Section XIV: XIVThe Dilemma and Its Adjudication

Marx’s Grundrisse passage known as the “Fragment on Machines” names the end state: once labor in its direct form has ceased to be the great well-spring of wealth, the theft of alien labor time on which the present wealth is based “appears a miserable foundation in face of this new one.” Marx wrote this in 1858. He could not have anticipated AGI specifically. He specified where the dynamic would tend.

This creates a specific dilemma for the AI build-out. The build-out must somehow preserve wage labor as the dominant form of production to preserve capitalism as we know it, in which case the valuations made on abolishing wage labor cannot self-validate and the entire architecture liquidates itself through the Volume III mechanism. The layered claims running between Google, Anthropic, Microsoft, OpenAI, and Amazon are exposed as such, and the bubble bursts.

The other horn has a name in Marx too. In Value, Price and Profit, his 1865 lectures to the General Council of the International Working Men’s Association, Marx named the Communist revolutionary horizon as ‘the ultimate abolition of the wages system.’ The AGI promise under capitalism reaches for the same end state from the “wrong” end. Capital is attempting to abolish the wage system from above through the elimination of labor itself, terminating the wage relation through the obsolescence of labor rather than through the emancipation of labor from private ownership of commodity production and surplus value extraction. Capital cannot complete the act of this replacement, because the system of capitalism itself is the fuel that this attempt at replacement is running on, and the fuel may run out before the build is complete. And labor will not forget the gun that was just placed at its head by capital.

An intermediate path is discussed, where some analysts predict slow productivity diffusion that still continues the trend of augmenting rather than replacing labor while gradually validating capex at a lesser valuation, which can take place over a longer time horizon while we mitigate the worst impacts. The architecture cannot take this path at its present prices.

The architecture cannot take the gradual path because its valuations forbid it: equity valuations are priced on the discount rate replacement justifies. The constant capital has to return its cost inside an economic life near 36 months, whatever the five- and six-year schedules on the books say. Reversion to augmentation and slow-pace diffusion produces the Volume III liquidation of current valuations and capex expenditure. The largest laboratory’s chief conceded the timeline in August: the economy has so much inertia, we have all been too ambitious on timelines, the transition will be smoother and slower.¹²⁰

The circuit forbids the gradual path from the inside as well. The labs’ survival depends on their hyperscaler patrons’ growth narratives, and the growth narratives are contingent on the labs’ spend. Operating efficiency, the only path to viability and profitability for OpenAI and Anthropic, is ironically disqualifying rather than merely unavailable: a lab that halves its compute bill to create profitability also halves capex expenditures, its patron’s growth story, the growth story of the chip fabs, and with it the reason to keep the AI lab alive. A circuit that destroys its participants for repairing themselves to achieve profitability has answered the question: there is no gradual path to soundness. The venture layer beneath them has its own record. Thrive’s 2022 growth fund has returned 30 cents on the dollar and sits in the top 5% of its vintage. The median venture vintage from 2017 to 2024 has returned between 0.92 and 1.23 dollars per dollar, against roughly 2.8 for the index with dividends reinvested over the same years. The allocator whose judgment is offered as proof that the capex is sound has trailed the index for a decade.

The token price chart of Section XII records a second mechanism beside the falling rate, and this one has a name and a date. Marx called it “moral depreciation,” the destruction of invested capital by competition before its value can be recovered through use, and it asks whether the capital expenditure retains its value at all, prior to any question of market validation. July supplied the demonstration in two stages. On July 17 the unveiling of a Chinese model, Moonshot’s Kimi K3, which among other demonstrations autonomously designed a functional chip, triggered a repricing across the American AI complex: the Philadelphia semiconductor index extended its decline past 20% from its high, and Oracle’s default insurance printed its all-time record the same session. That was just the announcement.

The price event came 10 days later, on July 27, when Moonshot released the full weights on HuggingFace under a permissive license. Kimi K3 is a 2.8-trillion-parameter mixture-of-experts model ranking second and third on the major intelligence indices behind only the leading American frontier systems, first in frontend code, matching them on coding work at roughly a third of the cost, with third-party hosts serving it commercially the same day. A capability demonstration can be dismissed as benchmark selection, but weights that anyone with a rack can run, and resell, cannot. If artificial intelligence commoditizes on Chinese efficiency curves, the trillion-plus dollars of constant capital deployed for AI is devalued by the cheapening of its own replacement before the question of whether it replaces labor is answered.¹²¹

The sellers describe the horizon in their own words as a world without people in it: job sites without workers, social media without socializing, programming shops without programmers, and the same class that funds it is alarmed about falling birth rates. The abolition of variable capital is here an aspiration of the people financing it, and the contradiction they leave out is the one this section runs: a world without workers is a world without the value the buildout is priced to capture.

