The Margin Call

The claim, stated narrowly

Open weights do not make compute free, and contested model pricing does not automatically crash GPU finance. The claim of this essay is narrower. The quality-adjusted selling price of adequate cognition may fall faster than parts of the surrounding capital stack can reprice, refinance and amortise. If effective supply outruns paid demand, the loss need not end the automation cycle. Where the discount is passed through competitively, it can finance the next phase of it.

Part I established the fact. Near-frontier capability has entered the open-weight layer and given buyers a credible outside option. Part II established the mechanism. Standard interfaces, delivery optimisation and routing lower the cost of comparing providers and push the delivered price of intelligence towards contested infrastructure economics, whether delivery is spread across many hosts or concentrated inside one integrated laboratory.

Providers are still not identical. Production deployments still need prompts, tool schemas, safety behaviour, latency, evaluation and compliance revalidated. Compute, power and reliable serving are still scarce. The claim is not that friction has gone. It is that exclusive control over frontier-like capability can no longer be assumed.

This essay asks the question the first two did not. Who lent money against the assumption that none of this would happen?

The build-out was priced before the premium became contestable

AI capital expenditure now runs into the high hundreds of billions of dollars. A Reuters dashboard reports more than $800 billion of estimated hyperscaler capital expenditure for 2026. Reuters’ analysis of 29 July reports that by 7 July, Amazon, Alphabet, Meta and Oracle had issued about $194 billion of bonds, 79 percent more than in the whole of 2025. It also records widening borrowing spreads, and cites a Goldman Sachs estimate that debt will fund about a third of hyperscaler capital spending in 2026.

That shows material exposure. It does not show a crash.

There is no single capital structure here, and precision matters. Cash-rich hyperscalers have paid for much of the build from operating cash flow and can absorb disappointing returns far longer than anyone downstream, although the size of their current commitments is pulling them into debt markets too. Utilities and data-centre landlords earn contracted returns. Nvidia can stay the biggest winner of the build-out, because open weights increase demand for accelerators. That can reverse if efficiency gains, overcapacity, slower frontier training and a secondary market in impaired chips weaken demand for new high-margin hardware. Nvidia can win the boom and still be repriced by the overbuild it supplied.

The vulnerable layer is at the leveraged edge: neoclouds, GPU lessors and project-finance vehicles whose debt was sized to particular rental rates, particular utilisation assumptions and particular collateral values, for hardware whose economic life is measured in years rather than decades. Those numbers were set when frontier-like intelligence looked scarce for long enough to repay the capital built around it.

Not every leveraged exposure is merchant exposure. In March 2026, CoreWeave closed an $8.5 billion non-recourse delayed-draw facility, secured on high-performance computing infrastructure and a related customer contract, maturing in March 2032. The company said its equity and debt financing commitments over the previous twelve months came to about $28 billion. Its first-quarter filing ties borrowing to depreciable equipment cost, projected debt-service coverage and project-level conditions. Contract backing may protect facilities like this from short-term falls in spot prices. The exposed group is narrower: structures where tenant credit, take-or-pay durability, utilisation, collateral value or refinancing assumptions fail together.

Nor is this a claim that frontier models fail to recover their training costs. Fully allocated research, serving and organisational economics for individual models are not public, and a laboratory can stay profitable through a succession of short leads. The hypothesis is narrower:

The quality-adjusted selling price of adequate cognition is falling faster than parts of the surrounding capital stack can reprice, refinance and amortise.

For buyers able to evaluate and deploy alternatives, procurement now happens in the shadow of downloadable weights that sat within three Intelligence Index points of the closed leader at K3’s launch, and that competing hosts with the infrastructure and licence rights can serve. The outside option constrains the premium whether or not anyone uses it. Contested model pricing can then flow downstream into the rental rates, utilisation and collateral values the leveraged edge borrowed against.

