Mistral CEO warns Europe faces two-year window on AI dependence

Arthur Mensch told French lawmakers that Europe has about two years to avoid becoming dependent on US-controlled AI infrastructure, arguing that the next phase of the industry will be decided by access to chips, energy and compute. According to Business Insider, the Mistral AI chief described the risk as Europe becoming an AI “vassal state” if it cannot secure the physical inputs that make large models cheap to train and serve.

The warning lands at a moment when European AI policy is often discussed in terms of rules and safety reviews, while the bottlenecks Mensch points to sit elsewhere: data-centre power contracts, scarce advanced semiconductors, and the cloud capacity concentrated in a handful of American firms. In practice, a European model that runs on US cloud credits or US-managed clusters is only as sovereign as the invoice terms allow; outages, export controls, pricing changes and contract disputes are governance events as much as technical ones. The pattern is visible across the sector: startups can build impressive software layers, then discover that scaling requires negotiating for GPUs and electricity in markets where incumbents already have long-term supply and political access.

Mensch’s testimony also reframes “AI competition” as an industrial siting problem. Compute is built where grid connections can be secured, permitting is predictable, and capital can be deployed quickly; regions that treat energy as a political rationing tool tend to get fewer server halls, not better ones. The result is a dependency that is easy to miss in consumer-facing debates: European firms may keep their headquarters and branding at home while the critical production inputs—chips, power, and cloud control planes—sit under foreign jurisdiction.

Mistral itself is part of the counter-move: a European AI company trying to stay relevant in a market where distribution and infrastructure often matter more than research prestige. But the two-year clock Mensch describes is not a research timeline; it is the lead time for power capacity, procurement and construction. Once the long-term contracts are signed and the clusters are installed, the bargaining power shifts to whoever owns the switches.

Mensch’s argument to lawmakers was not about a new model release or a regulatory filing. It was about where the servers will be plugged in, and whose name will be on the meter.