Power

Never before has the thinking of so many people hung on so few companies. And almost all of them are based in the US.

In June 2026, an order from Washington forced an AI model to be shut down worldwide, three days after its launch. Europe could only watch, and whether it comes back, nobody knows.

A handful of companies supplies the systems that hundreds of millions of people now write, research and decide with. Almost all of these companies are based in the US. The language models, the cloud they run on, the chips they compute on: at every one of these points, a few American corporations set what is possible and what is not. Infrastructure of this importance has never been concentrated in so few companies.

How far from abstract that is just showed itself with Fable 5. Anthropic introduced the model as its strongest, and three days later it was blocked. Not because of a bug, not out of a decision by the company. The US government issued an export rule that barred foreign users from access. Because such a rule can’t be enforced for individual user groups in a shared cloud service, Anthropic switched the model off for everyone, worldwide, paying customers included. Hours passed between the order and the global shutdown.

Anyone in Europe who had built this model into their work, their product or their research lost access from one moment to the next, because of a decision taken in Washington that Europe had no part in. There was no office where you could lodge an objection. The dependency is not a risk that might materialise someday. It is already here. Someone else works the switch.

And it runs deeper than the models. The data centres Europe’s AI runs on belong mostly to American providers too. The chips, without which no model computes, come overwhelmingly from a single US company. Even data that physically sits in Europe falls under American law the moment a US corporation runs the servers. Europe rents almost every layer of its AI, and the landlord sits on another continent.

That does not leave Europe powerless. The most advanced chips in the world can only be made with machines that a single company builds worldwide, and that company stands in the Netherlands. The tightest bottleneck in the whole chain is European. That is a real lever. But it is a single one, against a dependency that exists on many levels at once and cannot simply be dissolved by building a European model of our own.

The point here is not whether the US is a reliable partner. The point is that a dependency this deep can be cut from the outside at any time, and that the decision about it is not made in Europe. The Fable 5 case is just the one that became visible. The same possibility hangs over every model, every cloud, every service that someone here depends on.

We need to talk about this

What does this topic stir up in you?

In depth

What happened on 12 June 2026

Anthropic introduced Fable 5 on 9 June 2026, its most capable publicly available model. Three days later, on 12 June at 5:21 pm Washington time, Anthropic received an export-control directive from the US government: nobody without US citizenship, in the wording of the order no "foreign national", was allowed to keep using Fable 5 and its sister model Mythos 5, worldwide, inside the US as well as outside it. The government gave national security as the reason. Anthropic read the background this way: the authorities believed they had found a method to get around Fable 5's safety guardrails and reach the cybersecurity capabilities of Mythos.

Because a citizenship-based block can't be enforced cleanly in a shared cloud service, Anthropic switched off both models for all customers, including paying ones. The other models stayed available. The company called the order a misunderstanding and said it was working on restoring access; there was no timeline for it. How long the block lasts, and whether the model returns in this form, is open.

How deep the dependency runs

The concentration affects not only the models, but the entire chain beneath them. The leading systems come from a handful of US companies, alongside a few Chinese labs. Europe's only frontier lab of any standing, the French Mistral, raised 1.7 billion euros in September 2025, led of all firms by the Dutch chip-equipment maker ASML, which became its largest shareholder. That still leaves Mistral funded many times more thinly than OpenAI, Anthropic or Google.

In the cloud, around 70 percent of the European market goes to three US corporations: Amazon, Microsoft and Google. For the chips that AI is trained and run on, a single US company, Nvidia, holds a market share of somewhere between a good 80 and over 90 percent, depending on how you count. Each of these layers is highly concentrated on its own. Stacked on top of one another, they make a supply chain that runs almost entirely through a single country.

The data too: the CLOUD Act

It's not only about compute, but about data. The US CLOUD Act of 2018 obliges providers based in the US to hand over data on a lawful order from US authorities, regardless of where the servers physically sit. If data is held in a Frankfurt data centre run by a US corporation, US authorities can demand it be handed over. Legally, that collides with the GDPR.

In practice, according to the providers' transparency reports, this lever is rarely pulled so far. What matters is that it exists, and that it hangs on American law rather than European law. Whoever controls the servers settles the legal question, not whoever produced the data.

The one lever Europe holds

Europe is not entirely without a counterweight. The most advanced chips in the world can only be made with machines that use extreme ultraviolet light, EUV for short, to print the finest structures onto a chip. These machines are built by a single company worldwide: ASML of the Netherlands. Nikon and Canon dropped out of the technology more than a decade ago, and nobody else has ever brought it to production.

An EUV system consists of over a hundred thousand parts from thousands of suppliers, the result of decades of development and billions invested in research. So the tightest bottleneck in the entire AI supply chain sits in Europe. But it remains a single lever against a dependency that exists on many levels at once. And there have long been outside designs on ASML, too.

Sources