That transmission is not automatic, and the condition should be named. Open weights can destroy model-layer rent while increasing demand for GPU hours, because more hosts, deployments and inference can absorb more compute. A credit event requires effective compute supply, which is installed capacity multiplied by hardware and algorithmic efficiency, to outrun paid workload demand. Or it requires competition among hosts to push realised rental economics below debt service.

That is the prediction this essay adds to Parts I and II. If demand keeps merchant rental rates and utilisation at their underwriting levels, the model rent can die without the leveraged edge breaking. The efficiency curve, the overbuild and the routing layer make the failure plausible. They do not show it has happened. This essay dates the prediction rather than presenting it as a result.

The arithmetic is now visible: debt sized to scarcity pricing, serviced by contested pricing. The gap between those two numbers is where the bubble lives. Kimi did not create the gap. The release of its full weights on 27 July made it visible enough to date, not automatic enough to declare.

Volume is not solvency

The standard defence is demand. As machine cognition gets cheaper, firms will use more of it. But the variable that matters is not token volume. It is paid, accepted output. Better systems can complete more useful work with fewer tokens, as OpenAI says of GPT-5.6 on tool-heavy tasks, while weaker household demand can reduce the total output the economy will pay for.

Cheaper cognition expands the uses it is put to. That does not guarantee that paid demand outruns simultaneous gains in hardware efficiency, algorithmic efficiency and installed capacity, or that the rental rate, utilisation and collateral value assumed by the capital stack survive. Rising demand does not guarantee scarcity margins. Electricity, bandwidth and freight carry enormous volumes at infrastructure returns.

The distinction is where the margin call actually lives. Volume can service debt at utility margins. Grids, pipelines and telecoms are financed on exactly that basis. What volume at contested margins cannot do is preserve software valuations, venture returns, residual values for silicon that ages faster than its loans amortise, or refinancing terms written when capacity looked permanently scarce.

So the first loss lands on equity, wherever valuations priced reproducible cognition as a durable monopoly. It becomes a credit event wherever fixed obligations were sized to the old rental rate, the old utilisation or the old collateral value. That means merchant GPU capacity, leveraged neoclouds, project vehicles that depend on loss-making tenants, and hardware-backed lending whose collateral ages faster than the loan. The hyperscalers can absorb disappointing returns for much longer. The leveraged intermediaries may not survive refinancing.

The frontier laboratories can see this, which is why they are racing to become product companies, with agents, coding environments, enterprise platforms and devices. They are trying to own the system around the intelligence before the intelligence becomes cheap. Some will succeed, as product companies or as integrated providers able to undercut independent hosts while keeping value elsewhere in the stack. That weakens the case for a dispersed hosting market. It does not restore scarcity pricing to routine cognition.

Demand can fall while substitution accelerates

Cognition is an intermediate input, so demand for it depends in the end on final demand. If lost wages weaken mass consumption, paid workloads may grow more slowly than the demand defence assumes, or shrink. That strengthens the credit risk. It does not necessarily help labour.

Three quantities have to be kept apart: total final output, the absolute machine workload, and the machine share of production. Let Y be final output, c the cognitive input needed per unit of output, and s the machine share of that input.

Machine-handled cognition: M = Y × c × s

Human-handled cognition: H = Y × c × (1 − s)

Suppose final output falls from 100 to 90 and cognitive intensity stays at one. If the machine share rises from 20 percent to 60 percent, machine-handled cognition rises from 20 to 54, while human-handled cognition falls from 80 to 36. Output has fallen by 10 percent. Machine work has risen by 170 percent. Human work has fallen by 55 percent.

Even where machine work does not rise in absolute terms, human work can fall much faster than output. A firm serving a smaller market can serve it with one supervisor and a machine system instead of five employees. The displacement variable is not total compute consumed. It is the machine share of the cognition used to produce each remaining accepted output.

