Industry
A Moat Is What Does Not Get Retrained
Three years of debate on AI moats, and a thesis chased out every six months by the next. That instability is itself a clue: the moat is being sought in the wrong place.
We have spent three years debating moats in AI, and the debate has never stayed put. Some placed the defensible advantage in model size, others in training data, others in compute, others still in speed of execution or fundraising. Each thesis had its hour, drew convinced articles and considerable funding rounds, then was swept away by the next. This waltz is not the sign of a still-young debate that will eventually settle. It is the simpler, more disturbing sign that we are looking in the wrong place, and that no improvement of the gaze corrects an error of direction.
A thesis that falls every six months
When an explanation has to be replaced every six months, the problem is generally not in the explanation, but in where it is sought. If every moat candidate proves crossable in one or two model cycles, then none of them was a moat; they were leads. And a lead, by definition, gets caught up. The very rhythm at which the theses succeed one another should have raised the alarm sooner: it is not the market that changes too fast, it is the definition that is ill-posed. We looked for moats where there were only head starts, and were surprised to see them closed.
When a moat thesis falls every six months, the problem is not the thesis. It is where you are looking.
The confusion comes from calling a moat anything that provides an advantage, when an advantage and a moat are not of the same nature. An advantage is a favorable position at a given instant. A moat is what prevents a competitor from retaking that position, whatever its means. The first is measured today; the second is measured over time, against everything a well-funded rival might attempt. To confuse the two is to take a lead for a protection, and to think yourself safe at the precise moment you are most exposed.
The distinction is not merely academic: it changes how one invests. Whoever believes he holds a moat when he has only a lead rests on his position instead of defending it, and discovers, at the next cycle, that what he took for a moat was only a head start the competitor has just closed. Conversely, whoever knows he holds only a lead stays lucid about its fragility and seeks to turn that lead into something that, precisely, does not get caught up. The whole question is to know what, in enterprise AI, belongs to that second category.
The rule that sorts
A simple rule nonetheless identifies a moat in any technology sector, and it cuts short most illusions: a moat is what does not get retrained. Anything that can be reproduced by relaunching a training run, repurchasing a dataset, copying an architecture, is not a moat. It is a lead, sometimes comfortable, always temporary. The rule is convenient in that it tests itself: for each candidate, it suffices to ask whether it would survive a determined, well-resourced competitor relaunching the process.
Applied methodically, this rule eliminates the false moats one after another. A larger model? It gets retrained, and the competitor will do it. Superior training data? It gets acquired, licensed, reconstituted. A clever architecture? It replicates as soon as it is known, and it always ends up known. Massive funding? It is matched by a funding round. A time lead? It erodes with every passing month. An exclusive partnership? It expires, or is bypassed. For each candidate, the same question decides: can a determined competitor reproduce it by relaunching a process? If yes, set it aside. It was not a moat. The rule is brutal, and its brutality is what makes it useful: it lets no comfortable illusion through.
If a competitor can retrain it, repurchase it or rebuild it, it was not a moat. It was a lead.
One may ask why this reflex of seeking the moat in the technology itself is so tenacious, when the rule dismantles it so easily. The reason lies in the recent history of software, where technology was long the moat: holding a proprietary algorithm, a closed format, a brick no one could reproduce, was holding a defensible position. The reflex formed there, and it was rewarded. But AI overturns that regime: technical capabilities now spread at a speed that forbids any of them from staying rare. To keep seeking the moat in technology is to apply, to a world where everything replicates fast, a reflex forged in a world where technology stayed captive. The reflex is not absurd; it is simply obsolete.
What then remains, having survived the test, is instructive. What does not get retrained, in enterprise AI, is everything that accumulates in an organization’s real use and does not transfer with a mere change of tool: an operational memory built matter after matter, proven methods, coherent positions, a context that has sedimented over time. It does not get relaunched, because it was not manufactured; it was lived. And what was lived by one organization cannot be downloaded by another, nor reconstituted by a competitor, however rich. It is the only thing, in the whole stack, that answers the test positively.
The moat has changed owners
There is, in this shift, a reversal worth naming, because it inverts a decades-old habit. For a long time, the moat belonged to the vendor: it was the vendor who held the rare technology, the proprietary know-how, the asset the client could not reproduce. The client rented access to that rarity, and his dependence on the vendor was the counterpart of the advantage he drew from it. With AI, the rarity changes sides, and that change of sides is perhaps the most underestimated strategic event of the period.
Because what does not get retrained is not held by the model vendor, who is himself replicable, but by the organization that accumulates above him. The moat no longer sits in the tool; it sits in the use. It is no longer sold; it is built, and whoever builds it owns it. The vendor, in this scheme, is no longer the holder of the moat: he is the provider of the ground on which the client builds his own. It is a complete inversion of the value chain, where the defensible asset migrates from seller to buyer.
This inversion has an implication organizations should ponder before choosing their tools. If the moat is now built on the client side, then the right selection criterion is not the vendor’s power, but what the tool lets the organization accumulate and keep as its own. A tool that produces a great deal but lets nothing accumulate enriches the vendor, not the client. A tool that constitutes, on the client’s side, an accumulation layer that belongs to it, builds it a moat. The difference is invisible at purchase; it is decisive in the end.
The moat of enterprise AI does not belong to the model vendor. It belongs to the organization that accumulates above.
This reversal does not please everyone, and one understands why. It deprives vendors of the comfortable position they held, that of the holder of rarity to whom an access right is paid. But it opens, for organizations, a possibility they had never had: to constitute an advantage that belongs to them in their own right, that no change of vendor removes, and that no competitor can buy. They must still choose their tools accordingly, that is, according to what those tools let them keep, and not what they make them produce.
It is this conviction that structures the architecture of MAX. The Legal Semantic Layer is not designed to constitute a moat for a vendor’s benefit, in the classic sense of a rarity one rents, but to give the organization the place to build its own: the layer where what precisely does not get retrained accumulates, belongs to it and stays with it. MAX’s role is not to hold the firm’s moat. It is to provide the ground for it.
