When the Moat Is a Wall

The AI capital stack has a structural problem. It is financing exclusion and calling it defensibility.

A moat is a competitive advantage built into what you make. A wall is a barrier erected around it. These are not the same thing, and in the current AI moment, a great deal of money is being spent on walls while calling them moats.

Here is what is actually happening. The U.S. AI capital stack was financed on a specific thesis: that the companies producing the most capable models would capture durable advantage, because the performance gap between the best closed systems and everything else would widen over time. That thesis has a falsification condition — and it is arriving. The gap between closed and open-weight models is not widening. It is closing, at a pace that several serious observers now put at six to twelve months. DeepSeek demonstrated that a research team operating at a fraction of the assumed budget could produce a model competitive with frontier closed systems. This is not a surprise to anyone paying attention to the structure of the field. It is an anomaly the thesis was not designed to survive.

The response has a name, though it rarely goes by one: regulatory enclosure. Restrictions on chip exports, data localisation requirements, national security frameworks applied to AI infrastructure — each of these is being advanced, with varying degrees of sincerity, as a security measure. And some of them may well be. But the structural effect, regardless of intent, is to recreate through regulation the scarcity that technology is no longer providing naturally. When the model weights are approaching commodity status, the instinct is to build a wall where the moat used to be.

The great and prematurely dead philosopher of science, Imre Lakatos, drew a distinction between what a theory actually claims — what he called “the hard core” — and the apparatus of adjustments and exceptions constructed to protect it from inconvenient evidence. A theory in good health updates itself when the evidence pushes back. A theory in trouble keeps adjusting the apparatus while leaving the core untouched, and hopes nobody notices. The AI exclusion thesis is exhibiting this pattern. It has stopped producing predictions about what the technology will do. It has started producing predictions about how the regulatory environment will be arranged to make the technology’s current limitations irrelevant. That is a different kind of claim, and a considerably weaker one.

The auto industry spent decades behind tariff walls and produced cars that required those walls to survive. When the walls came down, partially and fitfully, the industry discovered that protection had not built resilience. It had purchased complacency. The companies that emerged strongest were not the ones behind the highest walls. They were the ones that had, in the meantime, built something that did not need the wall to function.

A genuine moat in an environment where model weights approach commodity status looks like this: two years inside a specific enterprise’s operational reality, learning what their problems actually are at the precision that only sustained contact produces. The outcome signal you get when you can observe whether the recommendation actually moved the metric. Relationships and accumulated context that do not replicate when a competitor downloads the weights and do not erode when the next open-weight release closes the capability gap by another notch. These compound because they are anchored to something that does not generalise: the named problem of a specific enterprise, in a specific context, at a specific moment.

The wall is not the moat. The question worth asking right now is not how to make it higher — it is what you are building inside it that will still be standing when it comes down.