The claim, before the rhetoric

This coda does not claim that every software incumbent collapses, that application revenue disappears, or that physical assets earn nothing. It makes a narrower, conditional claim. Models make cognition reproducible, and increasingly make application functionality reproducible too. The rent that survives tends to migrate towards whatever accumulated state, default coordination, enforceable right, loss-bearing capital or constrained physical capacity remains binding and excludable, and it tends to migrate towards the complements whose effective substitutability falls most slowly. Sometimes the destination is atoms. Sometimes it is permission or a balance sheet. And sometimes no complement holds, competition passes the margin through or burns it in the fight, and the rent dissolves. The market does not owe the technology industry a replacement margin.

Here, rent means the return above the competitive level made possible by constrained substitution. It does not mean all revenue, accounting profit or payment received by an owner. A business can survive, grow and earn an ordinary return after the particular rent discussed here has disappeared. Accepted, throughout, means accepted by the party whose recognition makes the output actionable: the customer, the auditor, the regulator or the court, depending on the task.

The escape hatch is temporary

The obvious response to model commoditisation is to move up the stack. If the underlying intelligence becomes cheap, the frontier laboratory becomes an application company. It wraps the model in a workflow, adds customer data, builds an interface, secures distribution and charges for the finished product rather than the cognition underneath it.

The trilogy treated that move as an exit from the contested layer. This coda treats it as a stay of execution.

If a sufficiently capable model can write software, operate interfaces, call tools, inspect its own output, repair errors and maintain a working system, then the application is another bundle of cognitive tasks. The laboratory can move from model to application. The model can follow it there.

A business asks for a lead-management system. The model generates the schema, interface, permissions, reports, email automation, analytics and integrations. It tests the result, deploys it and repairs defects. The cost of producing the functionality falls. The application layer has acquired its own proprietary half-life.

That does not mean every company abandons its existing software next quarter. It means the functionality stops being the scarce part of the product. The invoice can survive long after the feature that originally justified it has become reproducible, because the vendor is still charging for everything required to make that feature usable inside a live institution: access to authoritative data, integration with existing systems, implementation, security approval, support, continuity, contractual responsibility and recourse when the output is wrong.

Feature rent dies before the invoice does.

The application is not destroyed in one event. It is hollowed out from the interface inward. The visible functionality becomes cheap first. The surrounding institution decays more slowly.

Generated mechanisms and accumulated state

The distinction that governs everything below is not software versus atoms. Code is a stock. Data can often be copied. A reputation graph can be exported. The useful line runs between generated mechanisms and accumulated, recognised state.

Models are increasingly able to generate mechanisms: code, interfaces, workflows, controls, tests, integrations, monitoring, transaction logic, compliance documentation, even the software of a marketplace. But an economic system contains facts that are not recreated by reproducing its mechanism. These customers already use it. These sellers are already verified. This transaction history actually occurred. This record is recognised as authoritative. This organisation is licensed to perform the activity. This contract is enforceable. This balance sheet stands behind the promise.

A model can generate a marketplace. It cannot, by generating the code, generate the fact that everyone is already trading there. It can create a customer database. It cannot unilaterally decide which copy the organisation, the auditor and the regulator will treat as the real one. It can draft an insurance policy. It cannot produce the capital that makes the claimant whole.

The model can reproduce the mechanism. It cannot instantly reproduce the state on which the mechanism operates.

That is where rent survives longer. Not permanently, and not by law of nature. Longer.

The system of record becomes a database with contracts

The visible application is a collection of forms, dashboards, workflows and reports. Those are precisely the elements most exposed to generation. A model can build a new interface over the same business data, customise it per employee and replace a rigid menu with a conversational agent.

The incumbent’s durable asset was never the screen. It was the organisation’s decision that a particular body of data would count. When two records disagree, which one is authoritative? Which system controls access? Which audit trail does the regulator accept? Which database triggers the invoice, the shipment, the payroll or the legal notice? A system of record is an institutional declaration about which state is recognised when records conflict.