That creates a loop. Substitution weakens wage income. Weaker wages weaken mass demand. Weaker demand increases the pressure to cut unit cost. Cheaper machines make more substitution available. One contraction can therefore break infrastructure debt through too little absolute workload, while breaking labour through a rising machine share. The owner of compute needs enough absolute revenue to service fixed obligations. The worker needs human labour to keep enough of the relative share. Those are different requirements.

Earlier automation waves offer a precedent, not a forecast. Jaimovich and Siu found that losses of routine employment in the United States were concentrated in downturns, and that jobless recoveries were largely accounted for by routine occupations that disappeared. Robert Allen’s account of Engels’ Pause describes a long period in which productivity and output rose far faster than workers’ living standards before the gains were shared. Neither proves this transition will follow the same path. Both show that restructuring can concentrate in contractions, and that eventual recovery is no answer for a generation living through the delay.

The counterforce is real. Weak firms may lack the cash, the organisational capacity or the confidence to adopt new systems. Consolidation may keep lower infrastructure costs from reaching them. The acceleration claim fails if those frictions suppress adoption more than falling machine costs and cost pressure encourage it.

During the transition, final demand can still come from the wages that remain, from savings and credit, from government deficits and transfers, from capital income, investment, exports, and lower prices that stretch residual income further. None of these is automatically a permanent replacement for wage-based mass demand. Without redistribution or a new claim on automated output, the process can end in underused capacity, repeated demand shortfalls and concentrated production rather than general abundance. The absence of a stable replacement for wage demand is not a premise needed to start the transition. It is part of the successor-system problem the series leaves open.

A crash can speed up the displacement

This is the part the public conversation has backwards. The assumed sequence runs: the bubble bursts, the AI story dies, the pressure on jobs eases. Workers watching the capital expenditure figures with dread are quietly hoping for the crash, on the theory that a crashed industry stops buying machines.

The labour sequence may already have begun, quietly. Junior hiring can thin, graduate absorption can weaken and incumbent teams can be asked to do more, without a clean break in headline unemployment. The No-Scream Principle, that compositional damage appears before aggregate measures register it, predicts exactly this.

Credit behaves differently. It has payment dates. The prediction is not that no worker is harmed before a lender is. It is that financing stress at the leveraged edge becomes the first hard, dated, institutionally undeniable break, before AI displacement shows in headline unemployment.

In that sequence, the leveraged edge breaks. Merchant clusters are refinanced, sold or absorbed at steep discounts. Durable sites may hold their value. The acquisition cost of the compute inside them does not. Inference can then clear below the return originally needed to build the capacity, because someone else has already taken the capital loss.

One counterforce deserves real weight, because the acceleration claim depends on beating it. The most plausible buyers of distressed capacity may be hyperscalers that already own adjacent infrastructure. If a few incumbents absorb the written-down assets and hold the prevailing price, the lower acquisition cost becomes landlord surplus rather than cheaper intelligence.

Historic cost is sunk. A host with cheap assets charges the prevailing rate unless idle capacity and competitors’ price cuts force it lower. Standard interfaces and routing make price differences visible and switching easier, but they do not create competitors. The discount reaches customers only if capacity stays contestable after restructuring, through independent hosts, forced disposals, open access or some other mechanism that keeps survivors competing to fill their assets.

So there are two separate claims. A credit event can happen without pass-through. The acceleration claim needs both the write-down and competitive pass-through. If consolidation removes the second, the crash does not subsidise diffusion, even though lenders still take the loss.

If competition survives, the consultancies, insurers, law firms, software companies and government departments that found frontier API pricing hard to justify face a different calculation. The marginal adopter, the firm that ran a pilot and shelved it on cost, reruns the numbers and may adopt. Substitution can accelerate into the downturn rather than out of it, because downturns are exactly when firms cut headcount and hunt for cheaper inputs.

