AI dependence: how one letter exposed how vulnerable Europe is

On 12 June 2026 the US government sent a single letter. A few hours later, Anthropic had to switch off its most powerful models, Fable 5 and Mythos 5, for all of its customers. Not because of a data breach. Not because of a crash. Because of an export control measure on national security grounds, which banned access for every foreign national, both inside and outside the United States.

I was working with Fable 5 at that moment. And I noticed the difference: it had capabilities you could simply feel. From one moment to the next, that was gone. Not because I did anything wrong, but because as a non-American I fell under an American export rule.

Let that sink in. Access to frontier AI can be cut off for a European at the speed of an email, for reasons we have no part in whatsoever. That's not a technical problem. That's a dependence problem.

The dependence is measurable

Almost all the top models come from the US. According to the Stanford AI Index 2026, the United States released around 50 leading AI models in 2025, with China at around 35. Europe lags far behind and lacks a comparable frontier scale. The US has more than 5,400 data centres, more than ten times as many as any other country. And the investment gap is staggering: according to that same report, around $286 billion in private AI investment went to the US in 2025, compared with just over $12 billion to China. That Chinese figure is probably an underestimate, because a lot of funding runs through state funds. But even then, the gap remains huge.

In Europe we effectively have one real frontier player: Mistral, in France. Alongside that, a handful of smaller labs such as Aleph Alpha and Black Forest Labs in Germany. Each of them is valuable and often strong on open weights, sovereignty and EU AI Act compliance. But the bench is thin. If one player disappears or slows down, there's little behind it.

And don't forget the foundation: almost every advanced AI chip in the world is fabricated by a single company, TSMC, in Taiwan. On top of that, the accelerator stack is heavily dominated by NVIDIA. The whole chain hangs on a few non-European points. One geopolitical shock and the entire structure becomes unstable.

The part everyone underestimates: energy

You can't run AI without power. A lot of power.

AI runs in data centres. Worldwide, they consumed around 485 terawatt-hours in 2025. By 2030 that will double to around 950 terawatt-hours, says the International Energy Agency (IEA).

The driver is AI itself: the consumption of AI data centres will triple over the same period. And that pace is already visible. In 2025 alone, data centre consumption grew by 17% while global electricity demand rose by just 3%.

And this is where it pinches even harder for Europe: according to the IEA report Electricity 2026, Europe's energy-intensive industry paid on average around double the US electricity price in 2025. Bruegel and the OECD arrive at the same order of magnitude. On top of that come agonisingly slow grid connections and tight supply chains for transformers and turbines. Anyone who wants to train and run AI at scale needs not only chips but also an energy infrastructure built for it. We don't have that today.

This isn't a footnote to the AI debate. This is the debate. No energy, no AI.

Europe is moving, just not deep enough

Brussels isn't blind. Through the InvestAI initiative, €200 billion is being mobilised, of which €20 billion is for four to five AI gigafactories: colossal compute centres with around 100,000 advanced AI chips each, roughly four times the capacity of the current AI factories. The formal call for these is in preparation. On paper that's substantial and sensible.

But there's a catch, and we should be honest about it. If those gigafactories run almost entirely on NVIDIA GPUs, NVIDIA networks and a supply chain dominated by non-European players, then the infrastructure may sit on European soil, but the sovereignty doesn't. You can't buy independence by simply installing a lot of accelerators in European facilities. Sovereignty runs through the whole chain: chips, models, data, energy and talent.

So the goal isn't digital autarky. It's strategic resilience: knowing which layers are critical, where we're dependent and where we need to organise alternatives.

Why this should be a wake-up call, not a lament

The Fable 5 episode isn't an incident to grumble about. It's a warning with a date on it. It shows how narrow the margin is that Europe operates within, and how quickly that margin can vanish.

My conclusion is simple and urgent: Europe has to commit fully here, far further than what's happening now.

For organisations, this already means something concrete today: don't build your AI strategy on one model from one provider in one country. What can disappear overnight is not a foundation. Work with several models side by side, know your fallback options and treat model availability as a continuity risk, not a given.

I'm not watching this development from the sidelines. I want to be right in the middle of it and help organisations use AI wisely, safely and resiliently. Because the question isn't whether AI will shape our future. The question is whether Europe becomes a player in it or stays a customer.

One letter showed where things stand. Make sure the next letter doesn't bring your organisation to a halt.

Rob Ummels, AIAdopt. AI adoption, AI literacy and practical guidance for organisations that want to use AI wisely.