AI compresses the moat around that declaration. It can map schemas, clean data, rebuild connectors, identify dependencies, generate tests and rewrite downstream workflows, turning a multi-year consultancy migration into a much shorter project. But copying the data does not complete the migration. Duplicated data creates two records, not a new authority. The replacement becomes authoritative only when permissions, contracts, reporting rules, audit procedures and organisational behaviour recognise it as such.

The model can clone the database. It cannot unilaterally declare the cutover complete.

The old application therefore becomes something thinner and stranger: a database with contracts, permissions and historical recognition attached. Its feature rent falls. Its authority rent survives on an institutional clock rather than a code-generation clock. Portability rules, open standards and cheaper migration keep shortening even that clock. They do not reset it to zero.

Marketplace code is not marketplace power

The same distinction applies to marketplaces, and here the decay is visibly layered.

A model can generate the software for a marketplace: listings, search, ranking, matching, payments, messaging and reviews. That does not generate marketplace power. A marketplace earns rent because multiple groups have coordinated around the same place. Buyers go there because sellers are there. The software supports the equilibrium; it is not the equilibrium. The scarce state is the verified participants, the transaction history, the accumulated reputation, the payment relationships, the dispute rules and the shared expectation about where demand will appear.

AI attacks that rent one layer at a time, from the outside in.

Discovery rent comes under pressure first. A buyer’s agent can search several platforms, compare offers and identify the cheapest acceptable supplier without accepting any single platform’s ranking. The first take rate under pressure is the charge for helping buyers find sellers.

Identity and reputation last longer. The agent still needs to know whether the seller exists and whether the history can be trusted, and platforms can keep those records proprietary. A model can analyse the evidence it can access. It cannot infer its way around evidence deliberately withheld. Portability standards and regulation can weaken this moat; better inference alone cannot.

Settlement and recourse tend to last longer still. Who processes the payment, absorbs the fraud, decides the dispute and compensates the buyer when the item never arrives? An agent can automate the work of all four. The enforceable promise still belongs to an organisation with contractual authority and assets behind it.

Discovery is usually the first marketplace rent exposed. Standing behind the transaction is usually among the last.

The counter-evidence should be named now rather than discovered later. Several large marketplaces have so far preserved or increased effective monetisation. eBay’s reported take rate rose from 13.77 per cent in 2024 to 13.94 per cent in 2025, with advertising and shipping contributing to revenue growth. Booking Holdings’ revenue rose from 14.3 to 14.5 per cent of gross bookings, partly through payment facilitation. The evidence is not uniform, and some headline seller fees have fallen under competitive pressure, but marketplace tolls have so far proved capable of migrating rather than simply disappearing. The discovery claim is therefore a dated prediction, not a description of visible data. The test is compositional rather than the total take rate. Compression should appear first as falling discovery-related revenue per agent-mediated transaction, while settlement, identity, fulfilment and dispute-resolution charges prove more durable. A rising total take rate would not by itself falsify the claim if the toll had visibly migrated towards those slower layers. The claim is weakened if discovery monetisation per comparable transaction remains stable or rises after buyer-side agents mediate a material share of demand and can genuinely compare suppliers across platforms.

The agent becomes the new gatekeeper

If a personal or enterprise agent can search every marketplace and route each request to the cheapest acceptable source, the marketplaces become backend liquidity pools. The customer does not care which supplier completed the task. The agent controls demand, and that creates a new concentration point: whoever owns the identity the agent uses, the permission to spend, the connection to corporate systems and the definition of acceptable quality can tax the layers underneath.

This is the cleanest explanation of the current platform land grab. The laboratories and platforms racing to make their assistants the default surface, and to surround them with app directories, permissions and certification, are not merely diversifying away from the model layer. They are racing to own the layer that disintermediates the coordinators. The observable evidence is already public: model companies now operate directories through which third-party apps are published inside their assistants, and the major platforms subject AI apps and agents to marketplace certification requirements. Discovery, distribution and governance around generated functionality are being enclosed while the functionality itself commoditises.

