“You Pay for Intelligence Twice”: Microsoft CEO Warns of Closed Models
In a lengthy blog post, Microsoft CEO Satya Nadella has introduced a new term into the AI debate: the “Reverse Information Paradox.” His thesis: companies that use AI models are gradually giving away their most valuable asset — their proprietary knowledge. It’s worth noting that Microsoft itself is a major investor in both OpenAI and Anthropic.
Nadella builds on a classic concept from information economics. Nobel laureate Kenneth Arrow once described the paradox that the value of information only becomes apparent to the buyer once they know it — but at that point, they have effectively acquired it for free. The risk sat with the seller.
Artificial intelligence reverses this relationship, Nadella argues. In the AI age, it is the buyer who risks giving away knowledge — simply by using what they bought. “You essentially pay for intelligence twice,” the Microsoft chief writes: once with money, and a second time with the proprietary knowledge you have to feed the model to make it useful. The better you want the model to perform, the more of that knowledge you have to reveal.
“You essentially pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful. The better you want the model to perform, the more of that knowledge you have to feed it!” — Satya Nadella
“Intelligence Exhaust” as the Gateway
At the heart of Nadella’s argument is what he calls “intelligence exhaust”: the prompts employees write, the tools AI agents use, and above all the corrections people make when the model gets things wrong. Each of these corrections, he argues, distills institutional know-how — knowledge no competitor could ever buy, but which leaks away almost imperceptibly, “trace by trace.”
Over time, the information asymmetry keeps shifting: the provider continuously learns about the customer, while the customer learns very little about what the provider does with that knowledge.
Nadella also takes aim at the industry’s business practices — remarkable coming from the head of a company that, with Azure and its OpenAI partnership, ranks among the largest AI providers itself. He calls it ironic that model providers claim fair-use rights to train on public data on the one hand, while on the other imposing restrictive terms on distillation and reserving the right to learn from customer data. If learning flows in only one direction, he argues, economic value concentrates with the owners of the learning infrastructure rather than with the creators of the knowledge.
Five Principles for Enterprises
In response, Nadella calls for a hard “trust boundary”: a line across which nothing — not even the intelligence exhaust — passes without the company’s consent. Within this boundary, data, traces, evals, adapted model weights, and organizational memory should accumulate together. Specifically, he lays out five principles:
Control: Build your own private evals and retain ownership of traces, feedback, and institutional context — including the right to reuse model outputs from your own tasks.
Capability: Create proprietary learning environments within the tenant boundary, where models are trained or tuned on real workflows without exposing company knowledge.
Choice: Decouple the orchestration layer from any single model. Nadella’s test question: does the company’s capability survive if a given model is taken away?
Cost: Use that decoupling to combine context, models, and tasks as efficiently as possible.
Compound: Bring all four together into a continuous learning loop that compounds the value of AI investments over time.
Nadella also cites Palantir CEO Alex Karp, according to whom technical customers want control over their compute, models, data stack, and their “alpha” — the assurance that their own means of production are not being transferred to third parties.
Analysis
The post strikes a nerve in the current debate around AI sovereignty, one that is being waged intensely in Europe as well. What stands out is the tension with Nadella’s own position: as a cloud and model provider, Microsoft profits from the very status quo he criticizes — while simultaneously positioning itself, through its call for customer sovereignty, as an alternative to pure-play model providers. Whether the principles will be matched by corresponding contractual terms at Microsoft itself — for instance, distillation rights for Copilot and Azure customers — is a question the post leaves open.
