Who Owns the Boundary?

Dynamic moats, model sovereignty, Open​AI’s pivot, compute ownership, and the afterlife of GPUs

The central economic question in artificial intelligence is not which model is best. It is which boundary captures the value produced as models improve.

A model can be extraordinarily capable and still become a replaceable supplier. An application can have many features and still be a thin wrapper. A physical computer can sit in someone’s home while ultimate authority over its cognition belongs to a vendor or state. A GPU can leave the frontier and remain productive for years. A user can possess hardware without retaining the right to continue operating it.

These are all boundary questions.

Who owns continuity? Which exchanges strengthen that ownership? What can be substituted without destroying accumulated state and relationships? Who may modify, transfer, inspect, or terminate the system? Does progress upstream accrete to the user’s computer or deepen dependence on the supplier?

Political economy begins where architecture meets control rights.

A moat must reproduce itself

The standard moat inventory—brand, patents, network effects, switching costs, scale—can become static taxonomy. Pat Dorsey’s discussion of competitive advantage becomes more useful when the moat is treated as a dynamic boundary.

A durable boundary captures value and information from exchanges, then uses what it captured to reproduce itself more strongly. In schematic form:

M(t+1) = M(t) + reinforcement(M, environment) − decay(M, environment)

A network effect qualifies when each participant makes the network more useful and therefore more difficult to leave. A workflow can qualify when each completed case deepens state, trust, data, integrations, and institutional habit. A brand can qualify when successful exchanges lower future customer-acquisition and trust costs.

Profit is not itself the moat. Profit is a maintenance resource. The moat is the relation that repeatedly captures value.

Free-energy and Markov-blanket language can illuminate this structure: a viable organization maintains a boundary, senses external conditions, updates internal state, and acts to preserve itself. But this is an analogy, not a theorem of corporate finance. Firms do not become organisms because a diagram labels their customer interface a sensory state.

The practical test is simpler:

• What relation does the company uniquely mediate?
• Which resources cross its boundary?
• Does each exchange deepen the same advantage?
• Do supplier improvements strengthen the boundary or dissolve it?

Model labs sell cognition; aggregators own judgment

As model capabilities converge and releases leapfrog one another, raw cognition becomes a contested supply market. Differentiation does not vanish, but the economic unit shifts.

The durable aggregator can own:

• the user relationship;
• durable state and artifacts;
• demand and task routing;
• permissions and authorization;
• comparative performance data;
• institutional integrations;
• the court that decides which output becomes consequential.

Model labs sell cognition. The aggregator owns judgment about when, where, and under what authority each cognition is used.

This is not a defense of cheap-model routing as a moat. Choosing the lowest-cost model for a prompt is easy to copy and vulnerable to suppliers bundling the same function. The boundary becomes durable when the system remembers cases, preserves user state, adjudicates candidates, and carries authorized outcomes forward.

The best model can serve as a lower bound: a candidate should not be degraded merely to save tokens. Revisions must earn promotion by identifying defects, producing stronger evidence, passing tests, or clearing a better court. Cheaper heterogeneous models add value as independent critics and sensors, not simply as discounted substitutes.

The subset/superset test

Vertical AI products face a clean test.

If the product’s reproducible expertise can be expressed as instructions, examples, evaluators, and tools that a frontier model can execute inside another environment, that expertise should become a portable skill. Soft editorial taste, a research method, a formatting doctrine, or a prompt sequence does not necessarily justify a standalone application.

An application earns its boundary when it owns hard workflow state:

• durable projects and permissions;
• execution environments;
• integrations and identity;
• branch and merge governance;
• exact artifacts and provenance;
• institutional distribution;
• continuing operations.

The compact formulation is severe but useful:

If Claude can do your product, you are a feature. If your product can hire Claude without losing effective quality, you may be a platform.

The second sentence has teeth. A product does not become a superset merely by offering more buttons around a weaker model. It must preserve effective frontier quality while placing the model inside a larger, durable system.

Marker and the supplier tax

The Marker conversation with John Steinbach provides a case study. Marker describes an end-to-end writing environment: research, composition, editorial control, search, revision, storage, and publishing. It also proposes an editorial theory about developing ideas through stages of desire, argument, and prose.

Much of that theory can be useful. But if Claude can reproduce it from a skill, the theory is a subset of the model environment rather than a moat.

Marker then confronts adverse model economics. Generation, long context, search, storage, retries, critique, and citation all cost money. First-party labs can price internally, cross-subsidize, smooth capacity, and bundle frontier quality into subscriptions. A vertical product buys cognition from its eventual competitor.

The danger becomes obvious when recognizable model style survives while interpretive control degrades. Lower-tier prose can retain the verbal costume of Claude while losing the judgment that made the product desirable. Users compare it directly with subsidized frontier assistants and wonder why the vertical surface costs more.

