SpaceX has just provided a rather useful illustration of a problem that has been worrying me about artificial intelligence.
According to Reuters, SpaceX has introduced a restriction on the use of its Starlink service in Ukraine which means that the military-grade version stops working above 160 km/h. Ukrainian officials say they weren’t warned in advance, and that the change disrupted some drone missions.
There might be perfectly good technical or security reasons for the change, and we don’t yet know why SpaceX made it. But that isn’t really what interests me here. What matters is that Ukraine depends on an extremely useful technology which somebody else ultimately controls.
A few months ago I wrote about a closely related problem in Non-Proliferation, AI Edition. My worry then was that the United States might increasingly reserve its most capable AI systems for Americans and selected allies, while everybody else received weaker models, delayed access or expensive subscriptions to yesterday’s technology.
That had already happened briefly. In June, the US government ordered Anthropic to suspend access to Fable 5 and Mythos 5 for foreign nationals. The restrictions were later relaxed, but the episode demonstrated something important: Washington could decide who was allowed to use some of the world’s most capable artificial intelligence.
At the time, I was mainly thinking about what this might mean for countries excluded from the leading group. If the best AI lets an American company complete in a day what a competitor elsewhere needs a week to do, unequal access could gradually turn into unequal economic prospects.
What I hadn’t really considered was the possibility that Europe might respond by deciding that developing frontier AI itself wasn’t especially important.
That is more or less the argument discussed in a Politico article this morning. Europe, on this view, shouldn’t obsess about producing the world’s most advanced models. Instead, it should concentrate on applying AI in areas where it already has enormous strengths, such as manufacturing, healthcare, agriculture, energy and defence.
Ursula von der Leyen made the same point in her State of the Union speech last month. She used the printing press as an analogy: it was invented in Germany, but its value spread across Europe. “We do not need to be the ones who develop the frontier technology,” she said, “to be the ones who draw the greatest value from it.”
As an economic argument, this makes sense. No country needs to manufacture everything it uses. Denmark doesn’t need to design its own microprocessors before Danish companies can benefit from computers, and Europe doesn’t need a domestic equivalent of every piece of American software. Trying to become completely technologically self-sufficient would cost a fortune and would probably leave us producing inferior substitutes for many things we could simply buy.
Von der Leyen is also right that Europe has some excellent cards to play when AI moves out of the chatbot and into the real economy. Germany still has a huge industrial base, Denmark and other European countries have unusually good health data and there are European companies with deep expertise in everything from pharmaceuticals to wind turbines.
If the best model for discovering new drugs is American, I see no reason why Novo Nordisk shouldn’t use it. If Siemens can improve a factory by using a Chinese open model, that seems entirely sensible too.
The problem appears when using the best foreign model gradually turns into being unable to function without it.
Imagine, for instance, that in ten years’ time a Danish pharmaceutical company depends on an American model for much of its drug discovery. A German factory might use another one to design products, maintain machinery and control production. Governments could use them for economic modelling, translation and administration, while armed forces might depend on them for intelligence analysis, logistics and cyber defence.
At that point, access to frontier AI starts looking less like buying a useful piece of software and more like depending on infrastructure.
And infrastructure becomes politically interesting when somebody else controls it.
An American company could charge European customers more than domestic ones. The newest model might initially be available only in the United States, perhaps because regulators or the US government wanted more control over its release. Certain capabilities could be withheld because they were considered useful for cyberwarfare, biological research or weapons development. Access could also be withdrawn from a particular company or government.
I’m not predicting any particular one of these outcomes, but none seems especially far-fetched. Indeed, the more capable AI becomes, the less obvious it is that the United States will continue to regard its most advanced systems as ordinary commercial products.
If a future model can reliably discover unknown software vulnerabilities, design biological organisms or plan complex military operations, some form of export control would hardly be surprising. The logic behind AI non-proliferation becomes stronger precisely as the technology becomes more useful.
China presents the same problem more obviously. Europe should certainly use good Chinese models where that makes sense, especially open ones, but I wouldn’t want European industry or public infrastructure to depend on Beijing continuing to provide access on favourable terms indefinitely.
This is why the Starlink case is so instructive. The supplier isn’t an enemy.
Ukraine didn’t make a mistake by adopting Starlink. Quite the opposite: it was available when Ukraine desperately needed resilient communications, it has proved remarkably resistant to Russian interference and it has become extremely important to Ukrainian military operations. Refusing to use it because it was American would have been a peculiar form of strategic purity.
The difficulty is that there isn’t yet an equally capable alternative that Ukraine can switch to if Starlink suddenly behaves differently.
That, I think, is the distinction Europe needs to make with AI.
Technological sovereignty shouldn’t mean technological autarky. Europe doesn’t need to insist that every company uses a European model simply because its servers can be photographed with twelve blue flags behind them. We should buy and use the best technology available.
But we need to retain the ability to carry on without it.
Interestingly, von der Leyen herself seems to recognise this tension. Immediately before arguing that Europe needn’t develop every frontier technology, she said that Europe needed its own AI capabilities, particularly for national security: “We need to stay in the race. Because this is crucial for our independence.” She also called for much more European computing capacity and investment in European AI companies.
The Commission is putting some money behind this. Earlier this year it selected the EUROPA consortium to develop an open frontier model covering all 24 official EU languages on European infrastructure, explicitly with the aim of strengthening Europe’s ability to develop advanced AI itself.
So perhaps the real question is what it means to remain in the race.
Europe doesn’t necessarily need to produce the world’s best model every year. If OpenAI, Anthropic or a Chinese company produces something dramatically better, European organisations should be free to use it. Nor does Europe need to duplicate every American and Chinese research programme.
But it probably does need enough computing capacity, researchers, engineers and infrastructure that developing a competitive model remains possible.
There’s an important difference between having a European model which is somewhat worse than the American frontier and being unable to build one at all.
In the first situation, losing access to American systems would be expensive and inconvenient. In the second, somebody outside Europe ultimately decides what level of artificial intelligence European companies, governments and perhaps armed forces are allowed to possess.
The military case makes the problem particularly clear. It would be a peculiar form of strategic autonomy if European weapons, intelligence systems or cyber defences depended on an API which Washington could disable.
But civilian dependence could eventually matter almost as much. If AI really does become a general-purpose technology embedded throughout medicine, engineering, administration, science and industry, then accepting permanent dependence on foreign frontier models would mean accepting that somebody else controls an increasingly important part of Europe’s productive capacity.
Perhaps the best analogy is a spare tyre.
You don’t normally drive around on the spare simply because your proper tyres were made abroad. You buy the best tyres you can get and use them.
But when you get a puncture on a country road, that is a poor moment to discover that you abandoned the ability to change the wheel because buying tyres had always worked perfectly well before.
This is what worries me about the idea that Europe can specialise in applying American and Chinese frontier AI. It could work extremely well for years while quietly making it harder and harder to do anything else.
Europe should use American AI and Chinese AI whenever they are the best tools available, and it should exploit its existing strengths to apply AI throughout the economy. But underneath all this, it needs enough independent capability that losing access to somebody else’s models would be a serious nuisance rather than a strategic crisis.
The lesson from Starlink isn’t that foreign technology is dangerous. It’s that dependence becomes dangerous when there is no credible alternative.
Europe doesn’t need to make the AI it uses every day. But it does need an AI that nobody else can switch off.