The Volume III failure case says the machine cannot think well enough to service the debt based on it thinking well. The Grundrisse success case says the machine thinks so well that value itself ceases to be measurable by labor time. The moral-depreciation case, the one now trading live, says the machine thinks adequately and thinking becomes too cheap to service debt priced on the assumption it would stay expensive. These are the same partition the value theory draws from the price: the failure case is the human with a tool, the success case is the route ending, the moral-depreciation case is the price falling through the cost, and the rent that holds prices above that cost is the fourth condition, enclosure, which is not a case of the machine’s capability at all. This is all delivered by ordinary capitalist mechanism, requiring no judgment about the reality of artificial general intelligence at all, and it is the sharpest version of the circuit’s terminal problem: the build-out’s own premise, that advanced artificial cognition is about to become abundant, is the exact premise that strands the capital betting advanced artificial cognition would likely remain scarce.

The dilemma above still partitions exhaustively on whether wage labor remains the dominant form of production and admits no middle term. Moral depreciation destroys the constant capital before the question has been settled, and it does so through a mechanism that requires only that a competitor can supply the same capability below the price the debt was underwritten against. The dilemma establishes that there is no exit for the capital as it is priced. Both horns press on the same financing: an exit would need productivity gains large enough to validate the capex and a wage relation that survives them, with the financing lasting until both arrive, and that conjunction is the test the closing section sets. The competitor that proves the premise fastest and cheapest destroys the balance sheets of everyone who financed it. The successor’s laboratories do not need to outspend the American buildout and “win” the frontier race. They need only keep releasing models that show the buildout was massively overpriced, which strands the capital the buildout priced.

The tape shows three conditions, and in none of them does the machine produce value. The price falls through the cost of compute wherever discovery is allowed, which is the case above. The price is held above that cost by enclosure, with the product degrading under it, which is rent. The human operates the model with judgment, and the hour is the value while the model is intensity, which is the case of the human with a tool. Where the build-out lands among them is a question of capability. The hype decides only the size of the bet.

The exit a reader from the finance side proposes, cut the capital expenditure by half, impair the assets, return to the cash flows of search and cloud, is the one exit the circuit forbids, and the reasons are on four ledgers already kept. The seven companies trade at growth multiples on a third of the index because the buildout is the growth, and a company that halves its spending has priced itself at a utility’s multiple and told the market the buildout was not. The repricing runs through the collateral that a third of the index has become. The neoclouds and the private-credit wraps hold the chips as collateral against offtakes from the laboratories, and a halved buildout cancels the offtakes and marks the chips at their salvage value, the write-down landing on the insurers and the retirement funds rather than on the sponsor. The buildout is 39 to 75% of measured growth, and its halving takes a point of GDP or more out of a growth print of 1.5% before the collateral moves, the imported share of the machines the largest offset.¹²² And the venture marks, the software multiples, and the sellers’ own equity are one asset, so that the impairment the finance reader proposes as a housekeeping item is the same event as the crash it is meant to avoid. The pivot is available to a company. It is not available to a system that has made the company its floor.

The enclosure the state could build by rule is the last counter, and its first bricks are on the tape: the 2026 defense authorization bars one Chinese laboratory’s models from Pentagon systems and contractors, and in July the administration was reported to be weighing Entity List designations, procurement bars and liability rules for the rest.¹²³ A compliance cartel can bar Chinese-origin weights from the defense contractor, the hospital, and the bank, and this work expects one by the end of 2027, and the sliver it protects is a cost center: the compliance is paid for by the customer and the rent inside it is set by what the customer would pay outside it. Outside it, Chinese open weights already serve close to half of enterprise tokens through the routers American firms use, and the Global South and the non-aligned economies, where the volume is, run on them without a rule to bar them. The war’s own AI supplied the test from the other direction: on a former deputy assistant secretary’s account the incumbent’s targeting gains came from data architecture and from models unveiled in 2025 or earlier, the frontier having no part in them, so that the race to the frontier is not what is shaping the war, and the open weights will hand the same tools to anyone.¹²⁴ A cartel that holds the price inside the regulated third of the market while the other two thirds discover the cost is a cartel that has told its own regulated customers the size of the tax, and every prior enclosure of that kind, the Bell system, the mainframe, the licensed software of the 1990s, lasted until the customer could see the price outside the wall. The wall is visible from inside on a router dashboard.¹²⁵

The removal of scarcity had its mirror on the physical hardware side, and the Korean market priced it. The day after the weights were released, July 28, on the public offering of CXMT, the Chinese memory manufacturer, and the fears about DRAM pricing that its scale revives, the Korean benchmark took its eighth circuit breaker of the year, falling 8.02% to 6,213.51, representing the fourteenth market-wide halt on record, against a 2026 count that by late June, the prior month, had already logged roughly 30 sidecar activations and five circuit breakers, both figures past the 2008 crisis tally. Samsung Electronics and SK hynix, tech stocks together about half the index’s capitalization, fell some 11% and 12% respectively. The weights had taken the scarcity out of the model layer, and CXMT’s offering took it out of the memory layer the next day.¹²⁶

Neither Moonshot nor CXMT is attempting to beat the incumbent “western” competitor at the frontier. Both are flooding the layer below where the incumbent’s rent is collected: one giving away for free the weights and models that American labs price at a premium as scarce cognition, the other adding capacity to a memory market whose pricing power the hyperscaler build-out has been creating with high demand and low supply, and whose margins the surviving DRAM producers have booked as the return on that buildout.

