July 13, 2026
Paying Twice
Satya Nadella warns companies pay twice for AI, in money and in data. Days earlier, his own company swapped AI partners under the same button.
On Sunday, Satya Nadella published a piece warning AI users that they are paying twice. The first charge shows up on the invoice: tokens, subscription, usage. The second is what the team hands over: every prompt, every correction made when the model gets it wrong, becomes, in his words, "knowledge no competitor could buy," and that knowledge stays with the vendor.
The irony is that the one issuing the warning is the CEO of the company that invested billions in both OpenAI and Anthropic. And which, just days earlier, started routing part of Copilot's prompts in Excel and Outlook to Microsoft's own models, swapping out partners under the same button, without telling the people using it.
Nadella proposes two ways out: keep the data in your own learning environments, and build orchestration layers that let you swap models without getting locked to one vendor, which sounds great, since nobody wants to be held hostage by a single supplier. But it also sounds like someone hoping that layer runs on his own cloud.
Set aside the commercial interest of whoever is speaking, and the argument still holds. Every correction a team makes to a model is accumulated decision-making turning into another company's property. It is a question of architecture: who ends up owning what the team learned about its own business.
A company that outsources its learning outsources its competitive advantage. And it will not see that on the invoice.
Sources: TechCrunch · The New Stack