Mistral CEO Arthur Mensch warns that Europe is heading down a path toward “a complete trap.” If the continent does not develop its own foundation models, it will never be guaranteed access to this technology.
Mensch is the CEO of the only European AI model developer with a somewhat competitive LLM. Mistral Large 4 isn’t the absolute strongest AI model, but Mensch points to specific lucrative capabilities such as security and spatial awareness, categories that are extremely useful for industrial and critical applications.
The fallacy, as Mensch explains, lies in wanting to leverage technologies based on foundation models without making those models the focus. Export restrictions on Claude Fable 5 in June demonstrated that the United States can drastically restrict access to the most advanced AI at a moment’s notice.
A world of dependencies
This would not have been possible with Mistral. Although it is heavily dependent on chip suppliers such as Nvidia, that company, in turn, is heavily dependent on TSMC, which cannot manufacture the chips without the machines provided by Mistral investor ASML.
Yesterday, Germany presented plans to become a “sovereign AI nation” by 2030. This involves the construction of several AI factories, essentially massive GPU facilities. But the country also spoke of its desire to host three competitive AI labs, similar to the landscape in the U.S. or China. Mensch welcomes that rather loosely formulated idea, but adds: “There is no such thing as a national AI strategy. It must be a European AI strategy.” According to him, Europe has, above all, a balance problem and not so much a shortage of talented personnel. Capital and computing power are lacking, or at least not in sufficient quantities to be competitive.
Industry as the largest market
Mistral itself is focusing on industrial AI, even though its own foundation models are suitable for a wide range of purposes. The manufacturing industry is the French company’s largest market and includes partners such as Airbus and ASML. The Dutch chip equipment company recently invested an additional 400 million euros as part of a 3-billion-euro funding round led by Samsung.
Just before Mistral regained its competitive edge thanks to the Large 4 model, news broke that the French company would have preferred not to hear. Ecosia had parted ways with Mistral and opted for the Chinese competition, which was not only cheaper but also delivered better results. Mensch did not elaborate on that example, according to Politico.
Nevertheless, there are successful new applications for Mistral, especially now that it has a stronger AI model again. Orange Cyberdefense has already gained access to a security variant of Large 4. Meanwhile, Mistral is working on expanding its European computing capacity, which is set to grow to 1 GW by 2030. “The [new model] puts us right back in the race,” says Mensch.
Is Europe really in a bind?
In the short term, there is little for Europe to fear. Even if the best American AI models are inaccessible for whatever reason, China provides a significant counterbalance with open-source options. However, this carries the same risks as adopting U.S.-based LLMs when it comes to medium-term dependence. Without its own ecosystem of AI model developers, there can be no European AI strategy that can be called sovereign.
As is so often the case in the IT industry, the question is whether such autonomy can actually be achieved. For now, it is clear that Europe needs more capacity to support multiple Mistrals. It is also clear, for the time being, that no single AI lab has a technological lead of more than a few months. Since the early days of Claude and Gemini, there have been alternatives to the initial AI frontrunner, OpenAI. This was followed by Chinese LLM providers, of which DeepSeek has been the most productive contributor to the technology itself and beyond. It regularly shares not only model weights but also methods for running AI faster and more efficiently, for improving training, and for building agents.
The democratization of this knowledge may come to an end, but it is particularly concerning if it is accompanied by a consolidation of technological progress among AI labs. So far, the reality has been that every leading, advanced AI model has been surpassed within six months to a year by a more compact, open alternative. As long as that remains the case, Europe has no more to fear in the field of AI than it does elsewhere in the technology stack.
Read also: With Mistral Large 4, Europe is (slightly) competitive again