In an X essay, Satya Nadella argues the enterprise AI moat is the learning loop, not the model
Microsoft Chairman and CEO Satya Nadella has argued that the defining strategic question of the AI era is not which model a company runs but who owns the knowledge produced while running it. He warned that enterprises risk surrendering their most valuable proprietary assets through the ordinary act of using AI well.
In an article published on X on Sunday, Nadella set out what he calls ‘the Reverse Information Paradox’, an inversion of a 60-year-old economic idea that reaches a pointed conclusion: as foundational models converge toward one another and become a commodity available to everyone, the durable competitive advantage shifts away from the model and toward the learning loop a company builds around it.
If every enterprise can rent a comparably capable model, then the model confers no lasting advantage to anyone, and the only asset that compounds is the accumulated judgement a firm generates as it puts that model to work.
Nadella pointed out that today’s commercial arrangements route much of that compounding value back to the model providers rather than the enterprises whose daily work produces it.
Coming from the leader of the company most exposed to the enterprise AI transition through its OpenAI partnership and Copilot line, the article reads less as neutral commentary and more as an argument about how the economics of the industry should be restructured.
A Nobel- Prize winning paradox
Nadella anchored his case in the work of Nobel Prize-winning economist Kenneth Arrow, who described a paradox in which a seller cannot demonstrate the value of information without disclosing it, at which point the buyer has effectively obtained it at no cost. Arrow located the vulnerability with the seller. Nadella argued that AI transfers that vulnerability wholesale to the buyer.
“In the AI age, the buyer risks giving away knowledge, just to use what they bought,” he wrote, adding that “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 harder a company tries to extract value, the worse the mechanism becomes. “The better you want the model to perform, the more of that knowledge you have to feed it.” Here, Nadella effectively described an asymmetry that widens with every interaction, as the seller simply learns more about you as you use what you purchased, and on the flip side, you learn little about what the seller is learning in return.
The practical meaning is that a company's most diligent and sophisticated AI users, the ones extracting the greatest productivity gains, are also the ones exporting the most proprietary knowledge, which turns the usual logic of technology adoption against the adopter.
The asset at risk is not data but the mechanism through which an organisation learns
The essay's sharpest analytical move is its distinction between protecting data and protecting learning. Traditional information security guards static assets; Nadella argued the relevant asset is dynamic, generated continuously through use, and therefore invisible to conventional protection.
Models learn from what he called intelligence exhaust, “the prompts people write, the tools agents use, and especially the corrections people make when the model is wrong.” Every correction, he wrote, “is distilled into institutional know-how,” the kind “a competitor could never buy, and the kind that leaks almost imperceptibly: trace by trace, correction by correction, eval by eval.”
The knowledge in question is not a document that can be classified and locked, but a firm's tacit understanding of its own operations, which Nadella tied to economist Friedrich Hayek's insight about the irreducibly local knowledge of time, place and circumstance that no outside party can hold.
“In consuming intelligence, you are creating intelligence. And what you create should belong to you,” he wrote, describing a knowledge that “knows what you think, what you value, and how you measure success.” The implication for enterprises is that the leak is not an occasional breach to be patched but a continuous condition of use, one that accumulates fastest precisely where AI is embedded most deeply into real workflows.
Nadella indicts an asymmetry in who holds the right to learn
Nadella critiqued the prevailing commercial terms and said it is ironic that model providers rely on fair use rights to train on public data, an innovation he regards as necessary, and then “turn around and impose restrictive terms on distillation, and reserve the right to learn from customer usage and interaction data.”
The objection is not to learning as such but to its one-directional flow. “If learning flows in only one direction,” he wrote, “economic value converges toward the owners of the learning infrastructure rather than the creators of the knowledge itself.”
Nadella stated that better terms negotiated case-by-case would not alter the underlying direction of flow; only distributing the learning infrastructure itself would. To underline what enterprises stand to lose, he borrowed the language of Palantir CEO Alex Karp.
“What the technical customers want is control over their compute, their models, their data stack, and their alpha,” Karp has said. “They want to know they own the means of production, and it's not being transferred to someone else.” The current regime, Nadella argued, “does precisely the transfer Karp and companies fear,” a phrase that recasts a question of software licensing as one of who owns the means of producing intelligence.
A hard trust boundary and five disciplines for keeping the learning loop inside the firm
Nadella's proposed remedy is a trust boundary drawn inside each enterprise, a perimeter within which “an organisation's data, traces, evals, adapted weights, and memory accumulate and improve together,” and across which, in his words, “nothing crosses, not even the intelligence exhaust, without consent.”
He argued that enterprises will come to demand the right to use model outputs to train their own systems, describing this as “every firm's right to align models to their enterprise accountability obligations,” a formulation that grounds the commercial argument in a governance one.
He organised the practical response around five disciplines. Control means building private evaluations, since “evals define what ‘good’ looks like inside the organisation,” and retaining ownership of a firm's memory, traces and institutional context. Capability means constructing proprietary learning environments inside the tenant boundary, where “models learn against real workflows without exposing the company's knowledge.”
Choice means decoupling orchestration from any single model, so that a company can keep operating and optimising even if one model is withdrawn, a test Nadella posed as whether “your company ‘veteran’ capability” survives the loss of any given “‘generalist’ model.”
Cost follows from that decoupling, which lets a firm combine context, models and tasks efficiently without sacrificing quality. The fifth discipline, compound, is the argument's destination: bring the other four together and a company builds “your own continuous learning loop,” a hill-climbing machine that makes its AI investment appreciate rather than depreciate.
The strategic message beneath the checklist is that AI value does not accrue automatically to whoever spends the most on models; it accrues to whoever controls the loop through which use turns into improvement.
Nadella closed on the contrast that carries the point. “In the cloud era, enterprises accumulated data. In the AI era, they accumulate learning,” he wrote, and the boundary that protects a business must evolve “from protecting information to protecting the mechanisms through which organisations learn, adapt, and compound intelligence.” A company, he concluded, “should be able to use a model without giving up the knowledge that makes it unique. That is the reverse information paradox we need to confront.”
