Arthur Mensch, cofounder and chief executive of French AI lab Mistral, has issued a stark warning to enterprise leaders: abandon closed AI models or risk giving providers “immense leverage” over your business. In a recent LinkedIn post, Mensch argued that closed providers are now forcing data retention and leveraging customer data to gain competitive advantages. The sharpest part of his charge—that providers use customer information to pick acquisition targets—remains an inference without direct evidence, but the broader concerns are grounded in real-world events.
Data retention: a real anchor with caveats
Mensch’s claim about data retention has a factual basis, though with important qualifications. A U.S. court ordered OpenAI to preserve ChatGPT logs during The New York Times copyright case, though enterprise and zero-data-retention API customers were excluded. The blanket order was later lifted, but the incident highlighted how legal actions can force providers to retain data even when they promise not to. This creates a chilling effect for enterprises that rely on closed models, as they cannot fully control how their data is treated in the event of litigation.
The broader implication is that even with contractual promises, external forces can compel providers to retain data. For enterprises handling sensitive information, this uncertainty is a significant risk. Mensch argues that only open-source models, hosted on the customer’s own infrastructure or with strict zero-retention policies, can guarantee control.
Customer competition: a documented pattern
Mensch’s second major point—that providers inevitably compete with their most successful customers—is better documented. In 2025, Anthropic cut off coding startup Windsurf’s model access while building its rival Claude Code. This event underscores the conflict of interest inherent in large AI labs that both sell access to their models and develop their own applications. The Brookings Institution has warned that model providers increasingly compete with their own customers as they chase application-layer revenue. For enterprises building products on top of closed APIs, this creates a fundamental vulnerability: the provider can become a competitor overnight.
The pattern extends beyond Anthropic. OpenAI has invested in application-layer startups and developed products like ChatGPT Enterprise that directly compete with third-party tools. Google’s Vertex AI platform similarly offers services that overlap with those of its customers. Mensch’s argument is that enterprises must avoid building on platforms where the provider has both access to their data and incentive to copy their innovations.
The prescription: replatforming IT
Mensch’s proposed solution is comprehensive and radical. He calls for enterprises to adopt open-source models, open data stores, strict access controls, and a continuous training flywheel that improves systems on internal interactions. The goal is to turn the edges of a business into AI systems that vendors and competitors cannot replicate. However, he is candid that this amounts to a complete replatforming of IT and a fundamental change in how companies operate.
Access control is particularly challenging, according to Mensch, because AI models excel at surfacing information that employees were never meant to see. This requires robust governance systems that prevent data leakage while still allowing the model to learn from interactions. Mistral’s Studio platform is designed to address this, offering a control plane for building and governing AI systems.
Training custom models is no longer a fringe position. British startup Cosine has rallied BT, HSBC, and BAE Systems to build a sovereign UK frontier model. Palantir has published an AI sovereignty manifesto taking aim at the big labs. These initiatives show a growing recognition that control over AI infrastructure is a strategic imperative, not just a technical choice.
The European sovereignty push
Mensch’s argument resonates particularly strongly in Europe, where anxiety about dependency on U.S. technology providers has fueled a sovereignty push. The European Union’s AI Act, while focused on regulation, also encourages development of homegrown AI capabilities. Mistral itself has benefited from this trend, positioning itself as the European alternative to OpenAI and Google. The Paris-based lab is reportedly in funding talks at a €20 billion valuation and recently launched an industrial AI stack with Airbus, BMW, and EDF as launch customers.
The company’s products—Studio for governance and Forge for custom training—are designed to meet the needs of enterprises that want to retain control. Forge, launched in March, allows customers to train models on their own data using Mistral’s expertise, with the resulting model owned entirely by the customer. This model avoids the data retention and competition risks of closed APIs.
