Industry
What Remains When a Layer Commoditizes
Every technological wave follows the same trajectory: the lower layer commoditizes, value rises a notch. Models are entering this phase now.
There is a value trajectory that every major technological wave has followed, and that is worth naming clearly, because it is replaying before our eyes with language models. It holds in one sentence: as a technology commoditizes, value stops residing in it and migrates to the layer built above. What was rare becomes a given; what was above becomes the stake. To understand this trajectory is not to forecast: it is to read the present in the light of a pattern that has already repeated often enough to be trusted, and to be surprised that it remains so little anticipated.
The rule every wave has verified
The trajectory reads in the recent history of software, provided you look not at one wave, but at their succession. As long as hardware was hard and costly, it carried the value; when it commoditized, value rose toward operating systems and databases. As long as databases were a reserved art, they carried the value; when they in turn commoditized, value rose toward applications. At each floor, the same thing repeats: the lower layer becomes a service you consume without thinking, and attention, money, differentiation shift toward the immediately higher layer.
One must dwell on the mechanism, because it is what makes the trajectory predictable rather than fortuitous. A technology carries value as long as it is rare, that is, as long as mastering it distinguishes whoever masters it. As soon as it spreads, as it becomes available to all at the same cost, it mechanically ceases to distinguish anyone: everyone has it, so it no longer separates anyone. Value does not thereby disappear; it moves toward what, now, makes the difference, and that something is always one notch above, in the way the now-common layer is assembled and coordinated.
What strikes, in this succession, is its regularity. It is not an accident specific to one technology; it is a groundswell, almost a reversed law of gravity, where value rises instead of falling. And a law of such constancy has no reason to stop at the present generation. Whoever bets that models will be the exception bets against everything the history of software has shown, with no argument for why, this time, the rule would not hold. The exception would be remarkable; it would demand a justification; and no one, to date, has provided it.
At each wave, the lower layer becomes a given you replace. The upper layer becomes the asset you keep.
A clarification is needed, because the word commoditize often misleads. To commoditize does not mean to lose importance; it means to cease making the difference. A commoditized layer remains indispensable, often more indispensable than ever: nothing can be built without it. But since everyone has it at the same price, it no longer distinguishes anyone, and advantage is built elsewhere. Commoditized hardware did not disappear, it is everywhere; the commoditized database runs the whole world. Their commoditization was not their erasure, but their passage from the status of stake to that of base. This is exactly what awaits the models: not to disappear, but to become the base on which others will build what, henceforth, makes the difference.
Why models enter it now
Language models are entering this phase today, and the signs are now hard to ignore. The raw performance of large models is converging: the gaps that made headlines two years ago are narrowing, capabilities are spreading, what was a spectacular advantage yesterday becomes a shared standard. This does not mean models lose importance, on the contrary: they become a base infrastructure, omnipresent. But becoming a base infrastructure is precisely ceasing to carry differentiation, as the processor ceased to carry it by becoming universal.
When a capability becomes available everywhere, at a collapsing cost, it mechanically ceases to be what distinguishes one organization from another. All have access to it. The question is no longer who has the best model, but who makes the best use of it, and that use is built in the upper layer. There, for AI, replays the trajectory every previous wave has traced. The model follows the path of the processor and the database: indispensable, omnipresent, and precisely for that reason unable, on its own, to distinguish anyone.
One can even date the tipping point to a simple symptom: the moment the choice of model stops being a strategic decision and becomes a procurement one. As long as one model was clearly superior to the others, choosing it was an act of strategy. As models converge, that choice commoditizes: you take one, you will change it, that is no longer where the advantage plays out. And the day the choice of model becomes interchangeable, all strategic attention shifts to what does not interchange: the layer the organization has built above.
The asymmetry few buyers measure
This rise of value has a strategic consequence that is poorly measured, because it is invisible at the moment of purchase. Two organizations can start with exactly the same model, the best of the moment, and find themselves, eighteen months later, in situations with no common measure. The difference will not lie in the model, which they shared, but in what each will have accumulated above: one will have built an operational memory, methods, a coherence; the other will have produced, a great deal, while retaining nothing.
The scenario deserves to be unrolled, because it is in unrolling it that one grasps its reach. The two organizations start identical, with the same tool and the same promise. The first treats each matter as an isolated event: it produces, it delivers, it moves to the next, and what it has learned evaporates with the session. The second inscribes each matter in a layer that retains: decisions accumulate there, positions stabilize, methods consolidate. After eighteen months, the first has produced as much as the second, perhaps more. But the first has only a history of acts; the second has an asset that works for it, and that the first cannot procure, because it is not for sale.
There lies the asymmetry. What sits in the upper layer is not manufactured, it is accumulated, and accumulation needs time. That is what makes it formidable. A performance gap closes fast: just change models, and the gap vanishes in an update. An accumulation gap does not close, because you cannot buy eighteen months of operational memory; you can only have lived them. Time, here, is not a parameter you accelerate: it is the very material of the asset.
Two organizations, the same model at the start. At the finish, one has an asset, the other has a memory.
One can anticipate an objection, and it must be addressed because it always recurs: if model performance converges, why not wait for one of them to eventually integrate the upper layer itself, and get everything at once? The objection seems reasonable, but it runs into a fundamental difficulty. A layer owned by a model vendor ceases to be neutral: it is tied to a model, oriented by its interests, captive to its ecosystem. Yet an accumulation layer has value for the organization only if the organization owns and controls it. Entrusted to a vendor, it becomes a dependency again, that is, the opposite of what it was constituted for. To wait for the vendor to integrate it is to wait for him to take back with one hand what the trajectory had just placed in yours.
What accumulates does not get caught up
This is why durable value, in legal AI, will not be found in the model, however good, but in what the organization has managed to build and keep above. That performance lead rarely remains stable enough to become the foundation of a durable strategy; operational accumulation does. This asymmetry, invisible as long as you reason in capabilities, becomes the decisive factor as soon as you reason in duration. And it is in duration, not in instantaneous capabilities, that an organization’s position over several years is decided.
There is here a reversal of the question to ask at the moment of investing. The right question is not “what is the best model today,” because the answer will have changed in six months and will not, in any case, have decided much. The right question is “what does this investment let me accumulate, and to whom does it belong.” An investment that leaves nothing behind it, save matters produced and forgotten, has built nothing. An investment that feeds an accumulation layer the organization owns builds the asset that will matter when the models, themselves, no longer do.
That performance lead rarely remains stable enough to become the foundation of a durable strategy. Eighteen months of operational memory cannot be matched at all.
This is exactly the position MAX occupies: a layer above the models, not a product attached to one. The Legal Semantic Layer is designed to be the place where the organization accumulates what does not get retrained, its memory, its methods, its positions, independent of the model that, below, commoditizes. Because the only question that matters, in the end, is not which model you use, but what remains to you when that model is replaced.