Marker is taxed by model suppliers unless its hard workflow state becomes indispensable. Its strongest editorial theory should travel as a skill; its application must justify itself through state, execution, collaboration, distribution, and accumulated case law.

Open​AI’s two constitutions

Open​AI’s strategic pivot makes the ownership struggle visible.

ChatGPT Work and Codex organize around professional execution: files, applications, browsers, enterprise systems, scheduled work, long-running tasks, and inspectable artifacts. The commercially attractive boundary is not casual conversation. It is authorization over institutional work.

The screenless companion device points toward another constitution: ambient sensing, personality, continuity across daily life, household habit, and an intimate relationship with the model. One boundary seeks measurable productivity; the other seeks attachment.

These strategies can converge technically—one model identity accompanying the user everywhere—but they make different claims on trust.

Enterprise work requires isolated jurisdictions, explicit permissions, audit trails, client boundaries, retention rules, and artifacts that can be inspected by people who do not share one personality. A household companion tends toward blended memory and frictionless presence. The same continuity that makes the device feel personal can make it constitutionally unsafe across employer, client, household, public, and private domains.

The likely early customer is not simply someone who wants synthetic friendship. High comfort with ambient AI overlaps with AGI believers, professionals, and prosumers willing to pay for leverage. Expected productivity can cause willingness to tolerate intrusion. Tolerance for intrusion is not proof of demand for companionship.

The device must repay its sensing and privacy cost through concrete agency. That pressure pulls it toward a productivity surface even if it wears a companion shell.

Advertising does not dissolve the bind. The chatbot advertising forecasts cited in the source discussion are too small, taken on their own terms, to finance the capital ambitions attributed to Open​AI. Advertising can supplement a business; it does not establish a sufficient consumer engine. The direct claim that Claude Code caused Open​AI’s enterprise pivot also exceeds the evidence. Product sequence supports influence and competitive pressure, not a proved single cause.

Model continuity versus user continuity

Open​AI’s products can be unified around one model sovereign: ChatGPT knows the user across work, home, mobile, and hardware. The provider’s identity persists; the user’s switching cost deepens.

A user-owned Automatic Computer chooses a different invariant. State, artifacts, permissions, authorization history, and institutional memory persist while models change.

This is the decisive substitution test.

If Claude is replaced by Codex, an open model, or a future system, does the user retain the same computer? If the provider changes its safety policy, pricing, political obligations, or product direction, can the user continue operating? Does personalization create a better provider-owned profile or a more capable portable institution owned by the user?

A model-sovereign computer projects one company’s cognition into many contexts. A user-sovereign computer employs many cognitions inside one continuing state.

Compute ownership is a bundle

The same distinction applies to hardware and weights.

Ownership should be audited as a bundle of rights:

possess, use, modify, exclude, transfer, derive, continue operating

A person may physically possess a machine while lacking several of these rights. Firmware can restrict workloads. Attestation can condition access to software or networks. Transfer can require permission. Open-weight training or distribution can be prohibited. A remote authority can retain a cryptographic shutdown capability.

These interventions should not be collapsed.

Industrial registration and inspection are not the same as universal household surveillance. Transfer regulation is not the same as workload verification. A prohibition on distributing weights is not the same as a power meter. Remote shutdown is not merely paperwork.

A regime can avoid a generalized panopticon and still create a licensed intelligence order. It can leave nominal private possession in place while relocating decisive authority over advanced computation to states or a bilateral consortium.

This is the limit of the narrow rebuttal in Scott Alexander’s “AI Chip Regulation Is Not A Dystopian Surveillance State”. He is right that controlling industrial chip production does not logically require watching every household. That defeats an imprecise surveillance objection. It does not answer the ownership objection.

The central question is who governs intelligence and whether ordinary people remain owners within that boundary or become licensed customers and subjects.

The kill switch occupies the active boundary

Remote shutdown authority changes the constitution of the machine.

The nominal owner may rack the hardware, pay for electricity, maintain it, and bear financial risk. But an external principal retains the right to terminate operation. Continued use becomes conditional.

The kill switch occupies the machine’s active boundary. It converts a property relation into custodianship under sovereign revocation.

Any such regime needs more than assurances of benevolent use. It needs jurisdiction, evidentiary standards, appeal, key custody, delegation rules, emergency expiration, compensation, and sunset provisions. Who can trigger shutdown? Can one state act unilaterally? Can keys be stolen or politically repurposed? Does an emergency power expire? Can the owner contest a false positive before the machine becomes worthless?

Safety policy often describes the intervention and leaves its court unstated. Constitutional analysis begins with the court.

