The last moat isn't data.
It's your learning loop. Three Nobel economists explain the risk every enterprise runs when it buys intelligence. And the data shows the market already splitting into those who will pay it and those who won't.

In July 2026, Satya Nadella gave a name to a discomfort that enterprise leaders had felt for two years without being able to articulate it. He called it the Reverse Information Paradox. It is the sharpest framing yet of a real problem, and it deserves to be taken seriously because of who said it. It is also, in one crucial respect, not precise enough, and the imprecision is exactly where enterprises lose money. Read it as a diagnosis of symptoms, accurately observed. What it does not yet name is the mechanism, or the bill.

Three Nobel economists frame the whole question. In 1962, Kenneth Arrow described the original information paradox: the value of information cannot be known to a buyer until it is revealed, but once revealed it has effectively been acquired for free. The seller carries the risk. Artificial intelligence inverts that. To make a purchased model useful, an enterprise must feed it the knowledge that makes the enterprise distinctive: its prompts, its workflows, its evaluations, and above all its corrections. The buyer now carries the risk, and pays twice, once with money and again with knowledge.

The more important economist here is Oliver Williamson, who won the Nobel in 2009 for describing what happens next. Williamson showed that when a buyer makes investments valuable in one relationship but worthless outside it, a condition he called asset specificity, that buyer becomes vulnerable to being squeezed. He named it the hold-up problem. That is what the second payment actually buys: not the theft of your knowledge, but a growing dependency a supplier can later price against you.

THE SECOND PAYMENT  ·  The hidden cost of enterprise AI that contracts cannot prevent. The first payment is the invoice, and procurement already knows how to manage it. The second is paid in dependency: institutional judgment migrating, decision by decision, into a learning loop the firm does not own. Its mechanism is Loop Capture. Nadella named the symptom. What follows names the mechanism, prices the cost, and sets out the architecture that ends it.

The question of the decade is not who builds the best model. It is who owns the loop in which the model gets better at your work.

That this warning came from the chief executive of the company that sells more enterprise AI than any other is worth pausing on. It is credible precisely because it runs against interest, and Nadella is explicit that the problem requires more than data protection. It is also fair to note that the firms naming the risk are positioned to sell the remedy, in sovereign tenancy and governance tooling. Neither observation makes the paradox less real. Both argue for understanding it on the buyer's terms, not the vendor's.

01  /  The mechanismIntelligence exhaust, and the loop it feeds

The knowledge that leaves is not the data enterprises already know how to protect. It is a subtler byproduct, what Nadella calls intelligence exhaust: the prompts people write, the tools agents invoke, and most valuably the corrections people make when a model is wrong. Every correction encodes a unit of institutional judgment, what an organization considers good and how it measures success, and it drains away one trace, one correction, one evaluation at a time. This is what Nadella, reaching for Hayek, calls a firm's particular knowledge: the understanding of time, place, and circumstance that no one else can hold. These corrections do not sit still. They feed a loop, and the loop is the thing that matters.

02  /  The distinction"But they promise not to train on our data"

Here is where most treatments collapse, and where the argument must earn its keep. The strongest objection is factual and correct: on the major business and API tiers, from OpenAI and Anthropic to Microsoft and Google, customer inputs are contractually excluded from training, retention is short, and zero-data-retention terms exist for regulated workloads. For properly configured enterprise usage, "your data trains their model" is simply wrong.

Why learning compounds

Why does that flow matter more than the data it carries? Because data was never the scarce asset. Data tells you what happened. A model tells you what is probable. Organizational judgment decides what your company should do, and judgment is the one input on that list a competitor cannot buy.

Strategy research reached this conclusion long before AI made it urgent. Edith Penrose showed that firms grow on accumulated productive knowledge, not on the resources they happen to hold. Nelson and Winter showed where that knowledge lives: in routines, the half-tacit patterns of how work actually gets decided and corrected, which function as an organization's memory and are stubbornly hard to imitate. David Teece showed that advantage belongs not to the firm with the best assets but to the firm best able to keep reconfiguring them, what he called dynamic capabilities. Put the three together and the conclusion is uncomfortable for the way most enterprises now buy AI: knowledge is the growth engine, routines are where it accumulates, and the capacity to repeatedly improve decisions is the advantage itself.

The learning loop is exactly that capacity, made operational and accelerated. Each pass converts a correction into a routine, a routine into an evaluation, an evaluation into a better next decision. The causal chain is explicit and testable: AI adoption feeds the learning loop; the loop accelerates organizational learning; faster learning produces better decisions; better decisions produce better execution; and execution that improves faster than a rival's is what a durable advantage is. Break the chain at any link and no advantage forms. Own every link and it compounds.

One nuance keeps the claim honest. Owning the loop does not guarantee advantage; a firm can own its loop and still learn slowly. Ownership is the precondition, not the prize. It determines whose account the compounding accrues to, and a firm that rents its loop can execute brilliantly while building someone else's asset.

The pattern has a lineage every executive already knows. ERP captured transactions. CRM captured customers. Cloud captured infrastructure. AI captures organizational judgment, and judgment is the first captured layer that compounds with use. The earlier platforms took custody of your records. This one takes custody of your rate of learning. That is why this shift is different in kind, not merely in degree.

If the paradox were only about training, that objection would end it. It does not, for two reasons. Nadella himself points at the first: even where enterprise inputs are excluded, providers impose restrictive terms on distillation and reserve rights over usage and interaction data, so learning still flows one way at the ecosystem level. The second is deeper, and it survives even a perfect contract. Separating the two versions is the whole game.

