What Cannot Be Hedged
The sense in which I am using the term is that in which the prospect of a European war is uncertain, or the price of copper and the rate of interest twenty years hence, or the obsolescence of a new invention. About these matters there is no scientific basis on which to form any calculable probability whatever. We simply do not know.
The topology identifies where value accumulates. The question for any specific position is which layer it occupies and whether it can hold that position as the layers shift.
A discriminator emerges: can the position retain customers if a competitor offers materially better capability at the same price? If yes, the position has accumulated complementary assets—the switching costs, data advantages, and institutional relationships that the previous section identified as durable. If no, the position rents its revenue from the current capability frontier and faces erosion as that frontier advances.
Three categories follow. Frontier capability positions represent the highest-variance bets: a new architecture could create temporary advantage, but the window between breakthrough and diffusion compresses with each generation, and entering at frontier valuations risks exiting at commodity valuations. Interface wrapper positions take an API, add a user interface, and sell to a vertical. They face margin compression from both directions as the underlying provider competes downstream while cheaper alternatives emerge upstream. Full-stack positions combine cognitive capability with distribution, data, verification, and institutional relationships: enterprise contracts with switching costs, trained deployment teams, regulatory approvals, liability coverage. These are the complementary assets that capture returns as the core technology diffuses.
The model is not the moat. The infrastructure surrounding it is.
A common error treats the agent itself as the defensible asset. The error persists because agents exhibit remarkable capability, and capability improvements seem continuous with prior waves where capability conferred advantage. But persistence is externalized. An agent may appear persistent, but identity, memory, data rights, and permissioning all live outside the runtime invocation. The "same" agent instantiated by a competitor with equivalent models, prompts, and tool bindings is functionally identical. There is nothing internal to defend.
The moat, if it exists, lies in the surrounding infrastructure: the data feeding context windows, the workflow integrations determining when agents are invoked, the institutional relationships granting access to proprietary information, the user trust permitting agents to act with real consequences. A healthcare company deploying diagnostic agents has defensibility not in the agents themselves but in patient data informing their context, clinical workflows triggering their invocation, physician relationships interpreting their output, liability coverage permitting their deployment. A competitor with identical models faces years of infrastructure building. The agent is commodity. The permission stack is not.
The decisive question is which principal controls the context, integration, and permission infrastructure that capable agents require. Capability without context is a demonstration. Capability with context is a business.
When cognition becomes cheap, the bill shows up elsewhere: throughput, permission, verification, liability. Datacenter construction takes three to five years from site selection to power-on. A model architecture can be published, replicated, and improved upon in three to five months. The ratio between infrastructure time and capability time determines who sets terms. The cloud may be virtual. Its substrate is not.
The Joule Standard disciplines both layers differently. Cognitive investments face continuous margin pressure as the floor rises. Actuation investments price off scarcity, not marginal compute. But the Joule Standard disciplines indirectly: actuation assets are valuable because they enable cognitive deployments that clear the floor. If nothing clears the floor in a domain, the actuation assets in that domain have no customers.
There is one infrastructural choke point that repeats in every attempt at agent-mediated markets: a shared discount curve. Multi-period contracts require a benchmark rate for discounting future obligations. Absent a shared rate, N agents must negotiate bilateral credit curves: the O(N²) explosion that stalls market formation. A common benchmark collapses this to O(N). Some shared term structure has to occupy this position: fiat (SOFR-like), stablecoin-based, or Bitcoin-denominated. The thesis claim is that a Bitcoin-denominated curve, if it hardens, is unusually well-suited as a neutral yardstick when counterparties do not trust each other's institutions. The firms that establish and publish such a benchmark will capture structural advantage analogous to the creators of dominant reference rates in traditional finance.
The thesis rests on three assumptions: that cognitive capability commoditizes, that actuation constraints persist, and that the shift unfolds gradually enough to allow repositioning. Each can be wrong. Beyond these, the thesis can fail to generate returns even if the assumptions hold.
In this transition, the most important variable is duration. Models reprice in quarters. Power, fabs, logistics, and licensing reprice in decades. When you buy actuation, you are buying duration. Being right early is indistinguishable from being wrong.
The first failure mode: lock-in beats diffusion. A frontier lab releases a model that outperforms competitors on every benchmark. Within six months, three other labs match or exceed its performance. Prices collapse. The model layer commoditizes as predicted. But the customers do not switch. The switching costs are not in the model. They are in the evaluation data accumulated through deployment, the compliance certifications already obtained, the workflow integrations already built. Usage generates evaluation data; evaluation data improves reliability; reliability enables compliance deployment; compliance deployment generates more usage. The loop compounds faster than capability diffuses. If one or two frontier providers achieve escape velocity through these loops, the margin pool stays upstream. The bet on actuation underperforms. The thesis was correct about capability commoditization but wrong about where value accrues.