The same pressure reaches individuals, and it intensifies something already running. Freelancers already underbid agencies at today’s prices. Agencies underbid in-house teams. In-house seniors run departments without juniors. Juniors use the same tools to stay employable a little longer. Consultants sell transformation plans that automate the clients who bought them. Each actor adopts to survive the round, and widespread adoption removes the reason many of them were needed. That is the Multiplayer Prisoner’s Dilemma at the level of individual careers, and it does not wait for a crash. What a pass-through crash changes is the floor. Every move gets cheaper.

The structure has precedents. Railway investors were ruined in the 1840s, and the railways remained, their fire-sale capacity industrialising freight. Telecom investors were ruined in 2001, and the fibre remained, carrying the internet economy for twenty years at prices its original investors never planned for.

One honest difference matters. Rail and fibre were durable assets. This generation of accelerators is not. What survives an AI crash is the durable shell around the silicon: grid connections, substations, buildings, cooling, power contracts and permitted sites. The shell is why the capacity survives. It is not necessarily the discounted asset. The subsidy comes from the excess compute, the impaired chips and the capital claims written off above them, while the permitted shell takes years off the next owner’s deployment timetable.

A crash also pulls development and diffusion in opposite directions. Capital destruction may thin out the largest training runs and slow the frontier. It can speed the spread of capability already created, because the systems already shipped are enough for a large class of commercial substitution, and their delivery cost has fallen. The objection that bankrupt AI firms cannot advance AI is true and beside the point. The displacement does not need the next model. It needs the last one, cheaper.

A crash meeting these conditions may be reported as the end of the AI era while working as a subsidy to the replacement of labour, wherever the discounts reach employers able to use them.

The state’s four tax bases move faster than its tax code

The credit event and the fiscal squeeze run on different clocks. The first is a refinancing prediction for the late 2020s. The second compounds into the 2030s as compositional labour losses accumulate. Payroll and consumption receipts can weaken at the margin before headline unemployment breaks, but a state-wide cash-flow crisis is not the same-quarter consequence of the margin call.

The fiscal exposure may still be larger than the financial one, and almost nobody has added it up.

The state taxes wages. AI compresses wages, first through quiet non-absorption at entry level, then through restructuring in mid-career. One of the largest and most automatically collected tax bases in every developed economy erodes at source.

The state taxes corporate profits. The rents that survive do not vanish. They migrate out of the contested model layer into clouds, chips, energy, land and distribution, where they gather in fewer hands that are more internationally mobile and better placed to bargain with any single government. A broad, captive base becomes narrow and negotiable.

The state taxes consumption. Displaced households consume less, and consumption tax receipts follow wages down.

The state borrows against future growth. Here cheap AI may fulfil the GDP forecast while breaking the fiscal one. Output can rise while wage income, payroll receipts and mass consumption fall, if the gains go to a small, mobile capital base. Productivity is not tax capacity.

These are four gross pressures, not an accounting identity. Cheap AI could also raise taxable profits, lower prices, expand consumption and reduce some public costs. The claim is about timing and incidence. Payroll losses arrive automatically and domestically. Replacement rents are narrower, more mobile and slower to capture. Demands for transfers arrive before legislation, treaties and enforcement can redirect the surplus. The state does not lose because output disappears. It loses because the tax base moves faster than the tax code, and it must fund the transition before it has captured any of the gains.

The surplus may be mobile, but the scarce physical complements are fixed in place. That is why taxation will move towards what cannot relocate: land, generation, grid connections, water rights, data-centre footprint, extraction and territorial permission itself. States have other instruments, including wealth taxes, financial taxes, procurement and sovereign investment, and will use them. The most dependable base is the physical one, because it cannot move after the finance minister speaks.

That has a consequence for bargaining. A state whose dependable base is land, power and grid access is negotiating with the owners of land, power and grid access. It arrives at that table late, stretched, and needing them more than they need it.

The window for permits-for-equity is open and closing

The timing of the constructive move matters more than its design. The constraint essay identified permits-for-equity as the cleanest pre-commitment instrument under this framework. It is not the only one. Governments can invest directly, attach royalties, use development banks, keep public land, take rescue equity or tax gross flows. Its distinctive advantage is that it converts public scarcity into a permanent claim before capital is sunk.

