Executive Summary
- Half of the AI data centre capacity being built today will never earn back its cost of capital, and that is exactly how transformative infrastructure has always been built.
- Britain’s data centre debate is stuck on water, planning and village views while the real bottleneck, a grid connection queue four times the size of the country’s 2030 need, goes unaddressed.
- James Buckley-Thorp says the risk for the UK is not overbuilding. It is building too little, too late, and then holding an inquiry into why.
I build AI for a living, so let me say the quiet part loudly: half of the AI data centres being built today are stranded assets. They will never earn back what they cost. And I want every one of them built.
Start with the money, because the money is deranged. The five biggest tech companies on earth will spend more than $600 billion on capital expenditure this year, three quarters of it on AI.
Amazon, Alphabet and Microsoft are spending more on capex than their entire cloud businesses earn. They are borrowing to do it: Wall Street expects the sector to raise around $1.5 trillion of debt over three years. That is not a business plan. It is a hostage situation with better catering.
And the thing they are buying evaporates. An AI chip is state of the art for about three years before something twice as fast replaces it.
The industry writes the kit down at 20 per cent a year, which means the biggest companies in the world have signed up for depreciation bills larger than last year’s profits. The power contracts assume every business on the planet wants to run a frontier model. Most of them want a chatbot that reads their emails.
Every rack is an ice cube on a warm day. Everyone cutting a ribbon knows it. They build anyway, because missing the wave ends careers and overbuilding lands on someone else’s balance sheet.
Good. This is how it is supposed to work.
Bubbles build infrastructure
We are in the biggest capital misallocation since the railway mania of the 1840s, and that is good news.
Parliament waved through thousands of miles of track that were never laid, the shares collapsed in 1847, and Britain got the railway its economy still runs on.
The dot com boom buried oceans of fibre that sat dark for years, went bust, and handed us the cheap bandwidth the modern internet is built on. Bubbles are how infrastructure gets built by people who would never build it if they understood the odds.
The AI overbuild will go bust, and the survivors will inherit the cheapest compute in history. The only real losers are the investors, the insurers and the councils who believed the brochure.
Which brings me to Britain, the only country in the race arguing about whether it is allowed to enter. Every data centre objection here comes down to water, power, and a village that wants its view.
Water
Thames Water says a big data centre can drink as much in a day as 50,000 homes. Campaigners have found that 84 per cent of the water hungry sites proposed for Britain are in areas already water stressed or heading that way.
Water UK told MPs this summer that the government’s AI Growth Zone policy does not contain a single mention of water. All true. So yes, AI will drink your village dry. And it is worth it.
Because here is the part the objectors leave out. England has not built a major reservoir since 1992. Thames Water, the company warning you about data centres, leaks more than half a billion litres a day: thirty data centres’ worth, into the ground, every day.
A country that cannot build a reservoir in thirty years and cannot fix its own pipes does not get to blame a server hall for the drought. The data centre didn’t cause the deficit. It made it impossible to ignore.
Power
This is the real scandal, and nobody in Westminster understands it, because it does not fit on a placard.
The queue to connect to Britain’s grid grew tenfold in five years to more than 700 gigawatts, four times what the country needs by 2030, most of it zombie projects squatting on a place in line.
A data centre in London can wait seven to thirteen years for a connection; in competing European cities, three to seven. Ministers have spent 2026 rewriting planning rules and announcing AI Growth Zones as though the problem were a parish council rather than a substation.
An AI Growth Zone with no grid connection is a field with a press release. You can win consent in 2026 and still be waiting for first power in the mid 2030s, by which point the chips you designed the building around are three generations dead.
The village
A country that would rather protect a field than host the infrastructure of the next century has decided to be poor. That is a legitimate choice.
But make it in daylight, and be ready to explain to your grandchildren why the compute lives in Texas and Abu Dhabi, and the salaries went with it.
Britain’s national sport is the public inquiry. We are the only country that would hold one into the future before it arrives. So build them. Build too many.
Let half of them fail and let the next generation of British companies rent the ruins for pennies. The danger was never that we build too much, too fast. It is that we build too little, far too late, and then commission a report into why.
James Buckley-Thorp is founder and CEO of Atlian, an AI platform for construction insurance, and a serial founder with two exits, one to a large insurer and one to the group behind TK Maxx. He has built companies in Silicon Valley and London and is a regular AI commentator for BBC, Times Radio, LBC, iNews and Metro.
