Section XV: XVThe Failure of the Realization Circuit

The AI capex circuit is the shunt: an economic heart redirected to pump blood into a closed loop that does not connect to an actual body. Realization of value, in Marx’s framework, is the moment when a commodity is actually sold on the market. Production produces surplus value, realization converts it into money, the M’ in M-C-M’. A factory that produces commodities no one ends up buying does not realize the surplus value its labor produced. The capital invested in production cannot be recovered. The circuit must close on the demand side or it does not close at all, and it has three places to close, each of them compromised: the enterprises that buy the machines, the households that buy what those enterprises make, and the owners whose spending rides on the machines’ own valuations.

The American mass-consumption economy that anchored post-war global capitalism has meanwhile been hollowing for decades. The top 10% of US households now account for roughly half of all consumer spending. That half is not a floor. It rides on what its spenders own, and a third of what they own in the index is the build-out’s own valuation, so the spending that looks insulated from the bottom’s squeeze is the circuit’s collateral spent as income.¹²⁹ The bottom 50% accounts for less than 15%. Real wages for the bottom half have been roughly stagnant since the 1970s. The consumer-facing productive economy is reporting demand collapse. Wealth and income concentration has reached pre-1929 levels.¹³⁰

The household’s stocks have passed its houses as the main source of its wealth for the first time since the Second World War, equities a record 48% of its financial assets, above the dot-com peak, and the top tenth of households holds about 87% of them. The wealth effect that now carries consumption is the top decile’s, and the business cycle that once ran through the house now runs through the build-out’s share prices.¹³¹

The corporate earnings calls of 2026 have made the demand collapse explicit and quantifiable beneath the usual hedged and normalizing language. Whirlpool reported on May 7 that US appliance demand was the weakest since the 2008 financial crisis, its chief financial officer attributing the collapse directly to consumer-confidence damage from the war. Kraft Heinz reported customers “literally running out of money at the end of the month”. McDonald’s, that the consumer environment “might even be getting a little bit worse.” Consumer sentiment in May fell to 48.2, a record low, as the 30-year fixed mortgage rate rose above 6.5% and homebuilder stocks sat 20% off their 2024 highs.

The same 50 years that eventually produced this speculative pyramid also produced the rotting real economy underneath it, in measurable disrepair. The American Society of Civil Engineers’ 2025 Report Card graded American infrastructure at C overall, with 9 of 18 categories rated D or D+: roads, aviation, transit, storm water systems, drinking water, wastewater, schools, energy, dams. The estimated gap between current funding levels and what the system requires is $3.7 trillion over the next 10 years.¹³² The average age of US bridge inventory is 47 years against a 50-year design life.

The infrastructure even required to facilitate the AI build-out itself is also short. US data center power demand is projected to more than double from 31 gigawatts in 2025 to 66 by 2027, with utility forecasts adding about 90 gigawatts of data-center peak load by 2030,¹³³ and the Uptime Institute identifies power as the single defining constraint on data center growth globally. Sightline Climate’s April 2026 audit found data centers representing 12 gigawatts of new demand announced for completion in 2026 against roughly 5 gigawatts under active construction in 2026, a 7 gigawatt shortfall. The installed base says the same from the floor: about a sixth of the capacity Microsoft runs holds AI chips, on Bloomberg’s September reporting. Hyperscalers have responded by underwriting more than 10 gigawatts of new nuclear capacity in the past two years: Microsoft’s 20-year agreement to take the output of a restarted Three Mile Island unit, a $1.6 billion restart,¹³⁴ Amazon’s small-modular-reactor deployments with X-energy, Google’s with Kairos, Meta’s commitments to unlock up to 6.6 gigawatts, and most of it will not come online before 2030. Capital and labor that could have been deployed to maintain and grow infrastructure for actual human use and the productive economy is being actively routed into an unproductive and unprofitable build-out that consumes more infrastructure than it can conjure to even fuel itself.

Before the AI circuit’s output has to find a household it has to find an enterprise customer, and the enterprises buying it are substantially the software companies whose own revenue the buildout’s premise eliminates. Seat-based software prices a claim on the number of people employed doing knowledge work. The AI product sold to those companies is sold on the proposition that the number falls.

On the customers

An example of this: monday.com reported first-quarter 2026 revenue of $351.3 million, up 24%, net dollar retention at 110%, record operating income, and raised full-year guidance, and its shares fell from a 52-week high of $317 to roughly $70, about 3.3 times sales on 89% gross margins, carrying $1.2 billion of net cash, now 30% of its market capitalization. ClickUp, still private, cut 22% of its workforce in May while simultaneously evangelizing to the market the brilliance of AI adoption that removes the seats its own revenue is a count of.

The build-out’s own paying customers are being priced for extinction on the strength of the build-out’s promise, and their results show nothing a demand failure would show, yet: their operating numbers in the summer of 2026 were still good. What failed is the belief that the revenue persists medium term, and the equity market marked that belief down by 1/2 to 4/5ths across the category inside two quarters. These are the theoretical enterprise clients for AI. The print that would show it is net dollar retention below 100% at the seat-based vendors by the end of 2027.

