Mistral's wager on sovereignty is now yielding returns, though with some caveats.

      On Wednesday, Arthur Mensch joined a G7 working lunch in Evian alongside Sam Altman, Demis Hassabis, and Dario Amodei. Just days prior, the US government had mandated Amodei’s company to restrict access to its most advanced models for foreign nationals, leading Anthropic to implement a global ban.

      For the single European founder present, the timing couldn't have been more opportune. Mensch, the CEO of Mistral, has spent two years cautioning that such an occurrence was possible. Now that it has happened, the internet has turned his argument into a joke revolving around an imaginary large cat.

      The joke centers on 'Le Chaton Fat,' a fictional Mistral ‘frontier model’ that has gained significant attention in tech circles. It doesn’t exist. As confirmed by the French media outlet Numerama, it began as an elaborate running gag on Mistral's subreddit after the company renamed its Le Chat chatbot to Vibe, a change users despised.

      From there, it escalated: fake benchmark charts depicting the cat outperforming OpenAI and Anthropic, a mock EU notice prohibiting it for being ‘too heavy to regulate’, and imagined specs of ‘1000 meows per second’ and ‘maximum chonk’. Ethan Mollick from Wharton joked about expecting inquiries from corporate clients regarding ‘Mistral’s new ginormous cat model.’ Replit’s Amjad Masad joined in, while Mensch himself quipped: ‘It’s actually le gros chaton.’

      While lighthearted, the meme also reveals something important. It gained traction because it reflected a real sentiment: with American models suddenly able to be turned off, a European alternative that can’t be revoked became the punchline everyone opted for.

      Mensch articulated this argument long before it became amusing. At London Tech Week in 2025, he warned about US companies ‘having the keys’ to their models. ‘At some point, you need to be able to turn it off or on, and you don’t want to depend on another country,’ he stated. Last month, while speaking at France’s National Assembly, he warned that Europe had ‘two years’ to develop its own AI before becoming permanently reliant.

      The Anthropic shutdown transformed that concept into a stark reality. Mistral’s proposition of open-weight models that clients can operate on their own infrastructure—free from foreign government control—shifted from mere rhetoric to a compelling procurement argument. This week, Mensch reinforced the message on LinkedIn, declaring that Mistral’s mission is to keep AI ‘outside of centralized control exercised by states or corporations,’ framing the stakes using the terminology of the last century's crucial resource: ‘AI, akin to oil in the 20th century, is set to become the primary source of leverage and power in the world.’

      Europe's institutions are paying attention. The European Commission referred to the incident as ‘another illustration of why Europe must bolster its technological sovereignty.’ France took significant steps, dropping Palantir from its intelligence agency and committing to provide every civil servant with a Mistral-powered assistant. Moreover, Mistral has inked a five-year agreement to integrate AI into the nation's nuclear operations.

      The financial backing is also on the rise: Mistral is reportedly in discussions to secure around €3bn at a valuation of €20bn, nearly double its worth from nine months ago, with chip-tool giant ASML among its supporters.

      Talent is moving as well. This week, Mistral announced Brian Hall, a marketing executive with experience at both Amazon Web Services and Google Cloud, as its new chief marketing officer. His announcement emphasized its significance: he thanked ‘Anthropic and the US government for clarifying why Mistral is in such an exciting position.’

      However, there is a crucial point the meme neglects. Being a model that cannot be switched off does not equate to being the best model, and Mistral does not hold that title. It still falls short compared to Anthropic and OpenAI in terms of frontier capability, user numbers, and valuation—last valued at about €11.7bn against Anthropic’s roughly $965bn. Mensch acknowledges this reality himself.

      ‘Today, we do not yet possess the best language models,’ he said this week, though he mentioned that Mistral has ‘constantly narrowed that gap’ and promised a new open-weight model this summer, with early access in July. Thus, while sovereignty addresses a control issue, it does not resolve a quality one.

      A more concerning aspect involves safety. A study conducted by Estonian researchers, reported by the Financial Times, revealed that open-weight models are the least effective at filtering Russian disinformation, with Mistral’s top model ranking 47th out of 60 tested.

      For a model about to be integrated into the French civil service under sovereignty claims, the qualities of being ‘harder for an ally to switch off