Satya Nadella has thrown a grenade into the AI industry. In a blog post Monday, the Microsoft CEO warned that companies using proprietary AI models from labs like OpenAI and Anthropic are unknowingly handing over their most sensitive business data. He argues these firms pay twice: once for the tokens, and again by revealing proprietary knowledge that makes the models useful. "The better you want the model to perform, the more of that knowledge you have to feed it," he wrote.

Nadella’s concern centers on what he calls "exhaust"—the prompts, agent tools, and user corrections that teach a model the nuances of a business. Every correction becomes institutional know-how that a competitor could never buy, yet enterprises are feeding it directly to model makers. He calls it hypocritical for AI labs to freely scrape the internet for training data while blocking customers from studying—or "distilling"—those models in return. Distillation, the practice of using a model’s outputs to train cheaper alternatives, has already sparked tensions: in February, Anthropic accused Chinese open-source models of exploiting Claude for this purpose.

Nadella’s solution is conveniently aligned with his own business. He urges companies to retain ownership of their data, build proprietary learning environments on the cloud (like Microsoft’s Azure), and create "orchestration layers" to switch between models without lock-in. The subtext is clear: move to open-source models on your own premises. That trend is already underway. Solo.io CEO Idit Levine says her customers are ditching proprietary models for open-source alternatives that deliver 90% of the performance at a fraction of the cost, with full control. Vercel reports that open models now account for 29% of traffic through its AI gateway. With Microsoft’s CEO now openly cautioning against the very labs his company invests in, expect this shift to accelerate.

Source: TechCrunch