Microsoft owns 27% of OpenAI, but it’s not shying away from pointing out how reliance on AI model companies could have detrimental effects for large enterprises.
In a new essay, Satya Nadella has laid out what he calls the Reverse Information Paradox, and argued that enterprises need a mechanism resembling patents to protect what they reveal to AI models while using them.
Nadella built his argument off economist Kenneth Arrow’s old observation about the market for information — that a buyer can’t know the value of information until they’ve received it, at which point they’ve effectively acquired it for free. That was Arrow’s paradox, and it explained why sellers of information have always struggled to price and protect what they know. Patents were society’s answer. They let inventors disclose an idea publicly without losing their claim to it.
Nadella’s argument is that AI has produced the mirror image of that problem. In Arrow’s world, the seller was at risk. In the AI economy, Nadella says, it’s the buyer, the enterprise, that ends up exposed. The better a company wants a model to perform, the more of its own proprietary context, workflows, and corrections it has to feed the model, and every one of those corrections becomes something the model provider could plausibly learn from. Over time, he argues, the model provider ends up knowing more about the enterprise than the enterprise ever learns about how the model itself is evolving.
This isn’t a new concern for Nadella. He has talked before about the fact that he believes foundational models are getting commoditized, and separately warned that model companies could be sitting on a “winner’s curse,” where being first to build a great model counts for less than being able to defend that lead once it’s replicable. The patent framing extends that thinking to the enterprise side of the table. If models themselves are becoming interchangeable, then the only durable asset an enterprise can build is the compounding institutional knowledge layered on top of a model, and that asset is precisely what’s at risk of leaking back to the vendor through everyday usage.
The essay lands close to ideas Nadella floated in an earlier post about “token capital,” where he described human capital and AI capability compounding together inside a company’s own learning loop. This new framing sharpens that argument into something closer to a policy demand. Nadella lists what he thinks enterprises need to insist on: ownership of their own evals, since evals define what “good” looks like inside an organization; private environments to train or fine-tune models on internal workflows without exposing that data outside the company’s boundary; an orchestration layer that isn’t locked to a single model, so a company retains its own expertise even if a given model gets swapped out or discontinued; and the resulting cost efficiency of being able to route tasks across models rather than being captive to one vendor’s pricing.
Underneath the framework is a pointed critique of the terms model providers currently set. Nadella notes the irony of AI companies benefiting from broad fair use rights to train on public data, while simultaneously writing contracts that restrict enterprises from distilling their models or reserve the right to learn from customer prompts and corrections. He quotes Palantir’s Alex Karp, who has argued that enterprise customers want control over their compute, models, data stack, and “alpha,” and don’t want that ownership quietly transferred to someone else. Nadella’s point is that the current arrangement between most model vendors and their enterprise customers does exactly the transfer Karp warns about.
There’s an obvious reading of why Microsoft, of all companies, would make this case publicly. It has a 27% stake in OpenAI’s for-profit arm and, as Nadella has confirmed elsewhere, access to nearly all of OpenAI’s IP short of consumer hardware. A framework where enterprises demand ironclad boundaries around their own data plays to Azure’s positioning as the neutral infrastructure layer beneath a range of models, rather than a single frontier lab trying to own the entire stack. If enterprises start treating model portability and data sovereignty as procurement requirements, that favors whoever sells the orchestration layer, not whoever sells the model.
Whatever the motivation, the underlying warning is one enterprise buyers are already beginning to internalize on their own: every correction an employee makes to a model, every prompt refined through trial and error, and every eval built to measure a model’s performance against real business outcomes is proprietary knowledge in its own right. Nadella’s argument is that without a legal or contractual equivalent to a patent, that knowledge simply flows one way, from the enterprise to whoever owns the model.