Historical context: open vs. closed
The debate between open and closed AI models mirrors earlier battles in the software industry. In the 1990s, enterprises faced similar choices between proprietary software from vendors like Microsoft and open-source alternatives like Linux. Many companies initially chose proprietary solutions for their perceived ease of use and support, only to later face vendor lock-in and high switching costs. The rise of open-source software in the 2000s demonstrated that community-driven development can produce robust, secure, and cost-effective alternatives.
In AI, the stakes are even higher because the technology is integrated into core business processes and decision-making. Closed models create a single point of failure: if the provider changes its pricing, terms of service, or business strategy, the customer has little recourse. Open-source models, by contrast, can be forked, modified, and self-hosted, giving enterprises ultimate control. However, they require in-house expertise for deployment, maintenance, and fine-tuning—a barrier that Mistral’s services aim to lower.
Technical considerations for enterprises
Implementing Mensch’s vision involves several technical challenges. First, enterprises need robust data pipelines to feed the training flywheel. Data must be cleaned, labeled, and stored in formats that models can learn from. Second, access controls must be designed to allow the model to surface relevant information without exposing sensitive data. This requires integration with existing identity and access management systems, as well as monitoring for data leakage.
Third, the continuous training loop means that models must be updated regularly as new internal data emerges. This requires infrastructure for versioning, testing, and deploying models without disrupting operations. Mistral’s Forge platform handles many of these tasks automatically, but enterprises still need to invest in data infrastructure and AI governance.
Finally, enterprises must assess the total cost of ownership. Training custom models can be expensive, requiring significant compute resources and engineering time. However, for companies that rely heavily on AI for competitive advantage, this cost may be justified by the reduction in provider leverage and the ability to build unique capabilities.
The Windsurf case in detail
The Windsurf-Anthropic incident serves as a cautionary tale. Windsurf, a startup building AI-powered coding assistants, had relied on Anthropic’s Claude API. In 2025, Anthropic announced its own Claude Code product, which directly competed with Windsurf’s offering. Shortly after, Windsurf’s API access was cut off, allegedly for violating terms of service related to competitive use. Windsurf disputed this, arguing that the cutoff was anticompetitive. The case remains in litigation, but it has become a rallying point for advocates of open-source AI.
This incident highlights the asymmetric power relationship between model providers and their customers. Even if a provider promises not to compete, they have full visibility into how customers use their models, which trends are emerging, and which customer segments are most valuable. This information can be used to prioritize internal product development, giving the provider an unfair advantage.
Mensch argues that the only way to avoid this dynamic is to own the model and the entire stack. When a provider is also a potential competitor, any shared infrastructure is a vulnerability. Mistral’s business model, by contrast, is based on selling tools and services for self-hosted models, not on selling access to a proprietary platform. This aligns the company’s incentives with its customers’ interests.
Broader market implications
Mensch’s warning comes at a time when the AI market is maturing. Enterprises that rushed to adopt closed models are now facing integration challenges, cost overruns, and strategic dependencies. The initial appeal of closed models—ease of use and rapid deployment—is giving way to concerns about long-term sustainability and control. Analogously, many enterprises that adopted public cloud infrastructure have since moved to hybrid or multi-cloud strategies to avoid lock-in. AI appears to be following a similar trajectory.
The rise of open-weight models like Meta’s Llama and Mistral’s own releases has given enterprises credible alternatives to proprietary APIs. These models can be fine-tuned with domain-specific data, often achieving performance comparable to frontier closed models for specialized tasks. As the ecosystem of open-source tools and services grows, the cost and complexity of self-hosting decrease, making the sovereignty argument more compelling for a wider range of organizations.
Mensch closed his LinkedIn post by warning that frontier AI only accelerates your growth if it is in your hands. For Europe’s biggest open-weights lab, hands and business model happen to point the same way. Whether enterprises heed the warning will depend on their risk tolerance, technical readiness, and strategic priorities. But the debate Mensch has sparked is likely to shape the enterprise AI landscape for years to come.