The build-out still needs things only states can grant: land, planning approval, grid connections, water, transmission and political protection. Permission is not the state’s only leverage, but it is the cleanest. Before construction, a permanent public claim can be attached without a bailout, a retrospective tax fight or a threat of closure. It becomes a transparent condition of access that the developer can price before committing capital.

The leverage is uneven. A jurisdiction that demands equity can lose a sufficiently mobile project to one that does not. Permits-for-equity works unilaterally only where permission attaches to a binding, location-specific bottleneck: a constrained grid connection, publicly financed transmission, public land, sovereign procurement, guaranteed demand or rescue support. Where sites and jurisdictions are genuine substitutes, the claim needs regional coordination or it will be competed away.

The leverage does not vanish after construction. Expansions, power, water, tax treatment and regulation all continue. But it becomes costlier and more adversarial to use once the capital is sunk. The cleanest leverage exists on the way in.

The exchange rate matters. Trade permission for permanent public claims on the infrastructure. Do not trade permanent access to land, power and grid capacity for headline employment commitments. Construction jobs are temporary, and permanent operating payroll is small beside the capital value and public scarcity being allocated. Data centres do employ people in operations, networking, security, power engineering and their supply chains. The point is that automation will keep pressing down on a payroll that is already small relative to the concession. A state that swaps permanent physical access for temporary payroll promises will spend the next decade taxing an asset it forgot to keep a share of.

A crash narrows the window. Distressed compute changes hands quickly, survivors consolidate, and owners with sunk permits need less from the state. It is true that a crash gives the state new leverage as referee of the workouts. But workout leverage buys equity only if the state supplies rescue capital, guarantees or coercive law at the moment its own balance sheet is weakest. Permit leverage acquires the claim before the crisis. Workout leverage requires the state to buy it during one.

The bolder reply, that the state should wait and buy everything cheaply at the bottom, fails four times over. The fire sale is in the layer this series says is losing value: model businesses, leveraged shells and depreciating chips. The assets worth owning, the best grid connections, land, water and generation, keep or gain scarcity value through the crash, because the crash is value migrating into them. The buyer is fiscally and politically weakest on the day of the sale, with receipts falling and transfer demands rising, and weaker states may also face higher borrowing costs. Private capital clears a distressed auction in a weekend, while nationalisation needs a statute, a valuation and a court. And even a successful purchase buys depreciating machinery downstream of a chip supply chain the state does not control. A few currency issuers with cheap energy can play the bottom well. That is part of how states will sort themselves, not a plan open to the rest.

The leverage that turns permission into permanent claims exists cleanly now, while the concrete is still being poured. Afterwards the same claim has to be bought, taxed, litigated or rescued.

What would break this

This essay makes four separable claims, and the most exposed is pass-through. AI demand should not be treated as one variable. The evidence has to separate three groups of measures.

Infrastructure solvency: realised price per accepted output, paid workload revenue, cluster utilisation, debt service, refinancing conditions and collateral values.

Technological diffusion: deployed production workflows, machine-completed accepted outputs, the machine share of task execution, and cost per accepted output.

Labour displacement: human hours per accepted output, junior and graduate hiring, headcount relative to output, wage income in exposed occupations, and whether new occupations restore equivalent wage income.

Paid workloads can disappoint while the machine share rises. That is not a contradiction. It is how infrastructure investors and workers can lose in the same contraction.

The pricing claim is seriously damaged if neither independently hosted open-weight systems nor integrated volume tiers approach the best closed systems on accepted commercial outcomes at a materially lower quality-adjusted price, or if closed providers sustain their broad premium through 31 December 2027.

The credit claim fails if model-layer pricing compresses but realised merchant rental economics stay above debt service after the relevant borrowers have tested the refinancing market, through 31 December 2029. The sequencing claim fails if AI-exposed employment produces a clear break in headline unemployment before a composite credit turn at the leveraged edge.

