Announcing a data centre takes nothing more than a press release. Building one takes land, electricity, permits, a lot of money, and neighbours who agree.
In the United States, the second list is starting to catch up with the first.
A build-out without precedent
According to an investigation by the Financial Times, US spending on AI infrastructure is expected to exceed $10 trillion between 2025 and 2032, based on estimates by a Columbia University economist. That is roughly 3.6% of the wealth the country produces each year, far more than the great infrastructure waves of the past: railways, motorways or electrification.
But these projects are increasingly running into reality.
The figures that worry
47 projects. That is the minimum number of large data centres delayed, abandoned or blocked in the United States since the start of the year, according to a specialist firm, which describes it as a conservative estimate.
35%. That is the share of announced projects that analysts at the bank Bernstein consider genuinely credible.
More than 1,000. That is the number of US municipalities that have tried this year to introduce a moratorium or restrictions against these facilities, according to the Financial Times.
The money is starting to seize up
Investors are growing more wary. CoreWeave, a company that rents out computing power and counts Nvidia among its shareholders, has seen its share price fall by around a quarter in a week, and its bonds offer yields above 15%, a sign that lenders are demanding a high risk premium. Loans tied to a huge Oracle campus in New Mexico, intended for OpenAI, are trading below their value after permit delays.
One industry player's assessment, reported by the newspaper, sums up the mood: it is much harder to finance a data centre today.
In the 1840s, the United Kingdom experienced a veritable railway fever. Hundreds of lines were authorised, shares soared, everyone wanted to invest. Then, in 1847, the bubble burst and ruined many savers. Yet the tracks built during that period remained, and they formed the backbone of the British rail network for more than a century. The lesson holds for AI: a financial bubble can burst and still leave very real infrastructure behind.
Why it is sticking
Electricity, first. Connecting a giant data centre to the grid can take years.
Residents, next. Noise, farmland, electricity bills: as in this Virginia county with its 250 data centres, local resistance is organising.
Doubt, finally. Last week the International Monetary Fund ranked a fall in AI-related stocks among the major financial risks of the moment. Lenders, for their part, are starting to ask questions about return on investment.
France is not immune to this gap between announcement and reality: we recently showed that French projects reserve far more electricity than they consume.
What to take away
This is not the end of the AI boom. Data centres are still being built, and demand remains strong.
But it is the return of physics to a story told so far with software and promises. AI needs concrete, cables and electricity. And those cannot be ordered with a single click.