The Moat Is the Data, Not the Model

Building an AI operation is one thing. Keeping it defensible is another, and the honest answer to "what's the moat?" is usually not the thing companies reach for first.

It is not the model.

The model is the one part of the stack everyone can rent. That matters more than it sounds, because the economics have moved fast.

The model is a commodity now

Two things happened at once. The price of a given level of AI capability collapsed, and the capabilities themselves converged.

On price, a16z's analysis found that the cost of running a model at a fixed level of performance falls by roughly 10x a year. That decline is faster than compute during the PC era or bandwidth during the dotcom boom. The practical effect is that yesterday's frontier is nearly free today. Independent studies put the capability that was flagship-class a few months ago at around a fifth of the current flagship's token price. Whatever model advantage you have, time erodes it toward zero.

On capability, the frontier has bunched up. By Stanford's 2026 AI Index, the leading model as of March 2026 led the field by just 2.7 percentage points, and open and closed models have traded the top spot repeatedly. When the best models are within a rounding error of each other, and the one behind yours is cheaper, choosing a model is a procurement decision, not a strategy.

So a company built on having the best model is building on a moving target. Whatever edge the model gives, everyone can reach it, and reaching it gets cheaper every quarter.

But data alone isn't a moat

The usual answer is "our moat is our data," and most of the time that isn't true.

A dataset you collected once is a snapshot. It describes the past, a competitor can often assemble something similar, and incremental data runs into diminishing returns, a point a16z made years ago that has only gotten truer. Internal logs, support tickets, and raw transcripts are technically proprietary, but if they don't predict a business outcome, they're storage, not a moat. Almost every company can say it has a lot of data, which is exactly why having data defends almost none of them.

Having data is not the same as having an advantage from data. The advantage is in what the data does.

The moat is the feedback loop

The durable version isn't a dataset. It's a loop.

The pattern is simple to describe and hard to build. The system does real work, and that work produces proprietary signal: what converts and what stalls, which objection comes up and what answers it, which customer qualifies and which one wastes the hour. That signal feeds back into the system, which does the next round of work slightly better, which produces sharper signal. Usage improves the product, the better product earns more usage, and the gap between you and a new entrant widens with every cycle.

That is the asset a competitor can't copy by switching models, because it isn't in the model. It's in the accumulated, operation-specific knowledge of what actually works, learned by running the operation thousands of times. It takes years to build, which is the point. A well-funded rival can rent your model tomorrow. It can't rent the two years you spent learning from real operations, and it can't shortcut the learning, because the learning only happens by operating.

The strongest form of this is an AI that becomes the operating system for a particular business process, the place the work actually runs. Once it is, the data writes itself, the loop turns on its own, and the moat deepens as a byproduct of the business doing what it already does.

How we build

For Muan Group, this is why the portfolio is built around operating systems, not features.

Global Connect runs the outcome, so it learns the outcome. Every call MIA places, qualifies, books, or loses is a data point about what moves a real homeowner from a lead to a show-up. The model that powers the conversation is rented and replaceable. The record of what works, built from running the operation at volume, is ours.

ASIO compounds the same way across marketing. Each campaign a user runs through the platform teaches it a little more about what converts, for whom, and at what cost, and that learning makes the next campaign better. The software is the tool; the accumulating signal is the asset.

We optimize for the loop, not the model. We assume the model underneath will keep getting cheaper and better, for us and for everyone. What we invest in is the operation that turns each day of real work into a system that runs the next day better.

A quick note on what this data is and isn't. The signal that compounds is operational, what converts, what objection recurs, what a good outcome looks like, not anything that depends on holding personal records a competitor can't. The moat is learned know-how about the work, not a pile of private data.

The model layer will keep commoditizing. Capability will keep converging and prices will keep falling, which is good for everyone and decisive for no one. The advantage goes to the companies that have spent years running real operations and turning what they learn into operations that run better. The model is rented. The loop is owned. That is the moat.