Satya Nadella's July 2026 warning is blunt: companies may buy intelligence with cash and then donate their competitive edge through the prompts, corrections, and agent traces that make that intelligence useful.
What Nadella is naming
In his Reverse Information Paradox post, Nadella argues buyers pay twice: once with money, again with proprietary knowledge required to make models useful. Models learn from exhaust — prompts, tools, corrections, and evals — the kind of knowledge competitors cannot simply buy.
The enterprise response
Nadella calls for a hard trust boundary where nothing crosses without consent, including intelligence exhaust. For Canadian firms, that means contracts, private deployments, retention settings, and agent logging policies that keep institutional learning inside the company.
Practical controls this week
Classify which workflows may use public models. Route sensitive work to private AI. Disable training on enterprise data where possible. Store evals and corrections in company systems. Treat agent traces as IP, not disposable chat logs.