Whether the agent gate holds depends on which kind of state it accumulates. A user’s preferences, permissions and learned definition of acceptable are accumulated state, but they are individual state: exportable in principle, potentially reconstructible far faster than a two-sided network or certification regime, and an obvious target for portability regulation. The directory, the certification regime and the default placement are institutional state, and they sit on the slower clocks this coda describes. The agent gate is therefore two moats of different depths. Per-user memory is a shallow one. Distribution governance is not, and it may prove the most durable position the laboratories have yet acquired. Portability rules, multi-agent use and open protocols can still move the bottleneck past it. The coda cannot claim that they must.

If they do, the bottleneck moves again.

Trust performs four different economic functions

It is tempting to end the recursion with trust: the last moat is that people trust the incumbent. That is too vague to price. Trust contains four economic functions with four different half-lives.

Functional trust asks whether the system performs the task correctly. This is the most exposed form. Better models, evaluations, redundancy and formal constraints make functional confidence cheaper to produce. One model can test another. Functional trust is becoming a technical task.

Evidentiary trust asks whether the organisation can prove what happened: reproduce the decision, show which data was used, demonstrate that controls ran. Logging, monitoring and compliance documentation are highly automatable. But the production of evidence is automatable in a way its authority is not. A generated audit trail still depends on trusted provenance, recognised standards, tamper-resistant records and an institution willing to accept it. Evidentiary trust therefore sits across the boundary: the artefact is a mechanism; its admissibility is accumulated and recognised state.

Coordination trust asks whether the relevant participants recognise the same identities, histories, permissions and records. A newly generated reputation system has no reputation history. A copied identity network is not automatically recognised by banks, governments or trading partners. This is accumulated social state. It can be transferred or standardised. It cannot be produced by writing better code.

Financial and legal trust asks the residual question: if the system fails, who pays? Who owes the customer a duty, answers the regulator, funds the remediation, and possesses assets against which the claimant can recover? The surrounding work can all be automated. A model can perform the tests, draft the contract, prepare the audit and recommend the settlement. It cannot make the injured party whole unless a recognised entity has capital behind the promise.

The model can write the insurance policy. It cannot fund the claim.

Trust is not the last moat. The last institutional moat is credible recourse.

Recourse becomes the product

Organisations do not merely buy software. They buy someone to be responsible for it.

That sentence sounds like an objection to application commoditisation. It is the next stage of it. When the functionality becomes cheap, the commercial product shifts towards continuity, responsibility and loss absorption. A company pays a major vendor because the vendor promises the service will remain available, security will be maintained, regulatory commitments will be honoured, support will exist, and there is an entity to sue, fine or compel. The model can automate the tasks involved in meeting those commitments. It cannot eliminate the commitments themselves.

The application business therefore separates rather than vanishes. Some software companies survive as risk-bearing institutions with generated products inside them, their remaining rent resting on contracts, licences, recognised authority and a sufficiently large balance sheet. Others discover that the customer never needed the company once the functionality could be generated and the risk could be insured elsewhere. The split runs along liability rather than sentiment. Where losses are large, regulated or borne by third parties, the recourse premium holds. Across the long tail of low-stakes software, buyers will take the cheap option and carry the risk themselves, and nobody will be paid to stand behind it.

Functionality becomes cheap. Recourse becomes the product.

The clocks set the drift

Here is the operative part of the thesis, the part that makes it a prediction rather than a description.

Different complements become contestable on different clocks, and the resulting rents are repriced on a second set of clocks. Generatable functionality can become technically reproducible on model-release cycles, measured in months. Its commercial price may not reset until contracts renew, a credible substitute is deployable or migration becomes affordable. Integration and deployment become contestable on implementation cycles. Systems of record move on migration, procurement and audit cycles, measured in years. Networks and marketplaces move when enough participants can coordinate on an alternative. Legal obligations move through legislation, regulation and judicial decision. Loss-bearing capacity moves when insurers, capital markets or governments learn to price a new class of risk, often among the slowest clocks of all.

