When a language model accesses a company’s internal data to answer a question, the conversation does not stay confined within the office walls. Arthur Mensch, cofounder and CEO of French AI lab Mistral, made this point bluntly in a LinkedIn post: closed LLM providers are now forcing data retention and gaining ‘immense leverage’ over their customers’ businesses, precisely as companies connect models to their internal context.

Mensch’s vision is straightforward and disruptive: every interaction with a proprietary cloud-based LLM turns the provider into a guardian of your industrial secrets. It is not just about privacy, but about strategic dependency. The more an organization customizes the model with sensitive data, the more the provider sees, learns, and gains an asymmetric advantage.

Mistral has long championed open models and flexible deployment architectures, arguing that data sovereignty starts with the ability to run inference locally. For enterprises considering self-hosting, the logic has immediate implications: an on-premise, self-hosted LLM eliminates forced retention from the start and allows full control over interactions, aligning with regulations like GDPR without having to negotiate opaque clauses with a provider.

Of course, trade-offs exist. Running your own models demands in-house expertise and upfront hardware investment — CPUs, GPUs, storage — which can shift the TCO calculus from operational expense (OpEx) to capital expenditure (CapEx). But for organizations handling sensitive or regulated data, a cloud-only alternative risks a far higher cost: loss of control over intellectual property. Mensch cites no benchmarks or price lists, but his warning points to a principle long explored by AI-RADAR: before integrating an LLM into business workflows, companies should assess whether the convenience of a closed model is worth giving up a competitive edge.

Mensch’s post lands at a time when many enterprises are rethinking their AI strategies, driven by regulators increasingly focused on data residency. The question is no longer just ‘how powerful is the model,’ but ‘where do my prompts run, and who can read them.’ This is fueling growing interest in on-premise and hybrid setups, where the LLM operates within the corporate perimeter and internal data never crosses the firewall.

The clash between open and closed models is not new, but rarely has a CEO framed it as a matter of pure bargaining power. Mensch does not offer a ready-made solution; instead, he holds up a mirror to decision-makers, reminding them that outsourcing artificial intelligence to a closed provider also means outsourcing part of their strategic autonomy.