The European-sovereignty angle is what makes Mistral's strategy harder to read from the outside. If a meaningful share of revenue comes from EU government and defense contracts, the public benchmarks against OpenAI/Anthropic stop being the right scoreboard — they're optimizing for different procurement rules and a different definition of "trustworthy." Curious if anyone has visibility into how those contracts are structured: per-seat, on-prem, or something closer to hosted with sovereign keys?