Capability half-life is therefore not identical to rent half-life. The first determines when a complement can be reproduced. The second determines when its owner can no longer charge as though it cannot. The technical moat can vanish while the supplier keeps collecting through contracts, inertia, bundling or the absence of a deployable substitute. That is what feature rent dying before the invoice does looks like from the inside.

The rents do not decay on the same clock. Within an exposed sector, value tends to migrate towards the complements whose effective substitutability falls most slowly. But the order is conditional. Regulation, vertical integration, liability, capital structure and sector-specific institutions can bundle layers, skip them or reverse the apparent sequence. A medical application can jump almost directly from generated functionality to regulatory permission and liability. A consumer design tool can jump from functionality to distribution. Insurance capital can reprice overnight while a statute stands unchanged for decades.

The prediction is therefore not that every industry follows one ladder. It is that, as faster-reproduced complements lose pricing power, a growing share of the remaining margin will be justified by the slower-reproduced complements that still bind in that particular market.

That migration tendency is the falsifiable core of the coda. It predicts where price justification moves inside any given business, and it explains why the transition looks contradictory from outside. A software incumbent can remain highly profitable after its technical moat has disappeared, because it is living on the slower half-life of installed state, contractual position or customer coordination. A marketplace can lose discovery while preserving its take rate through settlement. A frontier laboratory can lose model-layer scarcity while gaining default-agent distribution. Revenue persistence is not proof that the original moat remains. It may be the harvest of a slower-decaying one.

The same tendency explains why policy is perpetually late. By the time regulators respond to model concentration, value has moved into distribution. By the time they regulate application stores, it has moved into identity and payment. By the time they mandate data portability, it has moved into liability, licensing or sovereign permission.

The policy response is usually aimed at last year’s bottleneck.

The lineage of this argument is worth naming, because it strengthens it. Teece showed in 1986 that when imitation is easy, the profits from an innovation can flow to the owners of complementary assets rather than to the innovator, with the allocation depending on who can control and contract over those assets. The Bottleneck Thesis is that result applied to a technology that can increasingly generate the complements as well, run at a speed Teece’s cases never approached, with one added branch he did not need: sometimes every complement becomes contestable, and the rent has nowhere left to sit.

Atoms are not the automatic winner

The trilogy’s final movement towards physical capital was correct in one sense. No model runs without machinery, electricity, fabrication, materials and logistics. Digital abundance rests on physical systems.

But necessity does not guarantee rent. Electricity is indispensable, and most electricity producers hold no exceptional pricing power. Servers are necessary, and ordinary server manufacturing is not a monopoly. Land is physical, and most of it is economically irrelevant to high-density computation.

Necessity determines what must be purchased. Scarcity and exclusion determine who captures rent.

Physical owners capture exceptional returns only where the complement remains hard to expand, substitute or route around: leading-edge fabrication, advanced packaging, constrained grid connections, scarce low-cost power, permitted high-density sites, critical minerals, restricted routes. Even those scarcities decay. AI improves chip design, grid operation, materials discovery and construction. Governments subsidise capacity. Demand can disappoint. A previously scarce physical input can become an oversupplied commodity, and the bottleneck does not stop moving when it reaches atoms.

In many cases the enduring constraint is not the machinery but permission to use it: the grid connection, the export licence, the environmental permit, the right to serve a national market, the legal authority to process the data. Sovereign permission can outlast the scarcity of the physical asset it governs.

The endpoint is not physical capital. It is the least contestable claim still required to complete the outcome.

The sovereign gate

At the bottom of most private rights sits the state. It recognises property, enforces contracts, defines legal identity, licenses regulated activities, decides which records and signatures count, and provides the courts through which recourse becomes more than a promise.

A model can generate a contract; the state determines whether it is enforceable. A protocol can issue a token; the state determines whether the token grants a recognised claim over a house or a bank account. A marketplace can build a private reputation system; the state can still remove the licence under which its participants operate.