The acceleration claim fails in either of two ways. If asset values fall but the cost of accepted output does not, demand or consolidation has stopped the write-down reaching customers. If the cost of accepted output falls but production adoption and machine share do not accelerate, cheaper supply has not produced the predicted diffusion. The labour claim weakens if human-complementary work restores displaced hours and wage income at comparable scale and speed.

If neither the headline labour break nor the credit composite appears after the relevant maturity window, that does not merely reorder the sequence. It falsifies the timetable, and weakens the wider claim that contestability is reaching the economy with the speed and force argued here.

Growing laboratory revenue would not falsify any of this. Commodity markets have large sellers, and revenue is not rent. Nor does a provider’s price cut establish post-crash pass-through. It shows delivered-price compression. The acceleration claim begins only when a distressed acquisition cost is passed into lower accepted-output costs after a credit event.

A proposed empirical test

What follows is a research proposal, not a completed preregistration. The task basket, borrower cohort and enterprise panel have not been published, recruited or frozen. The protocol becomes real only when those materials, the acceptance rubric, baseline model identifiers and the available data are disclosed. Until then it is the programme a funded research organisation would need to run. Any implementation should publish its baseline date in advance and use trailing ninety-day medians, so that no single promotional rate, distressed borrower or noisy employment release decides the result.

Pricing. Publish and freeze a basket of 48 commercial tasks across coding, research, document production, spreadsheet work, customer operations and regulated professional workflows. An independent panel selects the basket, or samples it from a published external task frame, under inclusion and exclusion rules disclosed before evaluation. Freeze prompts, tool access, latency limits and the acceptance rubric. At the first run, record the exact production model identifiers for the highest-scoring closed API from each of Anthropic, Google and OpenAI, plus the two strongest independently hosted open-weight models whose licences permit the test. An output is accepted only when two of three blinded domain assessors judge that it needs no substantive repair. Cost per accepted output includes inference, retries, orchestration and human verification. The human comparison uses task-specific, fully loaded labour costs from a preregistered external source on wages and overheads, with results also reported under common-rate sensitivity scenarios so that no single labour-cost assumption decides the outcome. Before evaluation, the rubric sets an absolute acceptance floor for each task category, so that risk thresholds can differ across coding, customer support and regulated professional work. An open option qualifies only if it clears that floor and reaches at least 90 percent of the best closed model’s acceptance rate. The open-weight contestability claim fails if no open release qualifies across the next two major open-weight generations. The broader pricing claim also fails if neither an open option nor an integrated volume tier produces material quality-adjusted compression and the broad premium falls by less than 10 percent from baseline in four consecutive quarterly runs through 31 December 2027. A fall of at least 25 percent from baseline, sustained for two quarters, counts as compression on schedule.

Credit. Publish a cohort of AI-infrastructure borrowers with debt of at least 30 percent of invested capital and a refinancing or maturity date between 1 January 2027 and 31 December 2029. Track paid workload revenue and realised quality-adjusted price per accepted output alongside financing measures. A credit turn requires at least three of the following, across two unrelated borrowers, within two consecutive quarters: a 25 percent year-on-year fall in quality-adjusted merchant GPU rental rates, a 15 percentage-point fall in disclosed cluster utilisation, a 250 basis-point widening in refinancing spread over the matched sovereign curve, a 20 percentage-point rise in collateral haircuts, a covenant waiver or take-or-pay renegotiation, or an impairment or distressed sale at least 25 percent below carrying value. The credit claim fails if pricing compresses and no composite turn appears by 31 December 2029, provided at least half of the cohort’s principal has passed a contractual refinancing or maturity date. If less than half has tested the market, the clock runs until it has.

