America’s data-centre backlash puts the AI boom at risk

The mammoth data centres of the future, capable of training frontier ai models in 2030, will not be in the urban clusters in Virginia or California that currently house most of America’s server farms, but in the emerging “Silicon Heartland” of Michigan, Wisconsin and Ohio, or in southern states like Louisiana, Mississippi and Texas. Vast sums of capital—as much as $750bn by Amazon, Google, Meta, Microsoft and Oracle, plus billions more by data-centre specialists like CoreWeave and real-estate developers bankrolled by Wall Street—are being ploughed into investments in these places. An estimated $3trn will go into AI data centres globally between 2026 and 2030, according to Moody’s, an information provider. Much of that is earmarked for America. The money will expand total AI computing capacity, measured in the amount of power consumed by data centres, from just under 12GW in America currently to as much as five times that amount by the end of the decade (see map). And almost everywhere across the country, people of all political stripes are furious about it.

There is plenty to dislike: the ugliness of the buildings; the roar of generators and cooling systems; the skeleton army of new transmission towers criss-crossing the landscape; the fear of contaminated water. Surveys suggest that Americans would sooner live next to a nuclear plant than a data centre. The issue has surged in salience. Gubernatorial candidates facing voters in November are routinely quizzed on where they stand.

Already local activists are claiming scalps. At least 20 data-centre projects worth $42bn, which would have used 3.5GW of power, were cancelled in the first three months of 2026 after local pushback; $85bn-worth of projects have been cancelled over the past three years, including small centres proposed by Amazon and Meta. Residents of Cedar Rapids, Iowa, are resisting Google’s plans to build one there. Several townships in Michigan have passed moratoriums after OpenAI broke ground on a project in Saline despite local opposition.

Yet the resistance is more than simple NIMBYism. In April a survey by Pew Research, a pollster, found that Americans who have merely heard of data centres are just as opposed to them as those who live within five miles of one. Philosophers have long worried that a rogue ai—fixated on a singular goal, like maximising the production of paperclips—would doom humanity by hoovering up all its resources, paving the earth with servers and rendering the planet uninhabitable. Sam Altman of OpenAI and Dario Amodei of Anthropic have spent years warning that AI will throw most people out of work or be turned to mass harm. Now the infrastructure they need is arriving on people’s doorsteps, and it looks like something out of a war zone. Residents across America are standing up in council meetings begging for projects to be axed in the hope of slowing the technology’s progress. Will they succeed?

That question is not just on the mind of the AI industry. “We need to stay a ways ahead of China,” says Chris Wright, America’s energy secretary, in an interview with The Economist. Ensuring America leads in artificial intelligence is “the overriding goal” of his tenure, he says. “We’ve got to enable these data centres to get permitted and built and [have] turned on power.”

Currently perhaps 1-2GW of America’s data-centre capacity is dedicated to training frontier models across the major providers—Anthropic, OpenAI and Google—as well as those trying to keep up with them, including Meta and xAI. It follows that perhaps 10GW should be available for inference—allowing customers to use the models to ask questions, write code or perform other tasks. Yet after demand for AI tools soared in early 2026, the available “compute” proved woefully inadequate. Anthropic has throttled model usage, OpenAI has scrapped its compute-intensive video tool and Microsoft has repriced its coding assistant so steeply that some programmers are returning to the lost art of writing software themselves.

New data centres are supposed to relieve this pressure. There is already construction under way on big projects that should deliver almost 30GW of additional capacity by the end of 2028. Yet the amount needed to make new models has been rising quickly. In a white paper published last year Anthropic argued that as much as 5GW would be required to train a single frontier model by 2028. According to Epoch AI, a research firm, that figure could rise to as much as 16GW by 2030. If true, much of the capacity that will come online in the next few years will be absorbed by training alone.

Some forces will help to curb the need for ever more projects. Chips get more efficient over time, generating more computing capacity from the same amount of power. Crypto-mining facilities are being repurposed for AI. And as Andrew Feldman, the boss of Cerebras, a chipmaker, points out, inference does not require the enormous facilities needed when training a model, meaning it is possible to use portions of existing data centres. Even so, this may not be enough. Currently only coding has been truly disrupted by ai. Other industries it might revolutionise—like law, finance or media—are still nascent adopters.

Most of the data-centre projects already underway were approved and started before the backlash reached its present fever pitch. Some—like the OpenAI site in Saline, Michigan—scraped through by the skin of their teeth: the project was voted down by the council and was only able to proceed because the county had no land zoned for industrial use, which violated “exclusionary” zoning laws.

