The US government’s decision this month to impose stricter export controls on Anthropic’s most advanced models, the Mythos 5 and Fable 5, made Europe’s reliance on US counterintelligence providers stark. Even if the US government were to lift these restrictions soon, it is obvious that this could happen again at any time. Similarly, the growing shortage of AI computing power makes political intervention to prioritize the needs of American consumers an ever-present possibility.
Without change of course, Europe risks becoming a technological backwater cut off from its most advanced capabilities, with potentially dire consequences for its security and prosperity. soon”Europe 2031” the state of Europe’s leading AI researchers and investors strongly points to such a future.
The US government’s decision this month to impose stricter export controls on Anthropic’s most advanced models, the Mythos 5 and Fable 5, made Europe’s reliance on US counterintelligence providers stark. Even if the US government were to lift these restrictions soon, it is obvious that this could happen again at any time. Similarly, the growing shortage of AI computing power makes political intervention to prioritize the needs of American consumers an ever-present possibility.
Without change of course, Europe risks becoming a technological backwater cut off from its most advanced capabilities, with potentially dire consequences for its security and prosperity. soon”Europe 2031” the state of Europe’s leading AI researchers and investors strongly points to such a future.
Unfortunately, the most famous initiative of the largest European autonomy, EuroStack, offers a plan that is at the same time impossible and not ambitious enough to deal with the possibilities of near-powerful AI. If things go wrong, his approach will bring not only partial success but a Europe exposed to danger and eventually fully dependent. Paradoxically, the most realistic path toward greater European capacity to act is to build closer ties with American industry leaders of the current technology paradigm while reducing Europe’s strengths in industrial AI, betting on alternative technologies, and building capacity by working with other powerful nations.
Any European AI strategy needs to account for the deep uncertainty about the future direction of the technology. It’s entirely possible that AI will continue under the current paradigm of large-scale linguistic models, or LLM, to stall. If so, modern frontier labs can even enter in an interesting fashion. But if the big financial bet on the current concept pays off (with almost 700 billion dollars invested alone in 2026) and what the CEO of Anthropic Dario Amodei calls “Powerful AI“can be achieved in the coming years, it is absolutely essential that Europe ensures the availability of leading US models. Second-best models may fail to fully protect against critical cyber and other security risks.
We got a taste of this mode with the release of Anthropic’s Mythos 5 this year. Some of the statements about its capabilities may be hyperbolic, but the UK’s AI Security Institute assessment proved that Anthropic’s heavily guarded initial release was not a deep event. The lesson for Europeans must be that having the most advanced AI control does nothing to protect you from such dangers if you don’t have access to the most advanced designs yourself. In this case, the availability of a large amount of computing power is also necessary to avoid large economic losses.
The EuroStack approach does nothing to address the possible future. With the motto “Buy Europe, Sell Europe, and Fund Europe,” its vision is to nurture European service providers from applications to chips and data centers, with public procurement as a key factor. In its defense, this plan is based on a dire estimate of Europe’s capabilities. It is remove characterizes the EU as a middle power and instead claims that Europe is “a SUPER power and we need to act as one” on AI policy.
The truth begs to differ. The best LLM in Europe, the Mistral, is currently far behind the American frontier models (and also the best Chinese ones) in terms of capabilities. Even if Europe gave more financial resources to Mistral right now, it is unlikely that the company could close the gap with the top models where Elon Musk and Mark Zuckerberg have failed so far despite the amazing mobilization of resources.
Indeed, EuroStack supporters confess that “Europe will not create great models of borders but we can still create models a few steps back that will be important.” That still seems promising, given that European AI labs will be at risk of being cut off from US hardware and computing infrastructure—unless Europe also manages to create frontier chip designers and completely change its game on data center construction at the same time. Europe currently has only 5 percent of the world’s computing capacity and is falling further behind. Public-led investment in AI industries has experienced delays and is ultimately based on little EU funding, while European industry players see great promise in leaner approaches to using AI and are largely reluctant to fund large-scale data center investments.
Instead of banking on the illusion of AI superpowers, Europe should assume its role as a middle power and focus on strengthening different strengths, building capacity together with other middle powers and testing alternatives. The goal should be to increase the cost to the US of preventing access to LLM cross-border structures from Europe and to increase the incentive for US companies to push for European access while at the same time placing bets that will pay off if technological alternatives succeed.
One way to achieve this is to strengthen Europe’s strategic imperative in the global AI framework. Europe already has power and real assets: for example, ASML lithography machines, without which no high-performance chips can be made, and Siemens Energy gas plants, without which the expansion of a fast data center in the United States is impossible. The main reason why they currently translate into less efficiency is because Europe remains dependent on the United States in the military field, making it necessary to provide its security as soon as possible.
In managing the critical relationship with the United States, EU members need to build joint capacity and coordinate strategies with like-minded central powers, such as Canada, India, Japan, South Korea and the United Kingdom. This cooperation could also put the central powers on a stronger footing against the AI superpowers—the United States and China—including with a view to urgently needed international agreements on managing the serious risks of AI.
Additionally, it can be used to pool resources and capabilities where appropriate, including government capacity building. For example, the UK’s AI Security Institute is setting an example for others to follow, including Germany, which has just decided to build its own version.
By expanding Europe’s computing power, publicly funded gigafactories could be a factor serving the basic needs of government and research, but facilitating private investment will be important. As a recent research and the Carnegie International Peace exhibition, one of the most important things here is the processes that allow fast and fast grid connections. Investments should be taken from Europe itself (such as Germany’s Schwarz Group), from partners including Japan (such as SoftBank’s recent 75 billion Euro investment in France), but also from the United States as a key part of the mix. Europe should invite US executives and a coalition facing growing public opposition at home to build data centers on the continent. On the other hand, companies will have every incentive to defend European access to models running on this infrastructure. Given the uncertainty of returns on large data center investments, bringing in American investors also reduces European risk.
Finally, Europe must focus on preserving and restoring its industrial capacity, in order to build capacity and capture a greater share of the economic value enabled by AI. European policy should focus on building specific capabilities in areas of competitive advantage, such as industrial AI and robotics. These areas are also closely linked to alternative technological concepts such as global models and learning chains, opening up opportunities to leapfrog into better scenarios. To enable such progress, European regulation must support rather than hinder leadership in industrial AI. It is a good move that the EU amended its AI Law to treat industrial AI systems differently from those for wider consumer use. It needs to do the same for industrial data in the European Union Data Act.
To benefit from its industrial base, Europe must avoid selling proprietary data resources and promising innovations to foreign buyers. This requires greater mobilization of European capital, including through the integration of capital markets and the reform of the pension system. Furthermore, European leaders need to make a real case for why Europe is the best place to build technology that can basically be built anywhere. In contrast, rule-based policy is increasingly a resource in its own right, but Europe must also build a more promising domestic market by strengthening the integration of the single market for services and adopting a more structured approach to risk management.
At face value, this strategy may seem less attractive than the grand promises of full European independence. But it is the most realistic way to achieve European independence in the sense of preserving and expanding Europe’s ability to protect its security and prosperity. It is also one that offers the best chance for the continent to take advantage of the opportunities that the age of AI offers.
–
Thorsten Benner is co-founder and director of the International Institute for Public Policy in Berlin.
Jakob Hensing is head of political economy at the International Institute for Public Policy in Berlin.