Labour and sequence. Preregister 20 highly exposed occupations and 20 controls, matched on baseline wage, education, sector and prior cyclical sensitivity, in both the United Kingdom and the United States. Track employment, advertised vacancies, the share of vacancies asking for no more than two years’ experience, median advertised real wage, human hours per accepted output, and headcount relative to output. Track whether new occupations restore equivalent wage income, not merely whether new job titles appear. A quiet compositional turn occurs when the exposed basket underperforms its controls by at least 10 percent in vacancies and five percentage points in entry-level share for two consecutive quarters. That can happen before the credit turn and does not falsify the essay. A headline labour break requires unemployment to rise by at least one percentage point year on year in either country while employment in the exposed basket underperforms its controls by at least three percent. The sequencing claim fails if that headline break comes before the credit composite.

Acceleration. The first credit composite sets the event date. A funded research partner would recruit a vendor-neutral panel of 500 enterprises, stratified by country, sector and firm size, and record each quarter the share of defined production workflows using AI, machine-completed accepted outputs, the machine share of task execution and human hours per accepted output, excluding pilots and individual ad hoc use, with questions and workflow definitions held fixed. The write-down becomes an adoption subsidy only if quality-adjusted cost per accepted output falls at least 20 percent below its pre-event four-quarter trend within the next four quarters, and the panel’s median production-workflow share grows at least 50 percent faster than its pre-event four-quarter rate for two consecutive quarters. The joint labour-acceleration claim also requires human hours per accepted output, exposed hiring or labour income to deteriorate faster after the event than before it. If asset values fall but prices do not pass through, consolidation has interrupted the mechanism. If prices pass through but adoption and machine share do not accelerate, cheaper supply has not produced the predicted diffusion. Either result falsifies the acceleration claim.

The redistribution, named

A successor system cannot rest solely on the monopoly windfalls of frontier laboratories that competition is already eroding. A durable fiscal claim has to attach to volume, ownership and scarce physical complements, not merely to the accounting profits of whoever currently has the best model.

Put the three parts together and the movement is one sequence. Near-frontier open weights create a credible outside option. Standard interfaces and routing transmit it through one of two industrial structures, competitive hosting or integrated undercutting. Both can make routine cognition cheap. Only the first depends on continued competition among many serving providers.

From there the model forks. Wherever either route lowers the all-in cost of accepted output below equivalent human production, unit cost dominance spreads through firms and careers. Separately, wherever effective supply outruns paid demand or rental economics fall below debt service, obligations written against the old scarcity take the loss. If the written-down capacity then reaches customers as cheaper access, the credit branch feeds back into the labour branch and speeds diffusion. The demand loop can tighten both branches at once. Machine workload and machine share are different quantities, and they can move in opposite directions.

Every arrow is conditional, and the tests above say where each can fail.

Where the conditions hold, labour loses bargaining power because reproducible cognition stops being scarce. Frontier laboratories need not become unprofitable, since some can prosper through successive leads, products and integration, while the scarcity rent on routine cognition compresses beneath them. Leveraged investors lose where obligations were sized to a premium that reprices faster than the loan. The state loses where its broadest and most automatic tax bases move faster than its tax code. The owners of whichever physical and legal complements still bind, including leading-edge fabs, constrained grid connections, scarce power, permitted sites, ports and enforcement, inherit the position, because the software layer prices itself towards utility economics and the machines still need atoms. That is the redistribution under way: from cognitive capital towards physical capital.

Part I opened with a shrug. The second DeepSeek moment arrived and almost nobody noticed, because the quiet parts of a discontinuity are easy to shrug at. Hiring that never happens makes no sound. Entry-level roles disappear without a press release.

Credit is different. Credit has covenants, coupons and dates. The first part of this sequence that nobody manages to shrug at will be the margin call.

A correction follows

The movement from cognitive capital towards physical capital is directionally right, but incomplete. Physical capital is one possible binding complement, not the economic endpoint. The coda supplies the corrected conclusion: rent follows the slowest-decaying complement that still binds, splits when several bind at once, and vanishes when none do.