BERLIN — Europe risks falling into a “complete trap” if it focuses on industrial AI while leaving frontier models to the U.S. and China, Mistral CEO Arthur Mensch said in an interview Monday.

On the sidelines of the Digital Summit in Berlin, Mensch said Europe needs to build its own computing capacity and advanced AI models to ensure security. He added that Europe cannot assume it will always have access to the most powerful foreign models, pointing to the risk that the U.S. and China could deploy export controls on AI.

“The idea that we can build anything meaningful without owning the foundation of it, which are the reasoning models, the brains that can operate the factories, that can help design better systems … is, I think, a complete trap,” he said. “And we’re lucky that people realize this now, that this is not a viable way.”

In recent weeks, Mensch has emerged as a leading voice against European leaders’ shift away from building frontier AI models. In September, top EU leaders advocated building applications on foreign models to reap the benefits, instead of chasing frontier models. 

His comments came as the German government unveiled plans on Monday to make Germany a “sovereign AI nation” by 2030. While industrial AI is one of the strategy’s main bets, it also calls for Germany to help build at least three European frontier AI labs by that date.

Mensch welcomed Germany’s ambition, but warned the challenge cannot be tackled at the national level alone. “There is no such thing as a national strategy on AI,” he said. “It needs to be a European strategy on AI.”

Europe’s problem isn’t a lack of talent, he argued, but a “balance sheet problem”: the continent needs the capital and computing power to train sophisticated models and build companies that can compete at scale.

The European Union’s 27 leaders will have their first dedicated conversation on AI next week at an informal meeting in Dublin. 

The competitiveness test

Mistral itself is betting heavily on industrial AI. Mensch pointed to his company’s partnerships with Airbus and ASML and said manufacturing is the French company’s biggest market, with applications ranging from designing aircraft to reducing machine downtime.

But industrial applications cannot make Europe technologically independent if others control the models and computing infrastructure behind them, Mensch said.

Building European alternatives, however, poses another challenge: convincing customers to use them when cheaper Chinese models can match or outperform their European rivals.

That tension has already surfaced for Mistral. German search engine Ecosia stopped using its models after concluding they did not meet its needs. Its CEO said that switching to Chinese models cut costs by half while improving overall performance.

Asked whether companies could be expected to compromise on price or performance for the sake of sovereignty, Mensch rejected the premise.

“There’s no such thing as a compromise in between competitiveness and quality and sovereignty,” he said.

Mensch did not directly address Ecosia’s experience. Instead, he argued that Mistral doesn’t need to lead on every benchmark, pointing to areas such as cybersecurity and geospatial imaging where he said its new Large 4 model outperforms Chinese open-weight rivals. 

He said Mistral deliberately focused on areas where it saw strong demand. 

“The quality of our technology finds a market everywhere outside of Europe, including in the U.S.,” he said. “Mistral Large 4 was a long-awaited release. It puts us very much back in the race.”

BERLIN — Germany is unveiling its plan to catch up in the global AI race, but key details remain unclear.

Germany should become a “sovereign AI nation” by 2030 and narrow the gap with leading AI hubs in the U.S. and China, according to a report drafted by the government’s AI task force and obtained by POLITICO set to be released Monday.

“The next 12 to 18 months will therefore be decisive in setting the course,” read the report, which was due to be presented to the National Security Council and also at a summit on AI in Berlin.

Germany wants to secure access to the world’s leading models while strengthening European alternatives, particularly through expanded cooperation with France and Canada, according to the report. 

“We are leveraging collaborations with like-minded partners in the EU and worldwide (especially France and Canada) to benefit the German AI ecosystem and ensure sovereign access,” the report noted. “We will also use intra-European collaborations,” such as with French AI company Mistral, “to achieve this leverage,” its authors added.

At the same time, “AI made in Germany” should stand out through its focus on industrial AI with “highly specialized models,” the report read.

“A key focus here is the development of innovative production systems to increase productivity through the use of AI,” the report noted. “The widespread use of AI and AI-based robotics can boost productivity and quality, optimize industrial processes, and improve control over complex systems.”

To help speed up the development of AI in Germany, the government plans to act as a “strategic anchor customer,” using public contracts and pilot projects to help European AI providers enter the market.

The plan comes as Germany struggles to catch up on AI while cultivating more strategic autonomy in the area. Much of the blueprint, however, lacks key details, including on financing. The measures “neither constitute funding decisions nor prejudge” future budget procedures, according to the report.

The government established the task force in June to better coordinate AI initiatives across several ministries. The blueprint proposes 18 measures such as plans to “strategically expand AI computing infrastructure and connect it as a European ecosystem” as well as “strengthening the semiconductor value chain.”

By 2030, the government wants at least 80% of large companies and 55% of small and medium-size businesses to use AI. Berlin is also aiming for 90% of federal agencies to have at least one AI application in production. AI agents are expected to prepare administrative processes “end to end” and, where legally permitted, complete them with full automation. 

The government is also working to increase the country’s AI data center capacity by 300% compared with 2025, the report read.

Jasper Bennink contributed reporting.

Mistral says its new model can afford organizations a kind of control over its cyber defenses that closed models can't.