The sovereign can create, allocate or capture rent wherever permission itself is binding and enforceable. Whether the public receives that rent depends on how the right is issued, whether through auctions, taxes, royalties, equity or revenue-linked claims, and on whether jurisdictional competition forces the state to surrender the value to the private operator. Its position can decay more slowly because the sovereign often defines the legal clock, although constitutional limits, supranational rules, political legitimacy and competition among jurisdictions constrain that control.

Even sovereign rent is constrained. States compete for investment and tax base, and a mobile activity can relocate. The state captures durable rent only where it controls something firms cannot cheaply replace: access to a large market, a constrained grid, public land, legal recognition, procurement demand, rescue finance, physical security. The state is powerful where its gate is real.

This revises Part III’s advice without withdrawing it. The case for permits-for-equity never actually depended on atoms holding their scarcity value through the transition. It depends on permission being the slowest-decaying complement the state owns, and on the state designing the capture mechanism rather than assuming that ownership of the gate automatically puts the rent into the treasury. The instruction stands, with sharper wording: the state should convert its permission into durable, contingent claims while the permission remains binding, and it should take those claims in the operating entity and the revenue stream rather than in the depreciating machinery, precisely because the coda cannot guarantee which physical assets stay scarce. Trade the gate for equity, royalties or revenue-linked participation. Do not trade it solely for an interest in machinery the recursion may commoditise.

Sometimes nobody captures the rent

The debate about AI rents usually assumes the surplus must have an owner. If the laboratory loses the margin, the application vendor gains it. If the application commoditises, the marketplace gains it. If agents weaken marketplaces, the platform gains it. If the cloud commoditises, the power company gains it.

That reasoning treats rent as a conserved quantity. It is not.

Suppose models are competitive, applications are easily generated, data is portable, marketplaces interoperate, identity is standardised, risk can be insured competitively and infrastructure is abundant. No layer holds a binding, excludable bottleneck. The previous margin is not inherited by a hidden final landlord. Some of the former producer rent is passed through as lower prices, higher quality or greater output. Some is dissipated through duplicated investment, strategic subsidy and capital write-offs. The essential claim is not that customers inherit every euro of the previous margin. It is that no successor producer is economically guaranteed to inherit it.

The market does not appoint a new rentier merely because the old one disappeared.

This is the most economically attractive outcome. It is also the politically dangerous one.

The fiscal paradox of consumer surplus

Consumer surplus is a real gain. If a service that cost €1,000 can be produced for €10, the customer is better off, and real living standards can rise even as nominal expenditure falls. Standard GDP is not a measure of total economic welfare and does not record consumer surplus directly; Brynjolfsson and colleagues built GDP-B to capture welfare contributions from new and free digital goods that conventional national accounts can miss. The fiscal implication drawn here is mine, not theirs: welfare that does not appear as a wage, profit or taxable transaction is not automatically converted into state revenue.

But consumer surplus is not a payroll. It is not a corporate profit, a dividend or a rent payment. The state cannot send a tax demand to the invisible difference between what a consumer would have paid and what the service now costs.

Some of the gain returns to the tax base indirectly, as released income is spent elsewhere. That transmission is not automatic. If lower prices arrive alongside displaced wages, compressed producer margins, reduced billable hours and falling nominal business expenditure, then real abundance can coexist with fiscal weakness. Society can afford more in material terms while the institutions responsible for funding the transition collect less in nominal claims.

The productivity gain exists. The fiscal claim on it does not arise automatically.

This gives the trilogy’s fiscal crisis a second, harder-to-tax form. In the first, rent migrates to concentrated physical, platform or sovereign bottlenecks, and the political fight is over who owns and taxes them. In the second, producer rent is competed away faster than the wage and tax systems can be rebuilt, and there is no equivalent corporate winner from whom to recover the cost of displaced wages and institutional adaptation. The output becomes cheap. The transition remains expensive.

