Microsoft CEO Satya Nadella just published a blog post that reads like a warning shot aimed directly at OpenAI, Anthropic, and the rest of the proprietary model crowd. His core argument? Enterprises are getting fleeced. They pay for AI tokens with cash, then pay again by feeding their most sensitive business knowledge into models that happily learn from it. ‘You essentially pay for intelligence twice,’ Nadella writes, ‘once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful.’
He’s particularly fixated on what he calls ‘exhaust’ — the prompts, corrections, and tool usage that distill a company’s institutional know-how into the model. Every time an employee corrects a bad output, they’re teaching the AI how their business actually works. Nadella calls this ‘the kind of knowledge a competitor could never buy,’ and he’s not wrong. The implication is clear: today’s helpful AI assistant could be tomorrow’s well-informed competitor.
Nadella also takes aim at the hypocrisy around ‘distillation,’ the practice of using a model’s outputs to train cheaper alternatives. He points out that AI labs freely scrape the entire internet for training data but then ‘turn around and impose restrictive terms on distillation’ when someone tries to learn from their models. It’s a pointed reference to Anthropic’s recent demands for U.S. export controls after Chinese developers allegedly pumped millions of prompts through Claude. Nadella’s stance: if you can learn from public data, others should be able to learn from you.
The proposed fix is self-interested but pragmatic. Nadella wants companies to build ‘proprietary learning environments’ on cloud infrastructure — preferably Azure, naturally — with orchestration layers that let them swap between models freely. But the real undercurrent here isn’t about Microsoft’s cloud. It’s about the accelerating shift toward open source models running on-premise. Solo.io CEO Idit Levine says her enterprise customers are increasingly asking: ‘Can I take an open source model and run it on-prem? It will do almost 90% of what the big one’s doing. It will cost way less.’ Vercel’s gateway traffic tells the same story: open models accounted for 29% of all routed requests last month. When the CEO of OpenAI’s biggest backer starts arguing that the proprietary model business is structurally dangerous for customers, the market is already moving.
