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 matters more than the winners themselves. Names change; topology persists. Every techno-economic paradigm creates a characteristic pattern of value capture, and the pattern outlasts the firms that first exploit it.

Every techno-economic paradigm creates a characteristic pattern of value capture. David Teece formalized the observation.(Teece 1986)David J. Teece, "Profiting from Technological Innovation: Implications for Integration, Collaboration, Licensing and Public Policy," Research Policy 15, no. 6 (1986): 285–305.View in bibliography When a core technology diffuses faster than any single firm can appropriate it, returns migrate to owners of complementary assets: the capabilities, relationships, and institutional positions that the technology requires but cannot replicate. The core technology becomes infrastructure. The complementary assets become the bottleneck. The bottleneck captures the margin.

Consider historical precedent. The electrification of American industry between 1880 and 1930 exhibits the topology with unusual clarity.

In the beginning, generation was the frontier. Edison's Pearl Street Station opened in 1882, and for a decade the technical challenges of producing reliable current absorbed the best engineering talent and the most adventurous capital. Dynamo design improved rapidly. The shift from direct current to alternating current, the development of the steam turbine, the scaling of generator capacity—each advance diffused within years of its demonstration. By 1900, multiple manufacturers could produce generators to comparable specifications. Westinghouse, General Electric, and a dozen smaller firms competed on price and service. The sea of generation capability was forming.

The membrane emerged where generation met consumption. Industrial customers did not simply plug machines into outlets. Factories had to be rewired, motors had to replace line shafts, workflows had to be redesigned. The consulting engineer who understood both electrical systems and manufacturing processes occupied a strategic position: the person who could authorize the conversion, specify the equipment, and take responsibility when the installation failed. Samuel Insull, who built Commonwealth Edison into the dominant utility of the Midwest, understood that selling electricity required selling the entire transition—financing the customer's rewiring, training the customer's workers, bearing the risk of equipment failure. The authorization layer was not the kilowatt. It was the relationship that made the kilowatt usable.

The apex crystallized around transmission and regulatory territory. A generator could be built in five years. A transmission network took decades. The corridors through which power flowed—rights of way, substations, interconnection agreements—were physical assets that competitors could not replicate by hiring better engineers. When state regulation arrived—Wisconsin and New York in 1907, most states by 1914—it cemented these positions into legal monopolies: exclusive service territories in exchange for rate oversight. The utilities that had built transmission infrastructure first received the franchise. Latecomers found the territory occupied. By 1930, a handful of holding companies controlled most of American generation, not because they generated power more efficiently but because they controlled the wires that delivered it.

The pattern was complete. Generation commoditized; the sea of capability expanded; margins compressed toward the cost of coal and turbine maintenance. The membrane—the consulting engineers, the industrial sales relationships, the financing arrangements—captured returns during the installation phase but eventually commoditized as electrification became routine. The apex—transmission networks and regulatory franchises—proved durable. The positions established between 1900 and 1930 persisted for the better part of a century. Commonwealth Edison's descendants still operate the grid Insull built.

That is where the margin went last time. The topology suggests where it goes again.

Cognition commoditizes; actuation concentrates. Value flows to the interface between them.

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 month by month, accessible at diminishing cost to anyone with electricity and network connection. Models multiply. Training cost per unit of capability has fallen along curves that have held long enough for serious capital to act on them. Inference becomes infrastructure. Work that constituted professional differentiation a decade ago is now available through API. Work that differentiates today will follow the same trajectory within its turn. The sea contains thousands of foundation models, hundreds of inference providers, millions of fine-tuned variants. No single firm dominates; capability diffuses faster than any moat can form.

At the surface lies an authorization membrane: credentials, licenses, and legal standing that convert cognitive output into binding consequence. A model drafts the contract; a licensed attorney signs it. A model recommends the prescription; a credentialed physician authorizes it. A model generates the trade; a registered broker executes it. This membrane does not expand with computation. It expands through the slow processes by which professional schools admit candidates, licensing boards confer authority, and legislatures define regulated activity. The United States licenses roughly one million physicians, one million attorneys, half a million CPAs. These numbers grow at low single-digit percentages annually while cognitive capability doubles. Thin by design (accountability constrains supply), and in its ideal form that thinness is feature, not inefficiency: the institutional mechanism ensuring consequential actions remain traceable to parties capable of bearing responsibility.

