The Growing Concern Over Closed AI Models
Arthur Mensch, cofounder and CEO of the French AI lab Mistral, has issued a stark warning to enterprise leaders: relying on closed AI models gives providers what he calls 'immense leverage' over their customers' businesses. In a recent LinkedIn post, Mensch argued that as companies integrate these models into their internal systems, providers gain visibility into sensitive data, learn from it, and may eventually compete with their own customers. This argument touches on fundamental tensions in the AI industry—between openness and control, innovation and privacy, and vendor dependency and self-sufficiency.
The debate over open versus closed AI models is not new, but Mensch’s intervention comes at a critical time. With enterprise adoption of generative AI accelerating rapidly, many organizations are connecting proprietary business data to large language models (LLMs) to enhance productivity, automate processes, and create new services. However, Mensch warns that this convenience comes at a hidden cost, potentially ceding strategic control to a handful of powerful AI providers.
Data Retention: A Real but Caveated Risk
One of Mensch’s most pointed claims is that closed providers force data retention on their customers. He references a US court order from the copyright lawsuit between The New York Times and OpenAI, which initially required OpenAI to preserve ChatGPT logs. Although that blanket order was later lifted, it illustrated the legal risks that can arise when enterprise data passes through third-party AI systems. Crucially, the order exempted enterprise API customers with zero-data-retention policies, highlighting that not all usage is equally vulnerable. Yet the incident underscores a broader reality: once data enters a closed provider’s infrastructure, control over its retention and use partially shifts away from the enterprise.
This concern is not hypothetical. Cloud computing giants like Amazon Web Services, Microsoft Azure, and Google Cloud have long offered data retention guarantees and compliance certifications, but AI models add a new dimension. Because LLMs can memorize and reproduce training data, enterprises must trust that their confidential information—trade secrets, customer lists, financial projections—is not inadvertently exposed or used to improve competitive services. Mensch’s warning amplifies this trust issue, urging leaders to reconsider the data flows they currently accept.
When Providers Become Competitors
Another pillar of Mensch’s argument is the risk that AI providers will directly compete with their most successful customers. He points to the case of Anthropic, which cut off API access to the coding startup Windsurf in 2025 while simultaneously developing its own rival product, Claude Code. A Brookings Institution study has similarly documented how model providers increasingly chase application-layer revenue, placing them in direct competition with the very companies they serve. This dynamic creates a fundamental conflict of interest: the provider has access to usage patterns, performance metrics, and even business strategies of its customers, information that could be weaponized to build superior competing offerings.
Mensch does not offer direct evidence that any provider has systematically used customer data to target acquisitions or poach markets, but the pattern is troubling enough to warrant caution. For enterprises building core business processes on top of closed AI platforms, the risk of eventual displacement is real. The history of platform dependency in other technology sectors—from operating systems to e-commerce marketplaces—shows that dominant players often expand into adjacent spaces, leveraging their control over the infrastructure.
The Open-Source and Sovereignty Countermovement
Against this backdrop, a growing number of organizations are advocating for open-source AI models and sovereign AI infrastructure. British startup Cosine has rallied BT, HSBC, and BAE Systems to build a sovereign UK frontier model, while Palantir has published an 'AI sovereignty manifesto' taking direct aim at the large labs. These initiatives reflect a desire to reduce reliance on US-based AI providers, especially in Europe, where data protection regulations like GDPR impose strict constraints on cross-border data transfers. Mistral itself has positioned itself as a European champion of open-weights models, deploying on customer infrastructure or through services that promise zero data retention.
The open-source approach offers several advantages: transparency in training data and algorithms, the ability to fine-tune models on proprietary data without exposing it, and independence from provider pricing and policy changes. However, it also requires significant investment in infrastructure, talent, and ongoing maintenance. Mensch acknowledges this trade-off, describing a complete replatforming of IT and a cultural shift in how companies operate. He specifically highlights access control as a minefield, since AI models excel at surfacing information that employees were never meant to see—an unintended consequence of open internal data stores.
Mistral’s Own Offering: From Warning to Product Pitch
Mensch’s argument aligns seamlessly with Mistral’s commercial interests. The company sells Studio, a control plane for building and governing AI systems, and Forge, a custom model training platform launched in March 2025. Both products are designed to give enterprises the tools to build and manage their own AI systems in house, either on their own infrastructure or through hosted services that Mistral says retain no data. This pitch targets the same anxiety about dependency that has powered Europe’s sovereignty push and Mistral’s own rapid rise. The lab is reportedly in funding talks at a valuation of €20 billion and recently launched an industrial AI stack with Airbus, BMW, and EDF as launch customers.
Of course, Mensch’s self-serving rhetoric does not invalidate the core concerns. Even critics of Mistral acknowledge that the risks of closed models are real, albeit sometimes exaggerated. The key question for enterprise leaders is whether the benefits of ease of use and rapid integration outweigh the potential long-term vulnerabilities. For those in highly regulated industries—finance, healthcare, defense—the choice may already be clear: control and sovereignty are nonnegotiable. Others may wait to see how the market evolves, weighing the convenience of closed APIs against the strategic independence of open alternatives.
The Bottom Line: Hands-On AI as a Strategic Imperative
Mensch concludes that frontier AI only accelerates growth if it is in the hands of the user. For Europe’s biggest open-weights lab, that statement is both a warning and a business opportunity. As enterprises increasingly treat AI as a core operational capability rather than a plug-in utility, the debate over open vs. closed models will only intensify. The winners will be those that navigate this transition with clear eyes, understanding the full implications of where their data goes and who controls the learning.
The push for AI sovereignty is not limited to Europe. In the United States, government agencies are exploring open-source models to avoid foreign influence, while in Asia, countries like Japan and India are investing in domestic AI infrastructure. The trend toward greater control over AI systems reflects a broader recognition that data and model access are strategic assets. Mensch’s LinkedIn post, however self-interested, captures a pivotal moment in the evolution of enterprise AI—one where the choice between open and closed may define competitive advantage for years to come.