Own the Loop

Four situations, one structure

A SaaS company runs a repricing engine. Every customer who changes a price feeds the engine a data point, the engine gets sharper, and a sharper engine attracts more customers who feed it more data. A product manager watches AI absorb the parts of the job that were writing tickets and cutting scope, and finds the surviving value sits in judgment that compounds with reps. A vertical AI integrator wraps a foundation model with proprietary workflow data and pulls ahead of competitors renting the same model. A freight network gets denser with every shipment, which lowers cost per lane, which wins more shipments.

None of these are about the same industry. Three of them are about AI and one predates the transistor. They are the same structure. Usage produces data, data improves an asset, the improved asset drives more usage, and the entity that owns both ends of that loop compounds a lead that nobody renting a piece of it can close. That loop is where value accrues. Everything outside it commoditizes. The only strategic question that matters for any product is which layer is the loop, and whether you own it or rent it.

The compounding loop, defined

Call it the compounding loop. The mechanism is precise: usage generates data, data improves a proprietary asset, the improved asset drives more usage, and the cycle tightens. A repricing model that sees more price changes predicts better. A logistics network that moves more freight routes more efficiently. A recommendation surface that watches more sessions ranks better. Each turn of the loop makes the next turn cheaper and the output better.

Two conditions separate a real loop from a story a deck tells about one.

First, the same entity has to own both ends. If usage flows to you but the asset that improves belongs to someone else, you are running a treadmill that makes a supplier stronger. This is the trap most AI applications fall into. They generate enormous usage, and every session improves a model they do not own. The loop is real. They are just not the ones inside it.

Second, the advantage has to widen with scale, not plateau. A dataset that gets you to 95 percent accuracy and then flattens is a feature, not a moat. A loop where the marginal data point still moves the asset at scale is a moat, because the leader’s lead grows while followers spend to reach a bar that keeps moving.

The loop is not a network effect. Network effects are one special case, where the improving asset is the network itself and usage is participation. Metcalfe’s law is a compounding loop with a specific asset. Nor is the loop a data moat in the usual sense. A pile of data is residue. It sits there. The loop is the mechanism that turns fresh usage into a widening advantage, and residue with no mechanism attached is just storage cost. Packy McCormick’s vertical integrators and Ben Thompson’s aggregation theory are both descriptions of specific loops: the integrator owns the workflow asset, the aggregator owns the demand-and-supply matching asset. Same mechanism, different layer.

Everything else commoditizes

Here is the corollary, and it resolves a tension that has been running through every argument about AI moats. If value accrues to the loop, then everything outside the loop commoditizes. Not eventually. Structurally. The moment a layer stops being the thing that compounds, it becomes a thing you buy at market price.

So when someone asks whether foundation models are commoditizing, the answer is yes, and that is the point. The model was never the moat for the company building on top of it. The model is the commodity input. The moat is the workflow data the application accumulates, the customer relationship it deepens, the proprietary asset it improves with every session. Anyone treating the model itself as the durable advantage is defending the layer that was always going to be rented, at rates set by whoever owns the loop underneath it.

This is what the corollary is for. It tells you what to let go of. You do not fight to own the commodity layer. You do not spend to differentiate on the thing that gets cheaper every quarter. You identify the one layer where usage compounds into something only you hold, and you concentrate everything there. The commodity layers are where you buy the best available option and move on. Founders burn years and balance sheets defending layers that were never defensible, because they mistook the layer they were standing on for the layer that was compounding.

The question that changes your answer

The diagnostic is one question, asked about your own stack: where does usage compound into something only I own?

Trace a single unit of usage through your product. A customer runs a workflow. That generates data. Does that data improve an asset you own, or one you rent? If it improves your asset, follow the next turn: does the improved asset pull in more usage, and does the advantage widen as it does? If yes at every step, you own a loop. If the data improves someone else’s asset, you are the treadmill.

Then the second half, which is the part most operators skip: if I rent the layer that compounds, whose ceiling am I underneath? Because renting the loop is not neutral. It means your best possible outcome is capped by the economics of the owner, and they can raise the rent, integrate forward into your business, or simply keep the compounding gains for themselves while you pay for access. Every company built entirely on a rented loop is running a business whose ceiling belongs to someone else.

Most product teams have an answer to “what is our moat” that names a feature, a brand, or a head start. Run the loop question and the answer usually changes. The feature is a commodity layer. The brand is a lagging indicator of a loop that already ran. The head start decays unless a loop is compounding it forward. What survives the question is the one layer, often unglamorous, where usage turns into a widening, owned advantage. If nothing survives the question, that is the finding, and it is more useful than the comfortable answer you walked in with.

Assemble fast, but not everything

The obvious objection is that owning the loop is expensive and usually the wrong early move. Building the proprietary asset from scratch costs more than renting a good one, ships slower, and often loses to a team that assembled off-the-shelf parts and got to market first. Crawl, walk, run. Assemble fast, integrate later. This is correct, and I have argued the fast-assembly version of it more than once.

It is correct with one exception that swallows the rule. The loop you skip in the crawl phase can be the exact moat you cannot build later. If a competitor owned that loop from the start, they have been compounding while you assembled, and by the time you decide to integrate the layer you rented, they are several turns ahead on an advantage that widens with scale. You cannot buy your way back to a lead built out of data you never collected. So the discipline is not “own everything” and it is not “own nothing yet.” It is: own the one layer that learns, from the beginning, and rent the rest without guilt. Assemble fast everywhere except the loop. That is the layer where late is a synonym for never.

This was never about AI

The freight network has run this play since before software existed. Careers run it: the operator whose judgment compounds with every hard decision owns a loop that the operator collecting titles does not. Marketplaces run it: liquidity begets liquidity, and the platform that owns the matching layer widens its lead over every reseller renting access to its demand. The compounding loop is not a technology. It is the shape value takes when usage feeds back into the thing producing it, and it has been the shape for as long as some businesses got better by being used and others just got busier.

AI did not create the loop. AI made it impossible to ignore, because AI collapsed the cost of the parts around the loop to near zero and left the loop standing alone as the only thing worth owning. When everything assembles cheaply, the only question left is which layer compounds, and that question was always the game. The tools change every cycle. The layer that learns from being used is where the value has always gone, and where it will go long after the current tools are the commodity nobody thinks about.