Above lies the apex: infrastructure ownership. Generators and transmission systems, fabrication facilities, data centers, physical plant converting capital into productive capacity over timelines measured in years. TSMC and Samsung fabricate over 80% of leading-edge chips. Three hyperscalers control the majority of cloud compute. A handful of utilities own the transmission corridors that determine where power can flow. Durable, physical, constrained by factors software cannot circumvent. Their owners set terms for everyone operating below.

Value drains from the cognitive sea as capability commoditizes, margins compressing toward the cost of electricity and silicon. Meanwhile the membrane and the apex do not get cheaper on schedule. They get scarcer in practice. The gradient of returns follows the gradient of constraint.

Foundation models improve through scale, data, and architectural innovation. Each improvement diffuses. Many of the techniques that define the frontier diffuse quickly enough to compress the window between breakthrough and commodity, even when labs try to defend the lead. Open-weight models closed much of the gap within a year or two, and the lag continues to compress. The frontier advances, but the distance between frontier and commodity shrinks rather than expands.

This does not render cognitive capability worthless. It renders raw capability insufficient as moat. What persists is integration of capability with distribution, data, workflow, and institutional relationships. A model embedded in an enterprise workflow with proprietary data and switching costs has defensibility that raw capability lacks. The moat is not the model but what surrounds it.

Base inference (raw provision of tokens per dollar) may come to resemble utility economics: necessary, competitive, modest margins, advantage accruing to scale and cost structure. Layers above base inference retain software economics: integrated workflows accumulating switching costs, vertical applications with domain-specific data, feedback loops where usage improves the product. Different layers, different dynamics, different investment implications.

The complementary assets are actuation constraints: physical throughput, trusted interfaces, verification infrastructure, liability capacity. They are institutional, physical, and regulatory assets that larger models cannot create through additional training. They capture returns as core technology diffuses.

At the margin, where the same interconnect and power can serve either load, a kilowatt-hour routed to inference must clear more value than a kilowatt-hour routed to mining, or the flexible capacity migrates to mining. This is the Joule Standard: a floor beneath which cognitive deployments are uneconomic. Applications that cannot maintain spread above the floor become uneconomic regardless of technical capability. Continuous discipline, not one-time shakeout.

Actuation layer investments face different calculus. Physical throughput, regulatory licenses, and liability capacity do not compete directly with mining for electrical capacity. Their returns are governed by scarcity of the complementary asset, not by the hurdle rate on energy conversion. The spread between actuation margins and cognitive margins determines which vertical integrations are profitable. That spread widens as cognitive costs fall while actuation costs remain sticky.

Coasean logic predicts firm structure. When cognitive transaction costs fall but enforcement costs remain high, firm boundaries shift toward the enforcement bottleneck. If verification is cheap and standard, markets thicken and activities move outside firms. If verification is expensive or bespoke, firms internalize the bottleneck.

Healthcare illustrates the integration case. Cognitive systems can diagnose, recommend treatment, generate documentation. They cannot prescribe, bill, or bear malpractice liability. A vertically integrated health system (employing physicians, owning facilities, managing insurance relationships, bearing liability) can deploy cognitive capability at scale while competitors face integration friction at every interface. The standalone cognitive provider, however capable, faces the question: who signs the prescription?

The boundary of the firm traces the boundary of the permission stack. Wherever liability concentrates, wherever regulation requires licensed entities, wherever physical assets must be controlled, vertical integration provides structural advantage over pure-play cognitive providers. Capital that mistakes the cognitive layer for the value layer will find itself holding commoditizing assets in a market that has moved downstream.

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, the term structure never arrives and the settlement infrastructure positions capture nothing. The observable tests are not whether AI succeeds but whether its success commoditizes.

Cognitive capability is necessary infrastructure. It is not where the margin sits. The margin sits at the bottlenecks: power, plant, licenses, liability, trusted interfaces. If you can own those, you own the gradient. At least until the gradient moves.