Microsoft chairman and CEO Satya Nadella has warned that enterprises adopting AI face a hidden cost beyond subscription fees, the proprietary knowledge they are forced to hand over just to make the technology useful. In a lengthy post on X, Nadella argued that using AI effectively means 'you essentially pay for intelligence twice,' first with money, and then by revealing institutional know-how that competitors could never otherwise buy.
Nadella calls it 'The Reverse Information Paradox'
Nadella framed his argument as an inversion of a decades-old economic theory. He cited Nobel Prize winning economist Kenneth Arrow's "Information Paradox," under which a seller of information risks losing its value the moment a buyer sees it. According to Nadella, AI flips that risk onto the buyer instead. He called this dynamic the "Reverse Information Paradox," writing that the better a company wants an AI model to perform, the more proprietary information it must feed it. Over time, he said, the resulting information asymmetry becomes increasingly skewed in the seller's favour, since the AI provider learns continuously from usage while the enterprise learns comparatively little in return.
Every correction becomes 'intelligence exhaust'
Central to Nadella's argument is what he called "exhaust", the accumulated prompts employees write, the tools AI agents use, and particularly the corrections made when a model gets something wrong. He said these corrections are quietly distilled into institutional know-how that leaks out gradually, in his words, "trace by trace, correction by correction, eval by eval." Nadella argued that this exhaust represents a form of "particular intelligence," invoking economist Friedrich Hayek's concept of localized knowledge of time, place and circumstance that no outside party can replicate.
Patents solved one paradox. This one needs its own fix
Nadella noted that patents historically solved Arrow's original paradox by letting inventors disclose ideas without giving them away for free. He argued the Reverse Information Paradox needs an equivalent safeguard, one that goes beyond conventional data protection to cover the mechanisms through which organisations learn and adapt using AI. He also referenced Palantir CEO Alex Karp, quoting him as saying enterprise customers want to "own the means of production" rather than see it transferred elsewhere, and said the current AI ecosystem does precisely the transfer companies fear.
Nadella's five-point framework for enterprises
To address the risk, Nadella outlined a five-part framework built around what he called the five Cs. Control involves enterprises building their own evaluation systems and retaining ownership of institutional memory, feedback and decisions generated through AI use. Capability calls for proprietary learning environments within a company's own tenant boundary, so models can be trained or tuned against real workflows without exposing sensitive knowledge externally. Choice requires decoupling the orchestration layer from any single AI model, so a company retains the ability to operate even if one model provider is removed from the stack. Cost follows from that same decoupling, allowing firms to combine context, models and tasks efficiently without compromising quality. The fifth, Compound, ties the previous four together into what Nadella described as a continuous learning loop that lets AI investment compound value inside the firm rather than outside it.
A shift from protecting data to protecting learning
Nadella argued that enterprises need a new kind of trust boundary, one that keeps not just data but "intelligence exhaust" inside the organisation without explicit consent to share it. He said that while it was reasonable for AI companies to have fair use rights to train on public data, it was inconsistent for those same companies to then impose restrictive terms on distillation while reserving the right to learn from customer usage. In Nadella's words, if learning flows in only one direction, economic value will converge toward the owners of AI infrastructure rather than the enterprises actually generating the knowledge.
Nadella summed up his argument by stating that a company should be able to use an AI model "without giving up the knowledge that makes it unique."
