Who pays when the data centre arrives

The story being told about AI infrastructure is a story about capacity: how much compute, how fast, how close to the grid. The story Americans are increasingly telling about AI infrastructure in their own counties is a different one. It is about who pays the water bill, who hears the substation hum at three in the morning, and who was in the room when the rezoning passed.

In the first three months of 2026, Data Center Watch logged at least 75 US data-centre projects, worth roughly 130 billion dollars, blocked or delayed by community opposition. That is the largest three-month total since the group began tracking in 2023, and it is already close to the 156 billion dollars logged across the whole of 2025.¹ The number of active opposition groups has more than doubled, to 833, spread across 49 states.¹

The framing in the trade press treats this as friction. A speed bump on the way to the buildout the hyperscalers have already promised their shareholders. I do not think that is what is actually going on. A doubling of opposition groups in a year, across nearly every state, with a shared playbook circulating between them, is not friction. It is the early shape of a civic veto on where AI infrastructure gets to land.

What the host community is actually being asked to absorb

It helps to be specific about what residents near a proposed data centre are objecting to, because the framing of "NIMBY" (Not In My Back Yard, the standard dismissal of local land-use opposition) does a lot of work to make those objections sound unserious.

The costs are concrete. A single hyperscale facility can draw tens of millions of gallons of water a year for cooling, in regions where residential water rates are already rising. It pulls hundreds of megawatts off the grid, which utilities recover through rate cases that fall partly on residential customers. It produces a continuous low-frequency hum that travels further than the developers' acoustic studies tend to model. It changes traffic, sightlines, and the tax base in ways that can take a decade to settle.

The benefits, for the host community specifically, are narrower. Construction employment is real but temporary. Operating employment per square foot is famously low: a 250,000-square-foot facility may run with a few dozen staff. The largest concrete benefit is property-tax revenue, and that is the case the industry leans on hardest.

The Loudoun County argument, taken seriously

The strongest version of the pro-buildout case is Loudoun County, Virginia. Roughly 53 million square feet of data centres now sit inside the county; data-centre property taxes fund close to half its budget.¹ Schools, roads, and public services in Loudoun are, in a real sense, paid for by the racks in Ashburn.

I want to take that case seriously, because it is the only one of its kind at scale, and dismissing it would be cheap. Loudoun is evidence that a community can capture significant fiscal value from hosting this infrastructure. The fiscal benefit is not imaginary.

It is also evidence of something the industry quotes less often. Loudoun's outcome took roughly two decades to build, ran through repeated environmental and zoning reviews, and was shaped by a county government with the planning capacity to negotiate from a position of some leverage. The current wave of AI-era projects is, by design, trying to compress that timeline. Expedited permitting, pre-approved sites, and master-planned campuses exist precisely to avoid the years of review that gave Loudoun whatever bargaining power it had.

A community that gets twenty years and a planning department can negotiate. A community that gets ninety days and a fast-tracked rezoning is not negotiating. It is being notified.

The honest reading of Loudoun is not "this can be done well, therefore the current buildout is fine." It is "this can be done well under conditions the current buildout is structured to avoid."

What 833 groups across 49 states actually means

The Data Center Watch numbers deserve more attention than they are getting in the AI press. Eight hundred and thirty-three active opposition groups is not a mood. It is civic infrastructure. Food & Water Watch's March 2026 guide, "How to Stop a Data Center Near You," documents a shared playbook: how to read a rezoning notice, which utility filings to challenge, how to organise a public-comment campaign, how to find a sympathetic state legislator.² The groups are learning from each other. The 40-plus communities that have already delayed or blocked projects are teaching the next 40.

New York's legislature passed a one-year statewide moratorium on new data-centre construction on June 4, awaiting Governor Hochul's signature.¹ One can argue, fairly, that a blanket moratorium is a blunt instrument that freezes both bad projects and good ones. I think that critique is correct. I also think it misreads the politics. Moratoriums are what you get when there is no functioning forum for the consent question. They are a sign that ordinary review has lost legitimacy, not a substitute for it.

The capex story the markets are not pricing

There is a second-order point worth naming. Hyperscaler capex plans for 2025 and 2026 are in the hundreds of billions, and equity valuations embed an assumed deployment rate for that spend. A 130-billion-dollar quarterly slip in projects blocked or delayed is not a rounding error against those plans. It is a real deployment-rate ceiling that no analyst model I have seen treats as structural.

The mobility argument, that capex blocked in New York will simply route to Texas or Virginia, is partly true and partly an evasion. It is true that developers will look for the path of least resistance. It is an evasion because the path of least resistance is shrinking. Forty-nine states, 833 groups, a shared playbook. The places willing to absorb a 500-megawatt facility with no questions asked are not infinite, and they are getting fewer.

What this is, actually

Strip the politics and the question this wave of opposition is asking is narrow and old. When a piece of infrastructure imposes concentrated local costs to deliver diffuse and distant benefits, who decides whether it gets built, and on what terms? American land-use politics has been arguing about this question, in different forms, for a century. The AI-era buildout is the latest case, not a new category.

What is new is the speed at which the question is being forced. The industry is trying to install, in three or four years, the physical footprint of a generational infrastructure shift. The communities being asked to host it are, increasingly, declining to do so on the terms offered. That is not a failure of public understanding. It is a working democracy noticing that it was not consulted, and using the tools it has.

The question worth asking is not how to clear the opposition. It is what terms a buildout would have to offer to be worth saying yes to. Loudoun, on its slower and more carefully reviewed timeline, suggests those terms exist. The current buildout, by design, is not offering them.

Footnotes

¶