How AI is redrawing market structure

The pattern underneath the noise

Every few months a new model arrives, the leaderboard reshuffles, and someone declares a winner. Step back from the release cycle and a steadier pattern appears: the gap between the best model and the tenth-best model, on the tasks most enterprises actually run, keeps narrowing. Open-weight models trail frontier systems by a shrinking margin on document extraction, classification, summarization, and routine drafting — the workhorse tasks that make up the bulk of enterprise AI. Meanwhile the price of a unit of capability has fallen by orders of magnitude in a short number of years.

When a critical input commoditizes, value does not evaporate. It relocates. Understanding where it goes is the difference between a coherent AI strategy and an expensive collection of pilots.

Four layers, four very different games

It helps to think of the market as a stack, each layer with its own economics.

• Compute. Capital-intensive, supply-constrained, and concentrated among a handful of chip designers and hyperscalers. Extraordinary economics for the few who hold position, effectively closed to everyone else.
• Models. Enormous fixed costs, rapid depreciation, and fierce competition — including from open-weight releases that reset the floor. A brutal place to try to hold a margin, which is why nearly every serious model provider is also selling something else.
• Infrastructure and serving. Turning weights into fast, cheap, reliable tokens. Real engineering advantage exists here, but it is a volume business with compressing prices.
• Applications and workflow. Where a capability becomes an outcome inside a specific business process, with the integrations, controls, and change management that implies. Fragmented, unglamorous, and where most of the durable enterprise value is being created.

The uncomfortable observation for anyone building at layers two and three is that the two layers above and below them are absorbing the margin. The encouraging observation for enterprises is that the layer where your advantage lives is the one that is not commoditizing.

What actually functions as a moat now

The "thin wrapper" critique — that anything built on someone else's model is trivially replicable — was overstated, but it was pointing at something real. The question is what survives when the model underneath you gets better, cheaper, and available to your competitor on the same terms.

• Proprietary data that is actually usable. Not a data lake. Data joined to entities, labeled by outcomes, and legally usable for the purpose you want. Most organizations have the raw material and not the asset.
• Workflow depth. A system embedded in how work is done — with the permissions, exceptions, and edge cases handled — is expensive to replicate precisely because that work is tedious and specific.
• Distribution. Being where the work already happens beats being better in a place nobody visits.
• Trust and compliance posture. In regulated industries, the ability to demonstrate control, lineage, and human accountability is a genuine barrier. It takes years to build and can be lost in one incident.

The falling-price trap

Per-token prices have dropped steeply and will likely keep dropping. It is tempting to conclude that cost stops mattering. In practice the opposite happens: cheaper inference makes previously uneconomic use cases viable, volume expands to fill the new budget, and the bill goes up while the unit price goes down. Reasoning-style approaches that spend more compute per request push the same way.

The metric that survives this is not cost per token. It is cost per resolved outcome — per underwritten file, per triaged alert, per handled contact — measured against what that outcome cost before.

What this means if you are buying

Three consequences follow for an enterprise strategy.

Buy what is commoditizing, build what is differentiating. Renting model capability is rational; the price is falling and the switching cost, if you architect for it, is low. Your data foundation, your workflow logic, and your evaluation suite are the assets worth owning outright.

Architect for substitution. Assume the best model for your task changes within a year. A gateway layer, a portable evaluation suite, and prompts and adapters that are not welded to one vendor's API turn that churn from a migration project into a configuration change.

Concentrate rather than scatter. The organizations getting real returns are not the ones with forty pilots. They are the ones that picked a small number of workflows where the economics were legible, instrumented them properly, and pushed them into production with the governance intact.

The honest uncertainty

Two things could reshape this map. If capability gains reaccelerate sharply at the frontier and the gap to open weights widens again, the model layer regains pricing power and the calculus shifts back toward the largest providers. And if the current build-out of compute capacity outruns durable demand, the correction would be felt hardest at the infrastructure layer.

What seems robust across both scenarios is the enterprise conclusion: the value you capture comes from putting capability to work inside a process you understand better than anyone else. That has been true of every general-purpose technology, and there is little reason to expect this one to be different.