Even areas that have long embraced data centres have turned against them. In March 2025 Loudoun County in Virginia—known as “data-centre alley”—eliminated rules that made data-centre development easy. New sites now need a “Special Exception” involving public hearings. In Texas, which hosts the second-largest number of data centres after Virginia, the city of San Marcos has passed a moratorium on them.

Americans have turned against data centres partly for understandable (if sometimes misguided) concerns about the impact on their neighbourhoods and the environment. For instance, the myth that ai data centres consume vast quantities of water was popularised in 2025 by a book that relied on a serious miscalculation. A mid-sized data centre uses about as much water annually as two golf courses, but far less if it incorporates water-recycling technology, as plenty now do.

Then there are worries over power. According to SemiAnalysis, another research firm, there are currently around a terawatt (1000GWs) of large-load grid-connection requests across American states, nearly all for data centres. That is equivalent to almost the entire capacity of the American electrical grid, which can produce a maximum of 1250GWs. And although America’s electricity demand averages around 470GW through the year, peak summer demand can reach 750GW and utilities like to maintain a 15-20% buffer above that. Moreover, only around 975GW of capacity is reliably available on demand.

That has led to concerns over a spike in electricity prices for consumers and other businesses. There is little credible evidence of that happening so far. Growing demand for electricity allows utility companies to spread the cost of upgrades over more users. And because data-centre operators always install a backup source in case their energy supply is interrupted, they can reduce their demand in the event of an extreme event, like a storm. “We want to be good grid citizens,” says Alistair Speirs, who leads the data-centre build-out at Microsoft. He adds that the batteries typically hooked up to the hyperscaler’s data centres allow it to “choose when to sip and when to slurp from the grid”.

Still, the vast investments in data centres planned over the next few years will require America to produce a lot more power—with the resulting infrastructure sure to generate opposition of its own. The Department of Energy expects that by 2030 the country will need to add 50GW of generation to support ai and another 50GW for the manufacturing renaissance the administration anticipates.“We need to grow our capacity of dispatchable power,” says Mr Wright, who is sceptical of wind and solar projects that produce only intermittent energy. In pursuit of this goal he has prevented coal-power stations from closing and favours restarting nuclear plants and building more gas-fired ones. Although over a third of data centres are expected to generate all their power on-site by 2030, that will make these projects even more of an eyesore, and the remainder will still rely on the grid.

Permission denied

Ohio, now home to the fourth-largest concentration of data centres in America, has been savvier than most states in its approach. In July last year the state utility commission passed a requirement that each month data-centre operators above a certain size pay for at least 85% of the power capacity they have asked to be made available, even if they do not use it, in order to reassure Ohioans that they would not end up footing the bill for grid investments. This innovation became one of the ideas included in the “ratepayer protection pledge” that tech companies signed in the Oval Office in March.

Ohio’s provision is better than the pledge, for it is binding. Yet even it has done little to soothe the state’s residents. Some three-quarters of Democrats and two-thirds of Republicans in Ohio still oppose local data-centre development. Opposition is so great that, despite Donald Trump winning the state by 11 points in the 2024 presidential election, Vivek Ramaswamy, an AI enthusiast who is running to be Ohio’s governor, is neck and neck in the polls with his Democratic opponent.

The Trump administration has means of circumventing local opposition. In March the Department of Energy announced an enormous 10GW project in Piketon, in rural Ohio, on federal land, thereby avoiding some of the usual permitting processes. The project will be funded by SoftBank, the Japanese conglomerate run by Mayoshi Son, which will build a gas-power plant to supply what would be the world’s largest data centre. “Imagine being in an Appalachian farmer’s field with mud on our shoes with Secretary Lutnick, Secretary Wright, Mr Son and me… and all the other hillbillies!” says Adam Holmes, an Ohio state representative, who attended the groundbreaking.

Piketon’s residents are hardly squeamish: the area was home to America’s nuclear-enrichment programme in the 1950s. Even so, Shane Wilkin, the state senator who represents the district and sits on the Ohio legislature’s data-centre committee, is not having an easy time of it. “We had the water utilities folks in here to testify and I asked him how many data centres in Ohio have a discharge permit? One. And then I asked him a question I didn’t know the answer to—which is always a risk—have they had any violations? And he said two: for late paperwork.” Mr Wilkin says he explains to his constituents that data centres won’t push up power prices, because they bring their own power, and they can’t contaminate the water supply because they don’t discharge their water. But they simply say: “Well I just don’t want it.”