What's this about?
French Public Accounts Minister David Amiel said in a press conference that the government intends to contract sovereign AI companies such as Mistral, and specifically that its future plans exclude OpenAI. Ethan Mollick's response supplies the counter-consideration: if your sovereign models are far behind the frontier, you pay a substantial penalty for using lagging models in frontier applications — and the penalty is largest exactly where the capability gap is widest, such as cyber.
What exactly is "sovereign AI"?
It's the position that a country should run AI systems it controls — built domestically, hosted domestically, subject to its own law — rather than depending on foreign providers. The motivations are real: data leaving your jurisdiction, a foreign government able to restrict access, and the industrial argument for a domestic capability. The cost is equally real: the frontier is expensive, and a national champion is usually somewhat behind. The policy question is not whether sovereignty is valuable but how large a capability gap is acceptable, and in which applications.
Why is this worth noticing?
- Mollick's cyber point is the sharpest version of the trade-off. In defensive security you are matched against attackers using the best available tools, so a deliberate handicap is most costly precisely where it's least affordable.
- It lands in a week of related moves: the Norway-should-buy-OpenAI thread, Anthropic's IPO, and Nvidia's reduced backstop — all versions of the question of who controls this infrastructure.
- Mollick flagged that he can't confirm the report, which is worth carrying forward: this is a press-conference remark, not published procurement policy.