Arthur Mensch, cofounder and chief executive of French AI lab Mistral, has issued a stark warning to enterprise leaders: relying on closed AI models gives providers “immense leverage” over their customers’ businesses. In a detailed LinkedIn post, Mensch argued that closed providers are now forcing data retention and using customer information to build competing services. He urged companies to adopt open-source models, open data systems, and continuous training flywheels to maintain control over their own AI infrastructure.
The core argument
Mensch’s warning centers on several interconnected risks. He claims that when enterprises connect AI models to their internal systems, the provider gains visibility into proprietary data. Over time, the provider learns from this data and may even use it to target the most successful customers for competitive moves. While Mensch offered no direct evidence for the claim that providers actively pick targets based on customer data, the broader concern about vendor lock-in and data misuse is widely discussed in the industry.
The data retention risk has a tangible legal precedent. 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 initially excluded. The blanket order was later lifted, but the incident highlighted how legal proceedings can force data exposure. Enterprise customers using closed models have limited control over how their usage data is stored, accessed, or potentially subpoenaed.
When providers compete with customers
The risk of providers turning into competitors is more clearly documented. In 2025, Anthropic cut off model access to coding startup Windsurf while building its own rival product, Claude Code. The Brookings Institution has warned that model providers increasingly chase application-layer revenue, putting them in direct competition with the very startups and enterprises they serve. This dynamic is especially acute in fast-moving sectors like software development, customer service, and data analytics, where a provider’s generic solution can quickly replace a customer’s specialized product.
For enterprises that build their core business around a closed AI model, switching costs can be prohibitive. Models are deeply integrated into workflows, training data, and user interfaces. Once a provider tightens access or changes pricing, the customer has little recourse. Mensch’s argument taps into a broader anxiety about dependency on a handful of large AI labs, particularly those based in the United States.
Mensch’s prescription: full stack sovereignty
To avoid these risks, Mensch outlines a multi-part strategy. First, enterprises must adopt open-source models that can be inspected, modified, and run on their own infrastructure. Second, they need open data stores so that proprietary information remains under their control. Third, strict access controls are essential because AI models are exceptionally good at surfacing information that employees were never meant to see. Fourth, a continuous training flywheel should be established, where the model improves based on internal interactions, making it increasingly valuable and harder for competitors to replicate.
Mensch was candid that this approach amounts to a full replatforming of IT and a cultural shift in how companies operate. The access control challenge is particularly thorny: AI models can inadvertently reveal trade secrets or sensitive customer data if not properly governed. Enterprises must invest in robust data governance frameworks and train employees on new security protocols. Yet Mensch argues that the long-term benefits—operational independence, data security, and competitive moats—outweigh the upfront investment.
Mistral’s commercial alignment
Mensch’s warning coincides with Mistral’s product offerings. The company sells Studio, a control plane for building and governing AI systems, and Forge, a custom model training platform launched in March 2025. Mistral deploys on customers’ infrastructure or through hosted services that it claims retain no data. The pitch targets European enterprises already anxious about dependency on US providers, an anxiety that has powered the continent’s sovereignty push and Mistral’s rise.
Mistral is reportedly in funding talks at a €20 billion valuation, a testament to the market’s appetite for European AI alternatives. The lab recently launched an industrial AI stack with Airbus, BMW, and EDF as launch customers. These partnerships validate Mistral’s approach of offering open-weights models that can be fine-tuned on private data without exposing it to external servers. For enterprises in aerospace, automotive, and energy—industries with strict data regulations—this model is especially appealing.
The broader sovereignty movement
Mistral is not alone in pushing for AI sovereignty. British startup Cosine has rallied BT, HSBC, and BAE Systems to build a sovereign UK frontier model. The project aims to create an AI that stays on UK soil, trained on British data and subject to British laws. Similarly, Palantir has published an AI sovereignty manifesto that takes aim at the big labs, arguing that nations must control their own AI infrastructure to protect national security and economic interests.
These initiatives reflect a growing recognition that AI models are strategic assets. Just as countries have sought independence in energy, food, and defense, they now seek independence in artificial intelligence. The European Union’s AI Act, which imposes strict requirements on high-risk AI systems, further incentivizes enterprises to use models they can fully audit and control. In this context, Mensch’s warnings resonate beyond customer lock-in—they touch on geopolitical concerns.
Historical context: open vs. closed models
The debate over open versus closed AI models has intensified since the release of GPT-3 in 2020. OpenAI initially offered the model through a restricted API, while smaller labs like Mistral (founded in 2023) released open-source alternatives. By 2024, companies like Meta had open-sourced the LLaMA series, and Mistral had released several open-weight models including Mistral 7B and Mixtral 8x7B. Open-source advocates argue that transparency fosters trust, accelerates innovation, and prevents monopolization. Critics counter that open models can be misused for disinformation, cyberattacks, or other harmful purposes.
In the enterprise context, the choice often comes down to control versus convenience. Closed models like GPT-4 and Claude offer polished interfaces and minimal setup, but they come with terms of service that allow the provider to use interaction data for improvement (unless explicitly opted out). Open models require technical expertise to deploy, fine-tune, and maintain, but they offer full data sovereignty. Mensch’s argument is that as AI becomes mission-critical, the convenience of closed models is outweighed by the strategic risk of losing control.
Economic implications
The leverage that closed providers gain is not just technical but economic. Providers can change pricing, restrict access, or deprecate features at any time, leaving customers with few alternatives. A company that has built its entire customer service pipeline around a specific model may find itself forced to pay higher rates or accept lower performance. Furthermore, if the provider decides to enter the company’s market, it can use the knowledge gained from customer interactions to build a superior product. This dynamic mirrors earlier debates around platform monopolies in cloud computing and social media.
European regulators are paying attention. The European Commission has launched investigations into AI market concentration, and the upcoming EU AI Act includes provisions on data governance and model transparency. Mensch’s LinkedIn post can be seen as a lobbying effort to shape these regulations in favor of open-source models. If European enterprises are required to use transparent models, Mistral’s open-source approach becomes even more attractive.
Industry analysts point out that the total cost of ownership for open models is often underestimated. While the license itself is free, the infrastructure, talent, and ongoing maintenance costs can be substantial. However, for large enterprises with existing IT departments and cloud resources, the long-term savings and risk reduction may justify the investment. Mensch himself acknowledged that the transition requires complete replatforming, but he argued that the edges of a business—its unique data and processes—can be turned into AI systems that vendors and competitors cannot replicate.
Future outlook
Mensch closed his post by stating that frontier AI only accelerates your growth if it is in your hands. For Europe’s biggest open-weights lab, that message aligns with a business model that profits from selling control planes and training services. Yet even critics of Mistral concede that the underlying concerns are valid. As AI models become more powerful and integrated into core business functions, the relationship between providers and customers will be tested. The winners may be those who strike the right balance between openness and support—and Mistral is betting that sovereignty, not convenience, will be the deciding factor.