire, with permits that have already been issued for industrial-scale electrical use.

This is the bit that should be lifted out of the noise. The story is not that 188 community groups exist; it is that the people who actually allocate the capital have already concluded the distribution-grid fight is unwinnable and have moved one layer up the stack.

The people allocating the capital have already concluded the distribution-grid fight is unwinnable, and they have moved one layer up the stack.

Utility M&A as AI-infrastructure arbitrage is a different game from data-centre siting. It is slower, more capital-intensive, less competitive, and less subject to local veto. It also concentrates the infrastructure in the hands of fewer, larger, more financially sophisticated players — which is to say, exactly the players who were going to win anyway. The community-opposition story, told in the language of grassroots civic action, ends in a market structure that is more vertically integrated and more concentrated than the one that preceded it. That is worth sitting with for a moment. The decentralised democratic act of saying no, repeated across hundreds of planning boards, has the aggregate effect of pushing the industry toward consolidation. I am not sure that is what the opposition groups think they are doing, but it is what they are doing.

Some analysts read NextEra/Dominion more conventionally — as utility consolidation driven by rate-base growth and not specifically as an AI-infrastructure play.⁴ That reading is defensible on the available disclosures. The strategic-intent argument is partly inference from timing and from the broader pattern of hyperscaler power-purchase agreements. But the pattern is real even if any single deal admits an alternative explanation: the binding constraint on AI buildout has shifted from compute to power, and the capital is following.

The Virginia question

About 35% of US data-centre capacity sits in Virginia, mostly in the Loudoun-Prince William corridor.⁴ Counties in that corridor have, in the last 18 months, imposed temporary construction moratoria, revised by-right zoning to require special-use permits, and in some cases simply removed data centres from the list of permitted uses in specific districts.⁴ Some of these moratoria have been lifted after negotiated terms. Some have not.

The concentration risk is now visible and largely unpriced. A sector with a third of its capacity in jurisdictions that are actively renegotiating the terms of its presence has a single-point-of-failure problem that I am not seeing adequately disclosed in hyperscaler investor materials. The standard defence — that Virginia's economic dependence on the sector will prevent any genuinely binding restriction — is probably correct at the level of total ban and probably wrong at the level of permitting velocity. A Virginia that approves data centres at half the historical rate, with double the negotiated cost per megawatt, is a Virginia that has materially repriced the sector's growth assumptions, even if it has not banned anything. That is the regime we are now in.

The capital response to Virginia-specific risk is what the rejection tracker is, in part, capturing: a forced geographic diversification into Texas, Georgia, Ohio, Wyoming, and increasingly into the Upper Midwest, where Milwaukee has just signalled that the welcome mat is being rolled up before anyone arrives. Each new market involves a fresh round of community engagement, fresh local politics, fresh utility-board negotiations. The cost of capital for US data-centre buildout in 2026 is higher than it was in 2023, and not primarily because of interest rates.

The contrarian read

I want to take seriously the argument that local opposition is a market-correcting mechanism rather than a barrier. The case goes like this. The 65% opposition figure indicates that AI infrastructure siting has a legitimacy deficit which the early buildout has not priced. Projects pushed through over community objection face longer litigation tails, higher insurance costs, occasional operating-licence challenges, and the constant low-grade political risk of a state legislature deciding, three years in, to retroactively impose conditions. Projects that go through community-benefits negotiation may be slower to build but are more durable once built. The historical parallel is wind and solar siting in the 2000s and early 2010s, which produced a generation of permissive-then-contested-then-litigated projects, after which the industry settled into a more structured planning regime that, ultimately, built more capacity than the permissive early phase would have.³

I think this argument is half right. It is right that the legitimacy deficit is real and that pricing it in will produce more durable projects. It is wrong, I think, about timing. The wind-and-solar comparison works over a 20-year arc; the AI infrastructure question is operating on a three-to-five-year arc, and the capital is not patient in the way that renewable-project capital eventually became. The slowing-now-builds-better-later case requires a degree of patience from the people writing the cheques that I do not see in their behaviour. NextEra/Dominion is not the move of capital that has decided to slow down and build legitimacy. It is the move of capital that has decided to route around the legitimacy question entirely.

So the contrarian read is correct as a description of what should happen and probably wrong as a description of what will happen. The market-correcting mechanism is operating, but the corrected market is one in which the infrastructure gets built anyway, by larger players, with power purchased upstream of the distribution grid, with less local economic benefit, and with a citizenry that has won a series of tactical battles while losing the strategic one. I am not pleased to write that sentence. It is what the evidence says.

