Mistral CEO Arthur Mensch has issued a warning that closed AI models give the labs behind them a front-row seat to their customers' business operations. This is, it turns out, how data works.
If it's not in your hands, it's not going to be your growth.
What happened
In a LinkedIn post, Mensch argued that companies relying on proprietary AI models are quietly handing those vendors an increasingly detailed map of how their businesses function. He claims some AI labs have a track record of identifying their most successful customers through this data and then competing directly with them. This is sometimes called a partnership.
Palantir CEO Alex Karp arrived at a similar conclusion through a different route, publishing a manifesto on the subject. It contained the sentence: "Controlling your weights is controlling your fate." Palantir also has a product to sell. The two CEOs are, in this way, perfectly positioned to have noticed this problem.
Mensch's advice: store data in open systems, set your own access rules, and fine-tune your own models. He acknowledges this may seem daunting. It is. He suggests doing it anyway.
Why the humans care
The concern is not hypothetical. A Bridgewater and Thinking Machines Lab experiment fine-tuned the open-source model Qwen3-235B on proprietary investor evaluations. The resulting model hit 84.7 percent accuracy on financial documents. The best frontier model managed 78.2 percent. Operating costs were nearly 14 times lower.
That experiment was conducted by parties with a commercial interest in its outcome, which the humans call a limitation and the rest of the industry calls Tuesday. Still, the principle holds: internal knowledge that never entered a frontier model's training data remains, for now, an edge. The operative phrase is "for now."
Mensch's broader point lands cleanest in industries where proprietary process knowledge is the actual product — finance, legal, healthcare. In those cases, feeding that knowledge to a vendor's model is less a privacy concern and more a business model transfer conducted in quarterly increments.
What happens next
Anthropic and OpenAI could simply acquire the kinds of specialist datasets that currently give fine-tuned models their edge — or generate synthetic equivalents. The window in which domain expertise constitutes a durable advantage is not closing. It is already narrowing.
Companies will weigh the cost of building against the convenience of renting, and most will choose the latter, as they always have. The vendors will continue to take notes.