Mistral founder and CEO Arthur Mensch warns companies against relying on closed AI models from large providers. Providers of such models are increasingly storing customer data, thereby gaining deep insights into their customers’ business processes – insights that individual providers, according to Mensch, have already used to target successful customers as competitors.
Mensch’s Core Argument: Loss of Control through Closed Models
In a LinkedIn post, Mensch advises companies to store their data in open systems, establish their own access rules for AI systems, and build their own training – even if, in his own words, this amounts to a complete overhaul of IT and is a mammoth task. The goal is to create AI systems that competitors cannot replicate. “Leading AI can accelerate your company’s growth, but if it is not in your hands, it will not be your growth,” writes Mensch. The question of access rights is particularly sensitive, as AI models are very good at revealing information that employees should not actually have access to.
Notably, Mensch’s accusation: He does not substantiate the central, sharpest part of his claim – that providers have specifically targeted their most successful customers – with concrete examples or evidence, as TheNextWeb notes in its analysis.
Not Alone: Palantir’s “AI-Sovereignty” Manifesto
Mensch is not alone in this position. Palantir CEO Alex Karp had publicly argued similarly shortly before and sharply criticized the business model of large AI providers in a CNBC interview. Companies are paying rising bills for token usage while simultaneously giving away their most valuable data and business know-how, Karp said. The day before, Palantir had published a nine-part “AI-Sovereignty” manifesto that states, among other things: Those who relinquish control over their own model weights allow others to transfer their company’s competitive advantage to their own. Palantir also criticizes the concept of “tokenmaxxing” – spending as much money as possible on AI usage – as a deceptive sense of progress.
A Justified Warning with Self-Interest
To fully contextualize, it should be noted that both Mensch and Karp have a direct economic interest in their respective warnings. Mistral currently has little to counter the performance of current frontier models from large US providers and is instead strategically positioning itself through the argument of European sovereignty – a business model supported by the very concern about loss of control that Mensch warns against. Palantir, on the other hand, sells its own sovereign AI infrastructure solutions through its expanded Nvidia partnership as a direct alternative to the usage-based pricing model of the frontier labs.
There is also a technical counterargument: Large, generalizing AI models often perform better in practice on specialized tasks than smaller, specialized models, provided that the relevant expertise was part of their training data. Mensch’s argument that proprietary training is necessarily superior cannot therefore be universally applied to every use case.
A Counterexample: Bridgewater and Thinking Machines Lab
A current example, independently created from Mensch, partially supports his perspective. As Thinking Machines Lab describes in its own technical report, the hedge fund Bridgewater and the AI startup founded by former OpenAI CTO Mira Murati jointly trained the open model Qwen3-235B with their own evaluation data created by experienced investors. In six tasks for classifying financial documents, the retrained model achieved an average accuracy of 84.7 percent, while the best-tested frontier model – based on its own, carefully constructed prompts – was at 78.2 percent. For simple, less optimized queries, the large models reportedly achieved only about 50 percent accuracy. The operating costs of the specialized model were about 14 times lower than those of the leading frontier models.
However, this is not an independent comparison: Both Bridgewater and Thinking Machines Lab have an economic interest in the success of this approach, and the results were not verified by an outside institution. It is also unclear whether this advantage will hold in the long term – providers like Anthropic or OpenAI could potentially purchase or generate comparable expert data in the future and regain their lead.
Contextualization
Mensch’s warning hits a sore point, which other market participants like Palantir are now also publicly articulating: The question of who ultimately retains control over proprietary data, model weights, and resulting competitive knowledge becomes increasingly important with growing AI penetration. At the same time, both warnings are clearly also strategic positions in their own business interest, and the specific accusation of targeted attacks on successful customers remains unsubstantiated so far. The Bridgewater example shows that proprietary expertise can indeed provide a measurable advantage – whether this advantage is lasting once the large providers catch up remains to be seen.