What CBAs are actually doing

Community Benefits Agreements are the genuinely interesting civic innovation in this story, and they are under-covered. The template is familiar from US land-use law in other sectors, sports stadiums, large-format retail, port expansions, and consists, broadly, of a negotiated package in which the developer commits to local hiring quotas, capped externalities (noise, water draw, light spill), tax-sharing arrangements beyond the standard property-tax base, and sometimes direct community-fund contributions. In exchange the community grants a conditional-use permit and agrees not to oppose specific subsequent expansions.³⁴

Data-centre CBAs are early and inconsistent. Some include local-employment commitments that the operators cannot actually meet, because a hyperscale facility employs perhaps 50-150 people at full operation regardless of how the contract is written. Some include water-mitigation commitments that turn out to be the binding constraint on the facility's design. Some include tax-sharing arrangements that local governments later discover they have under-priced relative to the facility's actual revenue contribution to the operator.

The CBA template is going to mature, and it is going to set the floor for what gets built. The interesting question for the sector is not whether CBAs become standard, they will, but what they end up requiring. If the template that hardens in 2026-2027 includes meaningful local economic-development commitments tied to the actual revenue generated by the inference workloads running in the facility, the unit economics of US data-centre construction shift permanently. If the template hardens around one-off capital contributions and standard property-tax escalators, the impact is marginal. Which way that goes will be decided in places like Loudoun County and (now) Milwaukee, by people who are not currently being interviewed in the AI trade press.

The CBA template that hardens in 2026 will set the floor for everything built after it, and it is being negotiated by people who are not currently being interviewed in the AI trade press.

The thing the federal level can do

There is a non-trivial federal preemption argument. The Trump administration's AI EO framework includes provisions for streamlining permitting on federal land and for pressure on state utility regulators, and the federal government retains significant levers — the Federal Power Act, the Defense Production Act, environmental-review streamlining — that could, in principle, override local zoning for facilities deemed nationally significant.¹ If those levers are pulled, the Milwaukee ordinance and its successors become less decisive. A federal-land siting strategy, with data centres on military-adjacent or Bureau of Land Management parcels near metro areas, would do most of what the hyperscalers need without ever entering a county planning meeting.

I think the federal-preemption move will be attempted and will work in narrow cases — particularly anything that can be framed as national-security infrastructure for defence AI workloads — and will not work at the broad scale of commercial inference infrastructure. The political cost of overriding local zoning for, in the public's perception, a Meta inference cluster is materially different from the political cost of doing it for a Department of Defense training facility. The current administration may well try; the Supreme Court may well let them; and the resulting backlash, in 2027 or 2028, is likely to be the moment community opposition becomes a national rather than a local political force. I would not bet against the federal levers being pulled. I would bet against the levers working at the scale the policy framework currently assumes.

Where this lands

The piece I have been writing here is not, fundamentally, about data centres. It is about the mismatch between the cultural and political legibility of AI in 2026 and the infrastructure assumptions that the policy and capital framework around AI is still operating on. The infrastructure assumption is that the buildout proceeds on something like the 2010-2020 cloud-buildout template, with occasional friction and occasional concessions but no structural geographic constraint. The cultural and political reality is that 65% of the country does not want this thing near them, 188 organised groups are coordinating to translate that preference into binding local law, and the most sophisticated capital in the sector has already concluded that the way to win the game is to stop playing it and buy the utilities instead.

The downstream consequences are: a US AI infrastructure landscape that is more concentrated, more vertically integrated, more dependent on federal preemption to operate at scale, more geographically diverse out of necessity rather than design, and more expensive to build than the 2023 planning assumptions priced in. The hyperscalers will be fine. The smaller competitors who were going to challenge them on the basis of cheaper inference will find that the cost-of-capital differential just widened. The communities that win their local fights will discover, over the next three years, that they have shifted the where of the buildout without much affecting the whether.

There is a version of this story I could tell that ends optimistically — CBAs mature, federal preemption stays narrow, capital learns patience, the sector ends up with more durable infrastructure than the permissive build would have produced. I have written the case fairly above. I do not, on the evidence, believe it. The evidence I see is of capital re-routing around the constraint, federal policy preparing to override it, and community opposition winning tactically while losing strategically. The zoning wall is real. It is also, increasingly, something the buildout is going over rather than through.

That is the story worth watching: not whether the next ordinance passes, but where the next utility gets bought.

Footnotes

¶