The Reverse Rollup

The whole market is bidding up the one thing in the stack that gets cheaper every quarter. The model layer is the worst place to own value in this cycle. The market believes the opposite. The market is wrong, and the economics that make it wrong are structural, not soft.

Everyone is trying to sell the shovels. The better trade is to buy the mines. The consensus is clean and seductive. AI is the defining technology of the decade. The labs sit at its centre. So the money should flow to whoever supplies the intelligence: own the frontier model, own the tokens, own the future. Pour capital into the layer closest to the silicon and let everyone downstream rent from you. A tidy story. It is mostly wrong.

The Surplus Leaks Downstream

Start with the work of Soren Larson on AI value accrual, one of the few models of frontier economics that does not collapse into "the machines take the jobs." His argument runs through the shape of the task-value distribution. If the most valuable tasks are rare but worth almost anything, frontier labs hold pricing power, because customers will pay through the nose for those outputs. If the tail is thin, pricing power evaporates: open-source converges on the same capability for free.

Model it as a Pareto distribution. Two things fall out. Pricing power is governed by the tail parameter, alpha. Heavier tails support fatter margins on inference. Here is the twist. The same heavy tails that sustain lab margins push most of the surplus downstream. At an alpha around 1.5, the lab captures roughly a quarter of the value created. The other three quarters accrue to whoever owns the application layer and the domain workflow.

Say it plainly. Thin tails commoditise you. Thick tails enrich your customer. The lab cannot win either way. Thin tails, and open-source eats your lunch. Thick tails, and your customer eats your lunch. To defend margins the lab needs persistent complexity. Wherever complexity lives, the value escapes to the operator who has embedded the model inside a real, messy, domain-specific process.

Thin tails commoditise the lab. Thick tails enrich its customer. The intelligence cannot win either way.

Hayek Already Explained Why

This is where the argument turns Austrian. In "The Use of Knowledge in Society," Friedrich Hayek argued that the knowledge an economy runs on is decentralised by nature. It does not sit in one head or one central planner. It is scattered across thousands of people, each holding a sliver the planner can never assemble. The man on the spot knows what the man at the centre structurally cannot.

Map that onto the AI stack. The operator running the workflow knows which tasks actually matter, which edge cases are expensive, where the customer genuinely pays and where they only pretend to. The lab supplying the tokens knows none of this. It sees an API request worth a few cents. It cannot see the workflow stacked on top of that request, the one earning thousands. The token-seller is the central planner. The vertical operator is the man on the spot. Hayek told us, eighty years ago, which one holds the knowledge that matters. The model is a commodity input. The context is the asset, and the context is exactly what the lab is walled off from.

The Reverse Rollup

Here is the move. Call it the Reverse Rollup. The ordinary technology instinct is to build a new thing and roll the market up underneath it: a venture-funded platform that fans out and signs up the incumbents. The Reverse Rollup inverts that. You do not build the clever layer and acquire distribution later. You buy the distribution and the domain knowledge first, while it is cheap, and you install the intelligence from the inside.

The targets are deliberately boring. Legal services. Insurance brokerages. Logistics intermediaries. Healthcare administration. Accounting. Field services. Industries nobody posts about. Look at what these businesses hold. Fragmented markets, often thousands of small operators, none big enough to draw a bidding war. Owner-operators frequently within a few years of retirement and short on succession. Decades of workflow and pricing knowledge that no model has ever seen. And valuations that look quaint next to AI multiples: a brokerage changing hands at four to six times earnings while a pre-revenue model wrapper raises at a number with more zeros than logic.

You acquire that operator. The cash flow, the customers, the tacit knowledge, the regulatory footprint, the trust. Then you rebuild it from the inside, running the frontier model as a swappable input that gets cheaper and better every year, while you keep the one thing that does not commoditise: the domain. The labs are racing to sell the most replaceable component in the stack. The Reverse Rollup buys the least replaceable one and lets the labs subsidise your margin with every price cut.

This is why the 2026 strategy looks less like venture and more like private equity with AI leverage bolted on. Not a fund chasing the next model. A buyer chasing the next fragmented vertical, applying intelligence as a margin lever rather than selling it as a product. PE has always understood that owning cash flow and distribution beats owning the clever idea. AI does not overturn that lesson. It sharpens it.

Everyone is building castles on the layer that commoditises fastest. Buy the land underneath, or rent it from the people who did.

The Objection That Cuts at the Thesis

Rollups are where returns go to die. Grant it. Integration risk is real. Stitching twenty small acquisitions into one operation is brutal, unglamorous work, and the graveyard of rollups that tried to financial-engineer past operational reality is large and well documented. Owner-operator culture resists the spreadsheet. These businesses run on relationships and habit, not org charts. Buy one, swap the founder for a dashboard, and the customers walk out the door behind him. And the sharpest cut lands on the thesis itself: if the value sits in tacit, decentralised, Hayekian knowledge, then AI cannot easily extract it. The very thing that makes the operator valuable is the thing the model struggles to absorb. You can buy the business. You cannot guarantee you can digitise its soul.

These are serious. They are not fatal. Integration risk is precisely why this is private equity work and not venture work. PE firms have spent forty years building the operating muscle for exactly this: the playbooks, the bolt-on discipline, the management bench. Pair that competence with a falling-cost intelligence layer and you have something neither a lab nor a venture fund is structured to do.

Culture is a reason to move slowly and keep the operators, not a reason to stay out. The thesis was never to fire the man on the spot. It was to buy him out at the right moment, hand him better tools, and keep him in the chair. Retiring owner-operators are not defending the fort. They are looking for a credible exit, which is the whole opening.

Tacit knowledge resisting transfer is not a bug in the plan. It is the moat. If the labs could absorb the domain through an API, the operator would be worthless and the lab would already own the surplus. They cannot. That is the entire reason the boring business stays expensive to displace and cheap to buy. You are not betting that AI extracts the tacit knowledge. You are betting that owning the business that holds it lets you point the model at the rest.

The Close

The market prices intelligence as the scarce asset. The economics say it is becoming the abundant one. What stays scarce is the context the model cannot see: the workflow, the relationships, the man on the spot. The durable edge is the boring business the model has to be pointed at, and it is on sale.