The AI race is about power, not just models - and Europe has five percent
When a researcher quit Anthropic this week and warned that AI could cost us our lives before the decade is out, the debate immediately turned to the models: are they getting too clever, too fast? But if you look at what money and political capital are actually being spent on in 2026, a different picture emerges. The race is at least as much about owning what the models run on - and about writing the rules for who may use them.
The bottleneck is no longer clever ideas
It's physics. Chips, power, cooling, square metres. Tech companies are on course to spend around 700 billion dollars on AI infrastructure this year alone. And the money moves in a circle: when an AI company raises capital, most of it flows straight back to the handful of cloud providers that own the machines. Whoever owns the computing power in practice also owns those who rent it.
Washington has discovered that access can be turned up and down
Export controls have become an instrument of foreign policy. In January the US allowed sales of Nvidia's H200 chips to China again, with a cap, and the following day imposed a 25 percent tariff on advanced AI chips. Licences are now granted case by case.
The controls hit American companies too. On 12 June an export-control directive required Anthropic to block non-US users from the company's two most capable models - a few days after their release. When the company couldn't verify nationality in real time, it switched the models off for everyone until the control was lifted.
That's worth holding on to if you're sitting in Europe building something on top of an American model. Access can vanish with a couple of days' notice, and it needn't have anything to do with you.
The companies have stopped merely building products
Last year the US Senate voted 99 to 1 against stripping the states of their right to regulate AI. The industry resumed the fight elsewhere: in the election campaign. Super PACs funded by the AI and crypto industries have raised more than 321 million dollars this election cycle and now rival the parties' own machines.
The industry doesn't agree with itself. One wing, funded by OpenAI people and venture capital, is fighting for a single national framework that overrides state laws. Another wing, where Anthropic has put in 40 million dollars, is fighting for the exact opposite. The disagreement is about the content of the rules - not about whether it's acceptable to buy influence over them.
The resistance comes from below, not from the ministries
What's actually slowing the build-out in the US right now is neighbours and electricity bills. At least 75 data-centre projects worth around 130 billion dollars were blocked or delayed by local opposition in the first quarter of the year, and electricity prices on the country's largest grid rose 76 percent in a quarter. Data centres have become one of the heavyweight issues of the midterm elections.
Whether the data centres are actually to blame for the higher prices is disputed. Goldman Sachs expects rises in the coming years, while a working paper from the US electricity research institute EPRI suggests that up to 2024 data centres actually pushed retail prices down. Politically that matters less: voters have already linked the two.
And then Europe: five percent
This is where the numbers get uncomfortable. The EU has around 5 percent of the world's AI computing power. The US has about 75 percent, China around 15. Projections to 2031 suggest Europe stays around five.
The rest of the picture looks similar:
- Three American providers - Amazon, Microsoft and Google - held 70 percent of the European cloud market at the start of 2026.
- Europe produces less than 10 percent of the world's semiconductors, and the AI chips in use are designed in the US and made in Asia.
- European data centres cost 20-30 percent more to run than American ones, driven by energy prices, planning permission and a grid with waiting times of up to ten years.
The EU is trying to buy its way in. On 30 July the Commission opened the tender for up to seven so-called AI gigafactories, each with around 100,000 advanced processors, for a total of around 30 billion euros in public and private money. The interest is there: 77 proposals from 60 different locations in 16 member states. But the decision won't come until early 2027, and the machines come after that.
At the same time Europe's strongest card - regulation - has been weakened. The rules for high-risk AI were due to apply from 2 August this year, but in July were postponed to December 2027, and for AI built into products to 2028.
Brookings' verdict on that combination is harsh: regulation without capacity of your own isn't sovereignty, it's dependency management. You can keep your independence on paper and still lose control of the infrastructure that defence, healthcare, research and industry rest on.
So what can be done
Chatham House sets out four realistic paths for countries of Europe's size: specialise in one link of the chain and use it as leverage, align fully with a superpower in return for guaranteed access, share sovereignty with like-minded countries, or spread across several suppliers so you can switch. The question isn't whether you want to be dependent - you will be - but on whom, how much, and on what terms.
There's also a market to meet. European companies increasingly demand local suppliers, and spending on sovereign cloud in Europe is expected to triple in two years. Even Amazon has opened a separate European cloud in Brandenburg. But the willingness to pay has a limit: at a European tech conference this year only a fraction of the room would pay more than ten percent extra for European computing power.
Back to the warning
The researchers warning about AI talk about systems getting out of control. That's one danger. The other is more mundane and already under way: that a few companies and two countries come to own the layer everyone else builds on - and that the rest of us find out the day access is used as leverage.
The two worries are connected. It's the race that makes slowing down hard, and it's the same dynamic that concentrates power with those running in front.
Sources: Brookings on middle powers' room for manoeuvre, The Future Society on European computing power, Chatham House on the four paths, the Commission's AI gigafactories call.