There is a third form between them, and it may be the most common. Where several weak complements bind at once, the surviving rent fragments into thin, contested margins spread across many holders, none dominant. That can broaden the ordinary tax base and reduce the political power of any single rentier, but it also leaves less concentrated surplus for the state to capture through targeted bargaining. Whether fragmentation is fiscally better or worse depends on mobility, reporting, profit allocation and the tax instruments available. The treasury no longer has one obvious counterparty. It is negotiating with a diaspora.

Cheap does not mean equitably distributed either. A generated service can be nearly free at the point of production while remaining inaccessible to anyone lacking a device, connectivity, recognised identity, payment credentials or the ability to bear residual risk. The productive mechanism becomes abundant; the rights required to use it remain scarce. The model can generate a legal argument, but only a recognised court can enforce it. It can design a treatment plan, but only a licensed institution may deliver it. The scarcity shifts from production to participation, and the important question stops being who owns the model.

Who is permitted to turn the model’s output into an enforceable claim on the world?

That owner may capture more rent than the model provider ever did.

The Bottleneck Thesis

The theory can now be stated plainly.

An accepted economic outcome requires a collection of complements: cognition, software, data, integration, distribution, identity, permission, settlement, recourse and physical capacity. Most can be supplied competitively. Actors controlling complements with low effective substitutability can charge above competitive cost, and that premium is the rent. Where several such complements are required at once, their owners bargain over the resulting surplus rather than one layer automatically inheriting it all.

Models reduce the scarcity of cognition. Code generation reduces the scarcity of application functionality. Agents reduce discovery costs. Interoperability reduces integration lock-in. Portable identity weakens network capture. Competitive insurance weakens liability rent. Infrastructure build-out weakens physical scarcity. Regulation can weaken or create any of these positions.

At each stage, one of four things happens. The previous owner retains the bottleneck. The bottleneck moves to another layer. The rent fragments among several owners who each hold part of the gate. Or no binding bottleneck remains and the rent is competed away or dissipated.

Not all difficulty earns rent; the asset must be binding. Not all scarcity earns private rent; the asset must be excludable. Not all control is durable; the customer must be unable to route around it.

Economic rent forms around binding, excludable complements that cannot be cheaply reproduced or routed around. Where several complements bind, the rent is divided according to control and bargaining power. As models make cognition, software and adjacent mechanisms more reproducible, surviving rent tends to migrate towards the accumulated state, coordination, enforceable rights, loss-bearing capital, physical capacity or sovereign permission that remains less substitutable. Competition can also fragment or destroy the rent without appointing a successor owner.

What would falsify the coda

This is a taxonomy with a migration tendency, so most tests are directional rather than dated. Freeze the predicates before reading the data.

Feature-level differentiation should become harder to monetise as generation quality rises, with vendors increasingly justifying price through integration, authoritative data, compliance and recourse rather than unique functionality. Systems of record should retain pricing power longer than feature vendors and lose it as migration and institutional cutover become cheaper. Marketplace discovery fees should compress before settlement, identity and dispute-resolution fees, and total take rates should fall towards the layer still bearing loss as reputation and payment state become portable. Model providers that fail to convert capability into distribution, state, permission or risk-bearing should compress faster than platforms holding those complements. Physical infrastructure should earn exceptional rent only while capacity or permission remains constrained, reverting to utility economics where supply expands and access is contestable. And where no layer retains a defensible bottleneck, prices should fall without an equivalent producer profit appearing elsewhere.

One near-term observable deserves its own line. Watch not merely whether AI-native insurance products emerge that price generated-software risk, but whether underwriting becomes standardised, coverage broadens, exclusions narrow and risk-adjusted premiums fall. If insurers can repeatedly price that risk without bespoke investigation or exceptional capital charges, the recourse clock, treated above as among the slowest, is visibly accelerating, and the taxonomy is being tested at its most load-bearing joint.

