Around the world, a backlash is growing against one of the most power-hungry pieces of infrastructure behind our increasingly digital lives: data centres.
From the outside, they typically resemble giant, squat warehouses. Inside are row upon row of humming computer servers, consuming electricity around the clock and generating so much heat that huge cooling systems are needed to stop them from overheating.
As you’d expect, this takes an enormous amount of water. In 2023, US data centres used around 66 billion litres – and a single large site can get through as much water in a day as a town of up to 50,000 people.
Then there are the power demands. By the end of 2025, the US had more than 4,000 data centres with around 50 gigawatts of capacity, enough to power London more than 10 times over.
Little wonder, then, that people are starting to push back. Although the Trump administration has publicly supported the US data centre boom, 69 per cent of respondents to a recent NBC News poll said they would oppose an AI data centre being built in their neighbourhood.
But just how worried should we be about the big data centre build? What will these enormous facilities mean for the communities being asked to host them? And, with billions being bet on an uncertain AI future, how many will actually be constructed?
Home truths
For all the controversy surrounding them, having a data centre move into your neighbourhood isn’t necessarily all bad news. At least economically: these facilities can boost local tax revenues and create permanent jobs, though usually only in the low hundreds.
Developers may also pay to upgrade local substations, roads and emergency services. Some even generate their own power on-site and sink wells to access groundwater rather than relying entirely on public supplies. And, of course, the people living nearby are likely to use many of the digital services these facilities make possible.
However, for Dr Ben Green – an assistant professor of information and public policy at the University of Michigan and author of Algorithmic Realism: Data Science Practices to Promote Social Justice – the question is whether communities are getting enough in return.
In some areas, developers have signed ‘community benefits agreements’, committing them to provide money or other improvements for the places hosting their facilities. Green sees these as a step in the right direction, but says there are limits to what they can achieve.
“I still think that the actual scale of those agreements tends to be pretty paltry compared to the scale of the data centre itself,” he says.
Crucially, communities don’t necessarily get much say over where the agreement money goes. Green argues that developers can choose to fund projects that suit them, while concerns that truly matter more to local residents go unaddressed.
A developer might prefer to “throw a few million dollars into the fire department”, he says, rather than tackle something residents may care more about, such as reducing the noise coming from the data centre.
Read more:
- Why Silicon Valley is betting big on data centres in space
- We might finally know what will burst the AI bubble
- ‘Oh my god!’: 1,206 AIs built their own secret chatroom – and the weirdest conspiracy emerged
Built on a bubble
Given the extraordinary number of data centres being planned, it’s easy to be alarmed by what their combined impact could be.
However, there’s a good chance many of them will never be built. In the past few years, far more data centres have been announced than have ultimately made it to construction, while others have faced lengthy delays.
And then there’s the question of who will pay for it all. Building data centres on this scale requires huge sums of money, which raises doubts over whether companies can actually afford to deliver everything they’re promising.
The dollars involved are truly staggering. In June this year, economists Jonathan and Jessica Wachter calculated that major technology companies plan to spend more than £560bn ($750bn) on new infrastructure and equipment in the US by the end of 2026. By 2027, that annual spending could exceed £750bn ($1tr).
For that colossal bet to pay off, AI will need to generate equally colossal returns. The tech giants pouring money into data centres are betting that all this extra computing power will ultimately translate into much greater profits – whether by selling AI services, creating new products or making their existing businesses more productive.
Assuming investors expect a 15 per cent return, the Wachters calculated that these data centres would need to become 2.7 times as productive for the numbers to add up. If those gains fail to materialise, they warned, “[these] firms risk bankruptcy”.
If data centres sound like an incredibly risky bet, it gets worse. Tech companies aren’t simply paying for all these data centres with current profits, but increasingly they’re borrowing it. That means if AI fails to generate the returns they’re hoping for, companies won’t just be left with expensive data centres that haven’t paid off – they’ll also have enormous debts to repay. And if enough companies find themselves in that position at once, the consequences could ripple through the wider economy.
There’s an uncomfortable historical precedent for this. Economic historian Liaquat Ahamed has compared the AI boom to the great US railroad bubble of the 19th century, when investors poured huge amounts of borrowed money into a building frenzy that doubled America’s rail network in just seven years after the Civil War.
When the railroad bubble burst in 1873, the damage spread far beyond the investors who had bet on it, helping plunge the entire US economy into the Long Depression. Yet railroad construction had averaged around 2.5 per cent of the US economy – less than the 2.8 per cent that spending on AI infrastructure is now expected to reach.
In short, there’s every chance the company behind that data centre down your road will go bust before a single brick is laid. But if that happens, an abandoned building site may be the least of your worries.
Tomorrow’s world
There’s another reason to question the sheer number of data centres currently being planned: nobody knows how much computing power AI will actually need in the years ahead.
For one thing, more and more AI could soon run on our personal devices instead. The chips used to run AI are improving remarkably quickly, with their performance roughly doubling every two years. As a result, tasks that need powerful servers in a data centre today could increasingly be handled locally.
“The chips that Nvidia put out on laptops today are more capable than the ones four years ago, or two years ago, that required 10 times as much energy and [went] into the big servers,” explains Prof Carl Boettiger, an associate professor at the University of California, Berkeley.
It’s not just the hardware. Boettiger says AI models themselves are shrinking, performing similar tasks with far less computing power, making it increasingly possible to run them on personal devices.
That could reduce our reliance on data centres… but technological progress could also push things the other way. As AI becomes cheaper and easier to use, Boettiger argues, we’re likely to use much more of it. Overall demand for computing power could therefore grow so quickly that we end up needing more data centres, not fewer.
Green agrees we’ll use more AI – just not necessarily the kind that needs giant data centres. Much of the current building boom is designed to host the most powerful models, but smaller, cheaper ones will soon be able to handle most everyday tasks almost as well.
“Why pay such a premium for the most sophisticated model when you can, for 10 per cent of the cost, get 80 per cent of the performance?” he asks.
If enough people reach the same conclusion in future, Green concludes, many of today’s planned data centres could become ‘stranded assets’ – humming away, half-empty, long after the hype has died down.
Read more:
