The warning came from the company selling the tools
The most pointed warning about buying frontier AI this month came from the man whose company sells it to you. On 13 July 2026, Microsoft chief executive Satya Nadella published a short essay arguing that enterprises adopting proprietary AI models are paying for intelligence twice: once in cash, and once in the proprietary knowledge they have to hand over to make the model useful. Microsoft ships Copilot and holds stakes in both OpenAI and Anthropic, which is precisely why the argument is worth reading twice.
Nadella called it the Reverse Information Paradox, a nod to the economist Kenneth Arrow, who observed that you cannot judge the value of information until you have revealed it - at which point you have already given it away. Applied to AI, the point is uncomfortable: the better you want a model to perform on your business, the more of your business you have to feed it.
Two prices, and only one shows on the invoice
The cash price is the one everyone budgets for; the second price is the one that compounds. Models improve on what Nadella calls exhaust - the prompts your people write, the tools your agents call, and above all the corrections your staff make when the model gets something wrong. Each correction is a small, distilled piece of institutional know-how that no competitor could buy on the open market.
Feed enough of it in, and the provider ends up holding a refined version of how your company actually works. Nadella's sharper claim is about where that leads. A firm that starts as an infrastructure supplier can grow into an application provider, and an application provider can end up competing with the very customers whose usage taught it the domain.
He also flagged the double standard. The large labs train freely on the public web, then restrict others from distilling their models in return. An enterprise handing over its hardest-won operational detail is on the wrong side of that same asymmetry.
Why this is a procurement decision, not a philosophy
Strip away the theory and this is a procurement decision, not a philosophical one. The question a buyer should ask is no longer only which model is smartest today. It is who keeps the learning that accumulates while you use it, because that learning - not the current benchmark score - is what is worth defending.
The test is simple and unforgiving. If you could not move your workload to a different model within a week without losing the corrections and context your team has built up, you do not own that value; your provider does. That is lock-in dressed as convenience, and it deepens every month you postpone dealing with it.
For a European operator there is a second edge to it. The same instinct that keeps your operational data out of a single vendor's training loop also keeps it inside your own jurisdiction, which is the harder half of any data-residency or GDPR conversation to satisfy after the fact.
Design the exit before you sign
The fix Nadella points to is boring, which is what makes it credible. Put an orchestration layer or gateway in front of the model so switching providers is a configuration change, not a rebuild. Keep your prompts, corrections and agent logs in a store you control, so the compounding edge stays yours. And run an open-weight model on your own hardware for the routine work - it will handle something close to 90 percent of the load at a fraction of the per-token cost, which in euro terms is the difference between a rounding error and a line item.
Reserve the expensive frontier model for the slice that genuinely needs it. Nadella did not tell anyone to stop buying AI; he told them to stop paying the second price without noticing. The owners who act on that keep the intelligence and the know-how. The ones who do not will find, a few years from now, that the smartest thing in their market was built out of their own corrections.
Read next: Buy the Wall, Not the Robot | Halo Killed the Engine That Ate Its Studio