To prevent the framework becoming a retrospective naming exercise, each sectoral test must identify the candidate complements before observing the resulting margins. For each complement, specify the exclusion mechanism, switching cost, portability, time to substitute and predicted source of pricing power. A previously unlisted hidden bottleneck cannot be introduced after the result without independent evidence that it constrained substitution during the test period.

The empirical claim is weakened if durable excess margins remain after the identified exclusion rights, switching costs and portability barriers have demonstrably fallen. The ordering claim is weakened if, across comparable markets and absent a separate regulatory or contractual shock, readily generated and portable layers retain pricing power more persistently than the less portable state, recourse or permission layers predicted to replace them.

The foundational statement, that constrained substitution permits rent, is a definition-level economic proposition. The falsifiable contribution of this coda is the proposed location and relative decay of those constraints as models climb the stack.

The corrected ending

The trilogy’s conclusion was that value migrates from cognitive capital to physical capital.

The complete conclusion is this:

Rent follows the slowest-decaying complement still binding, splits when several bind at once, and vanishes when none do.

Stated in full: economic rent forms around binding, excludable complements that cannot be cheaply reproduced or routed around. Where several complements bind, the rent is divided according to control and bargaining power. As models make cognition, software and adjacent mechanisms more reproducible, surviving rent tends to migrate towards the accumulated state, coordination, enforceable rights, loss-bearing capital, physical capacity or sovereign permission that remains less substitutable. Competition can also fragment or destroy the rent without appointing a successor owner.

Physical capital is one possible destination. It is not the economic terminus.

The model can generate the application. It can generate the marketplace software, much of the compliance machinery, the integration and the maintenance patch. What it cannot instantly generate is the fact that everyone is already there. It cannot generate the history on which a reputation rests, the shared decision that one record counts, the legal right to perform a regulated act, or the balance sheet that pays when the system fails.

That is where rent survives. Not permanently. Not necessarily in atoms. In the last binding claim the model cannot reproduce or route around.

When that claim becomes contestable, the bottleneck moves again.

When no bottleneck remains, the market does not appoint a new rentier.

The rent disappears.

Sources

1. David J. Teece, “Profiting from technological innovation: Implications for integration, collaboration, licensing and public policy”, Research Policy 15(6), 1986. The complementary-assets result: when imitation is easy, profits from innovation can flow to the owners of complementary assets rather than to the innovator, with the allocation depending on the appropriability regime and control of those assets.
2. Erik Brynjolfsson, Avinash Collis, W. Erwin Diewert, Felix Eggers and Kevin J. Fox, “GDP-B: Accounting for the Value of New and Free Goods in the Digital Economy”, NBER Working Paper 25695. The measurement framework for welfare gains that fall outside measured GDP. The fiscal-visibility inference drawn in this coda is the author’s extension, not a claim of the paper.
3. OpenAI, “Developers can now submit apps to ChatGPT”. Evidence of model companies operating discovery and distribution directories around generated functionality.
4. Microsoft, commercial marketplace certification policies. Marketplace certification requirements applied to AI apps and agents; evidence of platform governance enclosing the layer around generated functionality.
5. eBay Inc., 2025 Annual Report (Form 10-K) and Q4 2025 results. Reported marketplace take rate, defined as net revenues divided by GMV, of 13.94 per cent in 2025 versus 13.77 per cent in 2024, with first-party advertising penetration and shipping programmes contributing to revenue growth.
6. Booking Holdings Inc., 2025 results. Revenue equal to approximately 14.5 per cent of gross bookings in 2025 versus 14.3 per cent in 2024, with increased payment-facilitation revenue contributing to the rise.
7. Ben Luong, “The Second DeepSeek Moment”, Part I of the trilogy. The contestability of the model layer.
8. Ben Luong, “The frontier labs are building a product Hetzner will sell like bandwidth”, Part II of the trilogy. Accepted-output economics and the two routes to commodity-priced cognition.
9. Ben Luong, “The Margin Call”, Part III of the trilogy. Scarcity-priced obligations, the permits-for-equity window revised above, and the fiscal floorboards extended here.