Every enterprise board deck now seems to have an AI mandate. Companies are feeding customer records, internal strategy documents, and proprietary analytics into large language models at a pace that would have seemed reckless just three years ago. Most of this traffic flows into closed-source systems. They are convenient, powerful, and hosted by someone else. Arthur Mensch, founder of Mistral AI, thinks this convenience carries a hidden tax that could outstrip any subscription fee.
Mensch has been vocal about a specific danger: when you run your business through a proprietary model, you are not just renting compute. You are handing the lab that built it a detailed map of your operations. These providers store enormous volumes of customer interaction data to refine their systems. Over time, that data reveals patterns. They see how you price products, how you segment audiences, how you identify risk, and how you spot opportunity. Mensch warns that some labs have already shown a willingness to use this intelligence to target their own customers, entering markets once they understand exactly where the profits sit.
His advice is blunt. "If it's not in your hands, it's not going to be your growth." The remedy he offers is sovereignty: moving data into open systems, enforcing strict access rules, and investing in proprietary training pipelines. This is not paranoid fantasy. It is supply-chain security applied to cognition. You would not give your supplier your factory blueprints. Mensch argues that you should not give your AI provider the equivalent of your institutional playbook.
Why Model Weights Are Your Corporate Memory
The idea finds support from an unlikely corner of the tech industry. Alex Karp, CEO of Palantir, frames the issue with a phrase that is already making the rounds in defense and finance circles: "controlling your weights is controlling your fate."
To understand why this matters, it helps to think about what model weights actually represent. In machine learning, weights are the billions of tuned parameters that turn raw data into useful predictions. When you train or fine-tune a model on your own data—your customer churn history, your diagnostic images, your logistics optimization cases—that knowledge gets compressed into the weights. They become a kind of corporate memory, encoding decisions and expertise that took years to accumulate.
If you rent a frontier model and simply prompt it with your secrets, you never own that distilled knowledge. Worse, your provider absorbs every signal. Each query teaches their system a little more about your vertical. Their next model update could internalize the very edge that made you successful. Karp calls this your "alpha," the unique performance advantage that separates you from the pack. Outsourcing your weights means outsourcing your alpha. At that point, AI stops being a tool and becomes a conduit for transferring your expertise to a competitor.
The Fine-Tuning Advantage
The Bridgewater experiment shows there is another way. The hedge fund worked with Thinking Machines Lab to fine-tune the open-source Qwen3-235B model. Instead of sending proprietary investor evaluations into a black-box API, they kept the data in-house and adapted the model to their specific needs. The results were stark. On financial document analysis, their fine-tuned model hit 84.7 percent accuracy. The best frontier closed model managed only 78.2 percent.
That gap matters in finance, where a