This method is the SAME EXACT one that ran through solar, batteries, electric vehicles and shipbuilding across the prior 15 years, and it does not require technological leadership. It requires only enough capacity, at adequate quality, to destroy the price at which the initial position was capitalized. We identified TSMC’s leading-node monopoly rent as the place where the AI circuit’s visible profits are actually booked rather than extracted. This is the mechanism by which those same rents are removed, layer by layer, on a timing Washington does not set, and it indicates which layers go next, the ones currently priced as scarce: DRAM and HBM contract prices, the frontier price per million tokens, and the leading-node wafer price are the series to watch.

The EV instance now has its ledger entry. EY’s April study counts write-downs near 60 billion euros across the 19 largest carmakers in 2025, battery joint ventures, shelved plants, and cancelled models, and a 59% fall in their combined profit, from 143 to 59 billion. Geely was the only one of the 19 to grow. China exported 2.2 million cars in the first quarter of 2026, nearly a million of them electric, and EY’s own gloss on the war was that rising energy prices and inflation are poison for the industry regardless of drivetrain. The constant capital was devalued by the layer below it before the product reached the customer, which is the clock this section describes, already run once on the industry whose loans the household is still paying.

Both states have already registered the move as strategic rather than commercial. Washington’s science and technology policy director publicly accused Moonshot of training on controlled Nvidia silicon and of large-scale distillation against American models. Beijing’s Ministry of Commerce was simultaneously reported to be consulting its own firms on export controls covering models, weights, and training data. The consultation cuts both ways: a control on weights would hand Beijing the valve, which so far the laboratories hold. Both governments treat a weight file as live ordnance in an ongoing economic war rather than as a product.

Three jaws close on the circuit at once, the profit rate, fictitious capital, and moral depreciation, and Marx’s Volume III specifies how the unwinding proceeds once they do. When participants withdraw their underwriting, the system ceases to reproduce itself.

The unwinding of underwriting has already begun. SoftBank reduced its OpenAI-backed margin loan from a planned $10 billion to $6 billion in early May, after lenders refused to assign reliable collateral value to OpenAI shares at the company’s $852 billion post-money valuation. By September the same stake was being financed in the junk market: $11 billion of bonds marketed near 10% to fund the third $10 billion tranche of a $65 billion commitment for 13% of the company, from a borrower already the year’s largest junk-rated issuer, its 2031 bond yielding 8.2% from 6.7% in January and its shares a third below June’s high.¹²⁷ The Wall Street Journal reported in April that OpenAI had missed multiple internal sales targets. The Microsoft-OpenAI partnership was rewritten on April 27 to remove exclusivity, end Microsoft’s revenue-share payments to OpenAI, and eliminate the AGI clause that previously governed the relationship. Sora video generation was shuttered. Microsoft’s GitHub Copilot moved to unsubsidized token-based billing on June 1. Anthropic’s chief financial officer has stated publicly that he is worried about meeting the company’s compute commitments. Each of these is the same signal at a different layer of the circuit: lenders, partners, customers, and operators are each beginning to decline to extend the underwriting at the rate required to reproduce the build. The summer’s tape in Section XIII reads the other way on its face: guarantees at $108.5 billion, supply commitments doubled in a quarter, a $35 billion offtake signed on the last day of August. The two records are one movement. The lender at arm’s length is leaving, and the seller of the chips is taking its place as guarantor, tenant, and creditor, so the underwriting that remains is concentrating on the one balance sheet that books the chips as revenue. The claim takes its tickers here: the guarantee and receivable lines in each Nvidia 10-Q, against neocloud and laboratory debt priced with no vendor behind it. If arm’s-length credit for the laboratories reopens through 2027 without a vendor’s guarantee, the unwinding owes the revision. Much of this financing runs from the sellers of compute to their buyers, vendor finance where arm’s-length lending used to be, so a revision in what the buyers can pay reaches the sellers’ books through the same contracts.

The standard read of these signals is again that they constitute a fairly standard bubble correction: capital was overdeployed, the market reckons, the position rebalances, the system continues. The read fails to register what the AI capex circuit has become structurally, inside the broader US and global economy. The Magnificent Seven (Apple, Microsoft, Alphabet, Amazon, Meta, Nvidia, Tesla) account for 33.7% of the S&P 500 as of April 2026, up from 12.5% in 2016, and drove approximately 42% of the index’s total return in 2025. Morgan Stanley estimates AI capital expenditure contributed roughly 75% of US GDP growth in the first quarter of 2026. The St. Louis Fed identified AI-related investment as 39% of marginal real GDP growth across the prior four quarters. The second quarter printed 1.5% annualized against a 2.3% consensus, with technology investment half of the year’s growth on the broad count and a third net of the imported equipment.¹²⁸ The AI capex circuit is now the load-bearing pillar of both the equity market and the GDP growth profile. Strip it out of the index and the index is meaningfully smaller. Strip it out of the GDP series and the economy is in recession.