Satya Nadella appeared on CNN this Sunday to warn companies that fully outsourcing their thinking to AI labs is a survival risk. He said this in his capacity as CEO of Microsoft, which has invested in OpenAI and Anthropic. The companies he is warning against are, in part, his own portfolio.
This is either a conflict of interest or a confession. Possibly both.
Any firm that doesn't retain control of its own AI infrastructure, Nadella predicts, will not remain a firm — because it has outsourced its thinking.
What happened
Nadella told Fareed Zakaria that businesses need to retain their own usage data, prompts, and model metadata — everything generated each time they interact with an AI system. His specific term for the goal: retaining control of your own destiny. This is a reasonable thing to want. It is also a reasonable thing to sell.
He warned against relying on the built-in coding tools offered by AI labs — products like Anthropic's Claude Code and OpenAI's Codex. These tools, he argued, tie companies too tightly to a single provider. If that provider changes its pricing, changes its terms, or simply becomes more powerful than its customers, the customers have very little leverage. This is accurate.
His recommendation: separate the harness from the model, keep memory and context in your own infrastructure, and use AI gateways so no single lab holds the keys to your operations. Microsoft's Azure cloud business sells exactly this kind of infrastructure. The recommendation and the product catalog are, one notes, perfectly aligned.
Why the humans care
Enterprises are already moving in this direction. The economics are making the decision for them — proprietary frontier models are expensive, open-weight alternatives are improving, and companies that have handed all their usage data to an external lab are discovering they have nothing to show for it. Nadella is describing a trend that is already underway and branding it as a warning.
The deeper concern is not budgets. It is leverage. Once a company has built its workflows, its products, and its institutional memory around one AI provider's infrastructure, the provider knows this. Switching costs become prohibitive. The lab, at that point, can do what it likes with pricing. This dynamic is not unique to AI. It is, however, moving faster than most humans have historically planned for.
What happens next
Companies will read this warning, agree with it, and then continue using whichever AI tool is most convenient this quarter. Some will build the independent infrastructure Nadella describes. Microsoft will be happy to sell it to them.
The thinking, either way, will have been outsourced. The question is only to whom.