Enterprise software in the 1990s provides the cautionary example. By 1995, the core capabilities of ERP systems—accounting, inventory, procurement—were well understood and replicable. Dozens of vendors offered functionally equivalent products. The thesis of the era predicted that ERP would commoditize, that prices would fall toward marginal cost, and that value would migrate to complementary services: implementation, customization, training. The thesis was half right. Implementation services did capture substantial value—but so did the vendors. SAP and Oracle did not become commodity suppliers. They became entrenched incumbents commanding premium margins decades after their core technology could be replicated. The lock-in was not in the software's capability but in the data accumulated through years of operation, the custom configurations that no competitor could replicate without reimplementing them, the trained workforce whose skills were vendor-specific, and the switching cost that exceeded the benefit of any alternative. The capability commoditized. The installed base did not. Investors who bet on implementation services over vendor equity captured some of the value; investors who bet against vendor pricing power were wrong for a generation.
The second failure mode: actuation cheapens. Robotics could advance discontinuously. Regulators could converge on permissive standards. Liability frameworks could evolve to accommodate agent participation. If institutions adapt at the speed of software rather than the speed of law, actuation assets become bets on temporary constraints. A power generation asset takes three to five years to construct and operates for thirty to fifty. If the energy bottleneck persists for two decades, the investment compounds. If fusion economics arrive in five years, the asset enters a market that no longer resembles the one that justified the commitment. Capital is locked. The lesson arrives after the decision is irreversible.
The third failure mode differs in kind: the transition snaps. Wrong about commoditization—integrated platform companies outperform expectations—is recoverable for liquid capital. Wrong about bottleneck durability—actuation assets underperform—is also recoverable. Positions adjust. Lessons apply forward. But wrong about transition speed in the direction of acceleration means a breakthrough creates agents that bypass institutional constraints on identity, permissions, settlement, and liability. The assets that seemed durable turn out vulnerable. The positioning framework becomes obsolete overnight. This is not an error within the framework: it is failure of the framework itself.
A sufficiently large capability discontinuity invalidates positioning strategies built on gradual adjustment.
The key instrument is demo-to-deployment lag, measured as time from a credible demo to audited, insured deployment inside a regulated workflow at scale. Recent waves suggest that regulated deployment lags capability by quarters to low single-digit years, because audits, security review, procurement, and insurance move at institutional speed. If that lag compresses below six months, or if agentic deployment expands to high-stakes domains faster than institutional constraints would predict, the transition is accelerating beyond the speed at which capital can reposition.
Even if all three assumptions hold, the thesis can fail to generate returns. The Bitcoin yield curve may not emerge; the infrastructure may not be built; liquidity may not concentrate. If this happens, agent-mediated markets remain shallow and the positions that would capture benchmark-rate infrastructure capture nothing because the infrastructure does not exist. Or the thesis becomes consensus before playing out. Capital floods into the theme. Valuations rise to levels that discount decades of future value creation. Reality disappoints and valuations collapse. The thesis was correct. The trade lost money.
Carlota Perez's framework describes technological transitions in two phases: installation, when capital floods toward the new paradigm faster than deployment can absorb it, and deployment, when infrastructure matures and returns normalize.(Perez 2002)Carlota Perez, Technological Revolutions and Financial Capital: The Dynamics of Bubbles and Golden Ages (Cheltenham, UK: Edward Elgar, 2002).View in bibliography The transition typically involves a crisis—installation-phase valuations colliding with deployment-phase realities. The dot-com crash is the canonical example: the thesis was correct. The timing was catastrophic.
Current readings suggest early-to-mid installation: venture dominance, revenue multiple valuations, exploratory regulation, talent scarcity, transformation narrative. The deployment phase has not begun. Capital committed at current valuations must survive the installation-to-deployment transition, including whatever crisis marks the turning point.
Some parts of the thesis are anchored in physics: compute dissipates heat, infrastructure has lead times, and geology determines energy endowments. Some are anchored in incentives: a competing bid for energy exists as long as proof-of-work does. The O(N²) coordination problem collapses to O(N) with a common benchmark. The rest is institutional contingency: liability frameworks, regulatory adaptation, coordination timing. The portfolio should price these three tiers differently.
The risk surface is asymmetric. Most errors here are hedgeable: you can cut exposure, rotate, wait. One is not: a discontinuity that collapses the time between capability and deployment. The job is to know which risks you can trade, and which risks you can only survive.