Microsoft CEO Satya Nadella has warned enterprises that seriously adopting AI models means handing over the very knowledge that makes their business valuable. In a lengthy post on X on Sunday night that drew more than 5.7 million views, Nadella described the problem as a “Reverse Information Paradox” and called for the industry to find a fix.
Nadella built his argument on economist Kenneth Arrow’s classic information paradox, which holds that a seller of information cannot prove its worth without revealing it, at which point the buyer effectively has it for free. Nadella argued that AI inverts this dynamic: the buyer risks giving away knowledge simply to use what they purchased. “You pay twice, once in cash, once in proprietary knowledge,” he wrote, adding that the better a company wants a model to perform, the more information it must feed it.
He pointed to what he called “exhaust” as the sharper issue. This is not just a company’s raw data but its prompts, the tools its AI agents call, and above all the corrections employees make when a model gets something wrong. Every one of those corrections distills institutional know-how, and it leaks out gradually, in his words, “trace by trace, correction by correction, eval by eval.” Over time, Nadella said, the vendor learns a great deal about the customer while the customer learns almost nothing in return.
Nadella also flagged what he described as an irony in how AI companies operate: model providers rely on fair use to train on public data, then impose restrictive terms on distillation while reserving the right to learn from customer usage and interaction data. If learning flows in only one direction, he argued, economic value pools with whoever owns the learning infrastructure rather than with the people who created the knowledge. He cited Palantir CEO Alex Karp’s description of what technical customers want: control over their compute, models, data stack and alpha, and assurance that ownership of the means of production is not being transferred to someone else.
To address this, Nadella laid out a five-part framework for enterprises. Control means building private evals and retaining ownership of an organisation’s memory, traces, feedback and decisions. Capability means building proprietary learning environments inside a company’s own tenant boundary so models can be trained or tuned without exposing that knowledge externally. Choice means decoupling the orchestration layer from any single model, so losing access to one vendor does not mean losing the underlying capability. Cost follows from that same decoupling, letting a company match tasks to models efficiently without sacrificing quality.
The fifth element, Compound, ties the other four together into what Nadella called a continuous learning loop, allowing a firm’s human capital and token capital to accumulate over time. He said the goal is for enterprises to use AI models without surrendering what makes them unique, arguing that the trust boundary companies need must evolve from simply protecting information to protecting the mechanisms through which they learn, adapt and compound intelligence.