Sea, Membrane, Apex

Ephemeralization: doing more and more with less and less until eventually you can do everything with nothing.

The structure that generates winners can matter longer than the first winners themselves. Names change while rights of way, licenses, distribution channels, and other hard-to-reproduce positions remain. A technological advance may spread quickly even as control of what makes it usable does not.

David Teece gave this problem an enduring economic form.¹ When an innovation diffuses faster than any single firm can appropriate it, returns may migrate toward complementary assets: capabilities, relationships, and institutional positions required to commercialize what the innovation makes possible. The core technology need not become worthless. It becomes less sufficient.

Electrification offers a useful precedent without supplying a law. Generating equipment improved and spread, but electricity did not become economically consequential through generation alone. Factories had to be reorganized around motors. Transmission required rights of way, substations, interconnection, and regulation. Knowledge of the central technology could diffuse while positions around delivery and use remained difficult to reproduce.

The lesson is not that every margin migrates downstream. It is that diffusion of a capability can increase the value of complements that make the capability usable.

When cognition commoditizes, actuation can concentrate. An owner controlling a necessary route into use can demand a payment, a commitment or a change in the customer's plans. How much the owner can demand depends on the customer's alternatives and the obligations under which that route is supplied. The position matters before we know whether buying it was a good investment.

Actuation is the scarce ability to turn a decision into real-world consequence: licensed authorization, insured execution, physical throughput, and settlement.

The topology is crude but useful: sea, membrane, apex.

Below lies a sea of cognitive capability, expanding while its composition and control remain fiercely contested. Models multiply, inference prices fall for many tasks, and capabilities once confined to specialized firms become available through common interfaces. Yet diffusion is uneven. Frontier training, distribution, proprietary data, and installed workflows can preserve market power even when raw capability spreads. The sea is not perfectly competitive; the question is which differences endure.

At the surface lies an authorization membrane: credentials, licenses, insured status, and legal standing that convert cognitive output into binding consequence. A model may draft a contract, recommend a prescription, or propose a trade while an authorized person or institution remains responsible for letting the act proceed. The membrane can widen, but it moves through training, law, insurance, and institutional recognition rather than through model improvement alone. Its scarcity may preserve accountability or protect an incumbent. Often it does both.

Above lies the apex: infrastructure ownership. Generators and transmission systems, fabrication facilities, data centers, and the physical plant that converts investment into productive capacity over years rather than software-release cycles. These assets are durable, geographically situated, and constrained by conditions that another training run cannot remove. Where they are concentrated, their owners can set terms for activity below.

As some cognitive capabilities commoditize, margins at those layers can compress toward operating and capital costs. The membrane and apex follow different clocks. Licenses, grids, fabrication capacity, trusted interfaces, and liability capital may remain scarce after a model capability has become commonplace. Their persistence gives customers a reason to secure access in advance. It can give owners bargaining power while leaving them with expensive obligations to supply what was promised. Customers may have no convenient alternative while their supplier struggles to meet its obligations.

Foundation models improve through scale, data, and architectural innovation. Some techniques diffuse quickly; other advantages persist through compute, proprietary data, distribution, or deployment experience. The investment question is not whether a frontier exists. It is how long a particular advantage survives and which complements remain scarce after competitors reproduce the visible capability.

A differentiated model can itself preserve an advantage. Once competitors can supply comparable capability, integration with distribution, data, workflow and institutional relationships becomes more important to the position of its provider. Some of that integration earns its keep by making a service better or less costly to provide. Some makes departure expensive. Both can sustain a business, but the customer's willingness to pay for useful coordination and inability to escape an incumbent are different reasons for the payment. Neither requires us to pretend that the model has become irrelevant.

Base inference may come to resemble a capital-intensive service in which scale and cost structure matter greatly. Integrated workflows can follow different economics because data, evaluation, distribution, and switching costs bind the customer to more than a model call. The distinction will be made by observed margins and switching behavior, not by assigning every company to a predetermined layer.

The complementary assets are actuation constraints: physical throughput, trusted interfaces, verification infrastructure, and liability capacity. Additional training does not create them directly. Whether they capture returns depends on scarcity, ownership, substitutability, and the value of the deployments they enable.

At a site where the same power, interconnection, hardware, capital, and regulatory permission can be redirected between loads, proof-of-work mining may provide one observable outside option. An inference deployment that cannot beat the operator's expected alternative after conversion costs will not hold that flexible capacity indefinitely. The Joule Standard names this conditional comparison. It is neither a universal energy floor nor a market price for cognition: many sites cannot switch, mining returns vary, and hardware, financing, connectivity, curtailment rules, and risk all shape the spread.

Actuation investments face a different calculus. Regulatory licenses and liability capacity do not compete directly with mining for electrical capacity. A customer may pay for the scarce service they supply, but that payment must also support the work and exposure undertaken in supplying it. If cognitive costs fall while actuation costs remain sticky, integration across that boundary may become attractive. If the gate widens or the capability remains differentiated, the result changes.

Coasean logic suggests how firm structure might respond. Cheap, standardized verification can make market exchange easier. Bespoke evaluation, enforcement, and liability can favor integration. The boundary will reflect several transaction costs at once, not a single verification variable.

Healthcare illustrates the integration case. A model may contribute to diagnosis, recommend treatment, or generate documentation, yet prescribing, billing, and malpractice liability remain assigned through institutions. A vertically integrated health system can combine cognitive capability with physicians, facilities, insurance relationships, and an answerable principal. A standalone provider, however capable, still faces the question: who may sign the prescription, and who answers when it is wrong?

The boundary of the firm may therefore move toward the permission stack. Where liability concentrates, regulation assigns authority, or physical assets must be controlled, integration can outperform a provider that sells cognition alone. Capital that mistakes visible capability for the entire productive arrangement is exposed if the capability diffuses while its complements do not.

This breaks in three ways. If frontier models maintain differentiation through data network effects, distribution lock-in, or regulatory capture, returns stay in the cognitive layer and the framework mispredicts. If robots get cheap while models stay expensive, the complementary assets bind less tightly. If agent-mediated markets remain shallow, investments in their settlement infrastructure may find too little demand to cover their costs. The observable tests are not whether AI succeeds but whether its success commoditizes.

Cognitive capability is becoming infrastructure without becoming economically irrelevant. Its diffusion can leave the next investment dependent on an asset or permission that another party controls. The owner of that position may be able to require a long commitment, reserve capacity for a preferred customer, or make an alternative costly to reach. Such powers help determine what gets built and who can use it. An investor buying the position also acquires its costs and obligations, at a price that may already anticipate the advantage. The power to set terms and the return on acquiring that power require separate accounts.