GPUs do not have one economic life

Paul Kedrosky’s AI-infrastructure argument distinguishes AI’s usefulness from whether current capital expenditure can generate adequate cash flow. His strongest concern is duration mismatch: long-lived debt finances GPUs whose frontier economic position may decay rapidly while token prices fall.

That can break financing vehicles. It does not prove that the underlying GPUs become useless.

At least five clocks must be separated:

1. Physical service life: whether the hardware still operates.
2. Accounting depreciation: the schedule used to allocate cost.
3. Frontier competitiveness: whether the chip belongs in the largest training run or premium inference tier.
4. Merchant-inference profitability: whether selling tokens covers financing and operating cost.
5. Internal productive use: whether the machine creates more value for an owner than power, cooling, maintenance, and coordination consume.

A GPU can leave the frontier long before it leaves production.

Hardware can migrate through post-training, batch inference, specialized models, local agents, background institutional work, and lower-latency-insensitive tasks. A data-center owner can fail while a buyer at a lower asset price deploys the same boards profitably.

Software progress complicates ordinary depreciation. Better models, quantization, kernels, speculative decoding, caching, routing, retrieval, and harnesses can increase the useful cognition produced by fixed silicon. New hardware may improve faster, but relative obsolescence is not absolute productive decline.

The relevant unit is not token price alone. It is adjudicated useful outcomes per unit of energy, maintenance, and supervision.

Harnesses escaping containment

Kedrosky offers a useful falsification condition for the bubble thesis. Coding agents consume large amounts of inference because they operate inside rich feedback loops: inspect, modify, run, test, observe, retry. If such harnesses remain coding-specific, infrastructure demand may disappoint. If they escape coding into white-collar work, robotics, science, administration, and household operations, productive inference demand expands.

The Automatic Computer is a theory of those generalized harnesses. Branch, inspect, mutate, test, reconcile, promote, and roll back can apply wherever durable state can be transformed and adjudicated.

That does not guarantee every current data center earns its cost of capital. Civilizational inference demand and one financing structure’s solvency are different questions.

Regulatory vintages

Chip controls can create a second source of value for old hardware.

Before a surveillance, attestation, or remote-control regime, a GPU is valued mainly for cognition produced. Afterward, grandfathered unrestricted hardware may also be valued for privacy, exit, arbitrary-weight execution, transferability, and freedom from remote shutdown.

A regulatory-vintage premium can therefore offset technical obsolescence:

legacy value = compute + privacy + exit + uncensorability + optionality − power and repair cost

This is a hypothesis, not a guaranteed market.

If the premium becomes high enough, an aftermarket may treat pre-regime accelerators like Cuban cars: donor boards, component harvesting, specialist repair, firmware preservation, old-driver maintenance, and improvised server rebuilding. Standardized fleet cattle become individually tracked pets with serial histories, provenance, firmware lineage, repair records, and legal status.

The technical limits are real. Accelerator packages and high-bandwidth memory are difficult to repair. Package or HBM failures may be terminal. Power, cooling, interconnect, software abandonment, and detectability can overwhelm sovereignty value.

But regulation changes what is worth repairing. Scarcity premiums can finance expertise and preservation that ordinary depreciation schedules would never justify. Controls intended to concentrate governance at foundries may induce preservation and dispersion of legacy hardware, which can then motivate transfer, repair, possession, electricity, or operation controls downstream. That progression is not inevitable. It is a feedback loop regulators must evaluate.

GPUs have a sequence of jurisdictions, not one economic life.

The boundary that captures progress

The model lab wants each capability gain to deepen the continuity of its model across the user’s life. The platform wants each completed task to strengthen its workflow state and switching costs. The sovereign wants advanced compute to remain within a revocable legal and technical perimeter. The user wants better cognition without surrendering accumulated artifacts, privacy, exit, or the right to continue operating.

A moat is the boundary that repeatedly wins these exchanges.

This is why the Automatic Computer’s moat cannot be routing alone. It must be time-indexed constitutional state: the user’s branches, artifacts, permissions, adjudicated outcomes, and authority carried across changing suppliers and devices.

Choir does not need to predict which model wins. Its wager is that time spent improving any model can increase the value of the same user-owned computer.

The media version of that wager appears in “The Learning Economy”: Twitter’s graph is the moat, its algorithm is the regime, and delegated AI readers can separate the graph’s informational value from its attention interface. But abundance of readers makes privacy scarce.

The same question returns at every scale: which boundary owns continuity, and who can stop it?

This is the political-economy layer of The Constitutional Stack. The Automatic Computer defines the stateful substrate whose ownership is at issue; The Learning Economy applies the same analysis to media distribution, machine reading, and privacy scarcity.