Who Owns the Learning? Rethinking Model Selection After the Reverse Information Paradox

For two years the enterprise question about AI has been which model is best. That was always the wrong question, or at least an incomplete one. The better question is who keeps the learning that happens once a company actually uses a model.

Satya Nadella has given the problem a name. In a short essay published on X, he described what he calls the Reverse Information Paradox. The old paradox, from the economist Kenneth Arrow, was the seller's problem: to sell information you have to reveal it, and once revealed it can be taken for free. Nadella flips the risk onto the buyer. In the AI age, you have to feed a model your proprietary knowledge to make it useful, and the better you want it to perform, the more of that knowledge you have to hand over. So you pay twice. Once in cash, and again in the know-how that made your company yours.

What the Paradox Actually Says

The argument is sharper than the usual worry about data leakage. Nadella's point is not that a database might spill. It is that the exchange itself is asymmetric. As he puts it, the seller learns more and more about you as you use what you bought, while you learn very little about what the seller is learning in return.

The knowledge does not escape through a breach. It escapes through what he calls exhaust: the prompts people write, the tools agents call, and above all the corrections people make when the model gets something wrong. Every correction teaches the model something a competitor could never buy, and it leaks slowly, trace by trace, correction by correction, eval by eval.

There is some irony in hearing this from Microsoft, which poured billions into OpenAI and built Copilot to reach into a company's mail, files and chat. Nadella acknowledges the double standard directly. AI labs claim fair use to train on the public web, then impose restrictive terms that stop customers from learning from model outputs. He thinks the innovation from that fair use is worth keeping. He also thinks the one-way flow is the problem.

The Thing You Are Really Giving Away

The most valuable thing most organisations own is not data in the abstract. It is the way their people decide and operate: which cases are routine and which are risky, which approval is real and which is ceremonial, how a good outcome is judged. That judgement is exactly what accumulates in the corrections and evals that a closed model quietly absorbs.

As AI moves into sales, support, engineering, finance and operations, a company can export the very edge it is trying to build without ever noticing the moment it happened. The danger is not a single decision. It is the drift of institutional judgement onto someone else's platform, where it compounds for them instead of for you.

The Fix Is Architectural, Not Just Contractual

Nadella's answer is a hard trust boundary inside the enterprise. It is the line across which a firm's data, traces, evals, adapted weights and memory accumulate together, and across which nothing crosses without consent, not even the exhaust. His recommendations cluster around five points: control your evals, because evals define what good looks like inside the organisation; keep learning environments inside your own tenant; keep the orchestration layer independent of any single model; retain the right to use model outputs to tune your own systems; and let all of it compound over time.

Read plainly, that shifts what enterprises care about. Branding matters less than control, capability, the ability to switch providers, cost, and whether value compounds inside the boundary or outside it. Model selection stops being a leaderboard exercise and becomes a question about ownership.

Why Open Weights Change the Commercial Logic

This is where open-weight models matter. Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, has just released its first model, Inkling. It is a mixture-of-experts system with 975 billion total parameters and about 41 billion active per query, natively multimodal across text, image and audio input, and released under an Apache 2.0 licence with weights available to download. The company is explicit that Inkling is not the strongest model available, open or closed. That is the point. It is positioned as a starting point that organisations fine-tune themselves through Tinker, the firm's customisation platform, rather than a finished product they rent.

The distinction is commercial, not just technical. Instead of renting intelligence from a black box, a company can build a private layer of adaptation around a foundation model and decide what stays inside. That makes it possible to train on your own workflows, build domain-tuned assistants, and keep the feedback loop that a closed system would otherwise absorb. Thinking Machines points to work with Bridgewater Associates, where an open model trained further on the hedge fund's own financial expertise beat leading proprietary models on the firms' own financial reasoning tests at a fraction of the running cost. The result comes from their own evaluation rather than an independent one, so treat the exact numbers with care, but the shape of the claim is the interesting part: adaptation on private expertise can outperform a bigger general model.

The Market Fragments

The likely result is a more divided market rather than one winner. Frontier closed models will stay attractive where speed and convenience matter, and for general-purpose work where the task is not sensitive. Open-weight systems will pull ahead where governance, auditability and ownership of the output matter more.

Hugging Face's Clem Delangue has made much the same prediction: frontier models reserved for experimentation and high-value tasks, with most production work moving to private or open alternatives. For a lot of enterprises the sensible stack is hybrid. Closed models for commodity tasks, open models for sensitive or strategic workflows, and a private orchestration layer sitting above both so the choice stays yours.

The Fight Over Distillation

This also explains the growing tension over distillation. If a vendor trains freely on public and interaction data but forbids customers from learning from its outputs, enterprises will push back, and some already are. Alex Karp at Palantir made the same complaint in blunter terms around the same time. The buyer expectation is moving from "we use your model" to "we can learn from what we buy". A contract that blocks that will increasingly look like a reason to choose something else.

What Model Selection Becomes

The future of enterprise AI is not only bigger models. It is clearer ownership. The question a buyer should ask is no longer just how a model scores. It is whether the model helps the organisation keep its own intelligence, or whether it quietly monetises that intelligence somewhere else.

On that test, model selection is really a decision about where the learning loop lives. The organisations that treat a model as a one-off tool will keep paying the second price without noticing. The ones that treat it as a learning system they can hold inside their own boundary will be the ones still compounding an advantage a few years from now.

References

• Satya Nadella, "The Reverse Information Paradox," essay published on X, 13 July 2026.

• The Next Web, "Nadella says you pay for AI twice, and Microsoft helped build the trap," July 2026.

• diginomica, "Tokenomics: five C-words from Satya Nadella," July 2026.

• Thinking Machines Lab, "Introducing Inkling," and Inkling model card, 15 July 2026.

• TechCrunch, "Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling," 15 July 2026.