A definition, stated so it can be tested and quoted: Loop Capture is the migration of an organization's learning, its evaluations, corrections, memory, and decision heuristics, into execution paths that depend on a single provider, so that the learning keeps compounding but the organization no longer owns the compounding. Lock-in describes the cost of leaving. Loop Capture describes what you lose by staying.

03  /  The evidenceBoth versions, in the data

The evidence splits cleanly. Training Leakage shows up as shadow AI on ungoverned tiers. Loop Capture shows up as the depth of dependency being accumulated. Here is the shallow version first, because it is the better-measured of the two.

The deeper version is harder to measure, because it is newer and no one sells a dashboard for it. But its proxies are already loud: 76 to 81 percent of enterprises report concern over proprietary dependency in model, memory, and orchestration, and when a platform collapses, the migration bill arrives within weeks, as the Builder.ai case below shows. That is asset specificity presenting its bill.

04  /  The counter-caseWhy this thesis is falsifiable, and already contested

An argument that cannot be wrong is not worth making. So here is the strongest evidence against alarm: enterprises are not sleepwalking into capture. They are diversifying, fast. If that trend wins, Loop Capture recedes on its own.

So why is this not already solved? Because diversification at the plumbing level does not automatically protect the loop. Routing a query to whichever model is cheapest is not the same as owning your evaluations, memory, and corrections independently of any model. A firm can run twenty models and still have its institutional judgment tuned to, and trapped inside, one provider's environment. The bifurcation is real, but the dividing line is not "how many models do you use." It is "where does your loop close."

Now run the counterargument all the way to its limit, because its endpoint is instructive. Suppose the optimists are entirely right. Suppose every frontier model converges in capability and price until intelligence is a pure commodity, as interchangeable as electricity. What, in that world, differentiates one enterprise from another? Not the model; everyone has the same ones. Not raw data, which is increasingly abundant rather than scarce. The only remaining differentiator is the rate at which an organization converts commodity intelligence into better decisions, which is precisely the learning loop. Commoditization does not weaken the thesis. It completes it: the cheaper and more interchangeable intelligence becomes, the more the loop is the only durable asset left standing.

The thesis, stated so it can be tested: firms that own the loop compound advantage; firms that merely diversify their models do not. The next three years will show which group grows faster.

05  /  The precedentWhat the casualty list actually teaches

The public cases split along the same line, and reading them correctly matters. The 2023 bans were the reflex to Training Leakage, and they were largely resolved: Samsung, Apple, and the banks now run governed enterprise AI. The unsolved cases are the 2026 ones, and they are all Loop Capture.

06  /  The exposureFive risks on the board's desk

07  /  The India lensWhy the Global South pays a steeper Second Payment

The paradox is universal, but the bill is not evenly split. For India and much of the Global South, three forces compound it.

First, regulation arrives as capability scales. India's Digital Personal Data Protection Act reaches full enforcement on 13 May 2027, with penalties up to ₹250 crore per violation, and roughly 83 percent of organizations have not begun comprehensive implementation. A regime that puts primary accountability on the data fiduciary, on a strict consent basis with no separate category for sensitive data, turns uncontrolled exhaust into direct legal liability.

Second, the sovereign build itself runs on rented foundations. The IndiaAI Mission has committed roughly $1.25 billion and deployed tens of thousands of GPUs. Yet the silicon is almost entirely foreign, and the frontier models most enterprises deploy are trained abroad. Sovereignty at the infrastructure layer can coexist with dependency at the model and loop layers. That is Loop Capture wearing a national flag.

---

08  /  The trajectoryWhat this becomes by 2030

Two numbers, taken together, define the risk window. The market is racing to deploy autonomous agents, and it is nowhere near ready to govern them.

Follow the money, and the direction is unambiguous. The agentic and sovereign-AI markets are both being built to resolve this very tension.

Read together, these are not three trends. They are one: the market discovering, at scale, that it climbed onto the consumption rung of the AI Maturity Curve, and beginning to relocate capital toward the compounding rung, where the loop is owned. That relocation is the whole investment thesis of the next five years.

09  /  Five years outTwo enterprises in 2031

Extrapolate the two forces, capability converging and specificity compounding, and by 2031 the enterprise landscape has split in two. The dividing line is not size, sector, or which model a firm chose. It is the single decision made years earlier about where the loop would close. The same technology produces two opposite outcomes.

Both enterprises bought the same intelligence from the same vendors. The difference is entirely architectural, and it was decided long before it became visible on anyone's financials.

10  /  The diagnosticThe Second Payment Exposure Test

Concepts do not change behavior; diagnostics do. These are the six questions I put to technology leaders to locate an organization on the curve. None asks which model you use.

11  /  The responseThe trust boundary, in five disciplines

Arrow's paradox found its institutional answer in the patent, which let inventors disclose without surrendering. The Reverse Information Paradox needs an architectural equivalent, the trust boundary: a hard perimeter where an organization's evaluations, memory, traces, adapted weights, and institutional context accumulate and improve together, and where nothing, exhaust included, crosses without consent.

The reason the response is architectural rather than contractual is worth stating plainly. The loop runs on five assets: memory, evaluations, orchestration, retrieval, and the feedback path that turns corrections into improvements. Whoever owns those five owns the enterprise learning loop, regardless of which model executes underneath at any given moment. Models are the interchangeable part; the loop is the accumulating part. A contract can promise good behavior from the model layer. Only architecture decides who holds the accumulating layer, and that is the design brief the disciplines below execute.

None of this is all-or-nothing, and it should not be. Owning the loop is expensive, so spend the effort where it pays. Apply the full discipline to the handful of crown-jewel workflows that encode durable advantage, and rent freely for commodity tasks where switching cost is trivial. The test is simple: if a workflow accumulates judgment a competitor could never buy, the loop belongs inside your boundary. If it does not, let the market compete for your spend.