Own your denominator.

Two paradoxes, one gap

Economists have long known a strange thing about selling information: the buyer cannot judge its value without seeing it, and once he has seen it, he has effectively taken it. That is Kenneth Arrow's information paradox, and patents exist largely to soften it. Microsoft's Satya Nadella recently named its mirror image for the AI age, the Reverse Information Paradox: now it is the buyer who gives knowledge away, feeding proprietary context, corrections, and judgment into a model simply to make the purchased intelligence work.

Read through this book, what he is describing is a gap between two clocks, drawn not inside one system but between two parties. The vendor's clock of learning-about-you compounds with every prompt, every correction, every eval you run. Your clock of knowing-what-they-learned barely moves. Asymmetric learning rates, a widening gap, and value pooling on the fast side of it: this is the acceleration paradox restated as a commercial relationship. And as always, the danger is not either clock. It is the gap.

Your evals are your oversight half-life

The book's flagship instrument, the oversight half-life (H), measures how long your understanding of a system stays trustworthy. Ask what produces that understanding inside a firm using AI, and the answer is precise: your evals, which define what "good" looks like for your work; your corrections, which are your detection of drift, caught and fixed; your institutional memory of what was decided and why. These are not by-products of using a model. They are the firm's capacity to detect and correct, which is to say, they are the denominator of R = L / H.

Now run the equilibrium law over the standard arrangement. If those evals and corrections stream across the boundary and improve someone else's system, three of the book's five gears grind at once. Equity fails: the firm creates the learning and bears the exposure, while the upside compounds with the owner of the learning infrastructure. Steering fails: if the model can be withdrawn, repriced, or changed and your accumulated judgment lives inside it rather than beside it, you cannot change direction, only comply. And Resilience fails quietly: the institution that stops holding its own corrections eventually loses the ability to make them.

A boundary is a promise until you can check it

The emerging answer, in Nadella's telling and in the market at large, is a trust boundary: a perimeter inside which a firm's data, traces, evals, and adapted weights accumulate together, and across which nothing passes without consent. The instinct is exactly right. But the book has seen this clause before. It is the enterprise version of the pause clause in AI governance: a commitment that is empty until the other side can verify it.

"We do not train on your data" is today mostly a sentence in a contract. The firm cannot see the vendor's training runs any more than one nation can see another's. And the book's answer is the same at both scales: trust cannot be a virtue; it has to be infrastructure. Attestation of what actually ran. Proofs about bounds rather than contents. Visibility that flows both ways, so the party being trusted is also the party being seen. This is precisely the grammar of the Verifiable Compute Commons, pointed inward: the same machinery that could let rival labs verify each other's restraint is what would let a customer verify a vendor's, and turn the trust boundary from marketing into physics.

The Monday version

Strip the vocabulary away and the equilibrium reading of the enterprise AI question gives a checklist a leadership team can actually run.

Own the instruments, not just the data. Your evals, your correction logs, your decision memory: keep them in a form you control and could carry to any model. They are your H. Guard them like the crown jewels they quietly are.

Apply the steering test before you depend. The book's question for any system after launch: can you still change direction, or only watch? For AI vendors it becomes concrete: if this model vanished tomorrow, does your accumulated capability remain yours, and can your workflows run on the next engine?

Price the second payment. When you evaluate an AI contract, account for the knowledge that will cross the boundary, not only the fee. If the learning flows one way, the true cost is the fee plus a slow transfer of your alpha.

Demand verifiable boundaries, not promised ones. Ask the vendor the book's question: how would we know? What attestation, what audit, what proof about bounds? A vendor who cannot answer is asking for faith, and faith is not a gear.

The firms that thrive in the age of intelligence will not be the ones that used the most of it. They will be the ones that used it fast while keeping their power to detect, correct, and walk away. Velocity with vigilance was never only advice for labs and lawmakers. It is a balance sheet item now.

Attribution: the term "Reverse Information Paradox" and the trust-boundary framing are Satya Nadella's, from his public 2026 essay on enterprise AI and intellectual property; Arrow's information paradox is from Kenneth Arrow's classic work on the economics of information. The reading of both through the control ratio is the book's, and the argument here is independent commentary.