“Close Model Providers Are Gaining Immense Leverage on Your Business”
Arthur Mensch, co-founder and CEO of French AI company Mistral AI, has addressed enterprise leaders with an extensive plea for open-source models. In a LinkedIn post, the head of Europe’s leading AI startup issues a stark warning against dependence on proprietary (“closed”) AI models – and at the same time outlines how companies should, in his view, build their own independent AI infrastructure.
“Immense leverage” for closed-model providers
Mensch’s central thesis: providers of closed models, who are increasingly moving towards data retention, are gaining “immense leverage” over their customers. As soon as companies connect AI models to their business context, the providers see it and learn from it – and have a “track record of going after their most successful customers thanks to this information,” Mensch argues.
The Mistral CEO doesn’t stop at the model question, however. Companies also need to store their data in open systems, he says, as software vendors might otherwise block them from building AI systems outside the “walled gardens” they have set up. Those who cannot get complete data access from their vendors can now, thanks to AI, “migrate quite fast.”
As a further building block, Mensch points to access management: companies need to govern how AI systems access data on behalf of human users – “because you don’t always want Bob to see what Alice is doing in your company.” That is hard and merciless, he says, because AI models are very good at finding need-to-know errors. What’s needed are systems that check hard access rules, combined with models that check soft access rules.
A proprietary “training flywheel” as competitive advantage
Mensch sees the “most important part” in setting up a company’s own continuous training flywheel: companies should keep improving their AI systems based on interactions with employees and users. This is how the distinctive “edges” of a business become AI systems that neither vendors nor competitors can replicate. A side effect: models can be shrunk according to actual usage patterns, reducing increasingly substantial deployment costs. “We need to collectively become efficient if we want AI development to continue,” Mensch writes.
The Mistral chief concedes that this transformation is enormous: it amounts to both a complete replatforming of a company’s IT and a fundamental change in how software is developed and the business is run. “AI lifecycle management requires understanding human behavior and gradient descent – that’s a stretch.”
A pitch for his own products
The post is, of course, not just analysis but also a product pitch: Mistral provides all the necessary primitives in a single control plane called “Studio” and a training platform called “Forge,” Mensch says. The company deploys on customers’ infrastructure or via hosted zero-data-retention services – “so that your edges remain your edges, and the switch button can be fully in your hand.” His conclusion: “Frontier AI can accelerate the growth of your business, but if it’s not in your hands, it’s not going to be your growth.”
Context: real reference points, but open questions too
Mensch’s arguments have real anchors, but they should not be read without caveats. The reference to forced data retention goes back to a US court order that required OpenAI to preserve ChatGPT logs in the New York Times copyright case – enterprise and zero-data-retention API customers were exempt, however, and the blanket order was later lifted.
Better documented is the concern that model providers end up competing with their own customers: in 2025, Anthropic cut off coding startup Windsurf’s model access while building its rival product Claude Code, and the Brookings think tank has warned that model providers are increasingly competing against their own customer base as they chase application-layer revenue. For the sharpest claim, however – that providers deliberately use customer information to pick their targets – Mensch offers no evidence.
Mensch is not alone with his warning: Palantir CEO Alex Karp also recently cautioned – from a different angle – against vendor lock-in in enterprise AI. For European companies already concerned about the dominance of US tech giants, Mensch’s arguments are likely to fall on fertile ground – even though the Mistral founder is ultimately promoting his own company’s business model, which is built precisely on sovereign deployments on customer infrastructure.