Meanwhile, realizable market is being repriced, and the repricing is visible exactly where price discovery is allowed to happen. On the neutral routing marketplaces where models compete head to head, Chinese open-weight share has gone from under 2% in late 2024 to roughly 1/2 of all tokens served by midsummer 2026: 14 consecutive weeks in the lead, 9 of 10 most-used models Chinese, at prices running from a 50th to a 20th of the American frontier’s. Captive channels still carry the majority of global volume, the enterprise workloads locked to the US hyperscaler clouds, which is the bifurcation as infrastructure: a high-margin sliver defended by law, contract and compliance, and a mass volume business wherever the buyer is free to choose. That majority is the rent condition of Section XIV: the price held above the value by enclosure, and the product degrading under it, for which the popular name is “enshittification.”¹³⁵

And this is under subsidized pricing. The first time enterprises were charged something near actual API cost, the vendors’ own people describe the customers screaming about it.¹³⁶ The following 80% price cut was priced against Chinese competitors, and the pricing is being held down and the rationing is on the buyer’s side¹³⁷ and canceled deployments. It has since been run across a market of paying users. Within twelve months the coding assistants moved from flat rates to metered tokens, the largest on June 1, its seller saying in the announcement that it had absorbed much of the escalating inference cost and that the old pricing was no longer sustainable, and developers reporting allotments burned through in hours and projected overage bills in the hundreds and thousands.¹³⁸

The price has a floor the vendors are still below, and holding it down is a wasting asset. Inference runs on fixed costs, the data centers and the power contracts, spread over the users who pay. Every user who declines the price at $200 a month raises the share of the fixed cost the remaining users must carry, and the price that clears the remaining pool is higher than the price that cleared the whole one. A subsidized supersonic shuttle between two cities would be full at 10 pence a seat, and its passengers would stop investing in the railway, and the question that decides its future is who pays the difference when the subsidy ends. The frontier is the shuttle. The open weights the challenger ships are the railway.

The objection this invites has a name, Jevons, and it does not apply to the frontier. The paradox says a fall in the cost of a service raises its total consumption by more than the price fell, the steam engine’s coal and the light bulb’s electricity, and it requires the cost to have fallen. The frontier’s price is falling and its cost is not: the price cuts of July and August came from customers refusing the price, with the laboratories’ costs rising and $230 billion raised in six months to cover them, which is subsidized dumping and not efficiency. The paradox does apply to the challenger, whose cost has fallen, and the volume it releases runs on open weights and pays the frontier nothing. For the frontier it is a paradox about cheaper work applied to work sold below cost. And where the volume would have to come from to rescue the revenue line, the physical ledger kept above has already set the ceiling: the 31 to 66 gigawatts the standard series allows and the 7 gigawatt shortfall against them, so that the volume the paradox needs cannot be served at any price the sellers can charge.¹³⁹

Financial Times reporting of Ramp payment data in late August found the costliest frontier vintage at roughly 11% of enterprise spending on its maker’s models two months after launch, undersold by cheaper sibling models released weeks later, the first break in the pattern of enterprises migrating to the newest model on arrival. That is Section XIV’s moral-depreciation clock running inside the leader’s own price list. The buyers said it in their own words at the software industry’s largest gathering in September: models a year or two old do the sales and service work, Salesforce’s own agent product runs on none of the newest releases, the large buyers route each request to the cheapest model that answers it and keep the frontier for the hard cases, and one analyst put the token bill where it lands, on the software vendor’s gross margin, from above 85% toward 45%.¹⁴⁰

The laboratories’ accounts also show what Marx called moral depreciation at its fastest. A frontier model is constant capital that must be replaced every six to eighteen months, and the argument over the prospectus is whether its cost belongs in the operating loss at all: on an inference-only count the laboratories may already run a positive cash flow, and on the full count Anthropic spent about $2.75 for each dollar of its 2025 revenue. Paul Kedrosky put the dilemma plainly. Stop training and the moat goes, leaving a producer of commodity tokens against an industry in China that already sets the floor price on cheaper power, and keep training and the losses keep pace with the moat. Either way the rent the valuation assumes has nowhere to come from, which is the lesson of the solar panel applied to the model itself.¹⁴¹ The seller has begun to buy the commons the commodity runs on. In September it agreed to pay $12.9 billion for Hugging Face, the hub where more than three million open models, Chinese ones among them, are shared and downloaded, betting on open models as its largest customers build chips of their own, in a season when a payments company paid $7.5 billion for a router of open-weight models. If the model becomes a commodity, the profit moves to whoever owns the market it is traded on.¹⁴²

On the wage

The August 7 report supplies realization failure’s necessary condition without proving it, and the proof depends on its composition rather than its headline. The load-bearing line is not the headline, which the September revision turned from a loss of 23,000 into a gain of 21,000, but the wage. Annual wage growth slowed to 3.2%, the lowest since May 2021 and below June’s CPI inflation of 3.5%.¹⁴³ After the revision wage growth stood at 3.1%, still under the price index, with net hiring over the trailing year near 31,000 a month.¹⁴⁴ Wages rising slower than prices shrink what each job can buy, which is what the realization condition turns on.

The triple discriminates between the two available explanations. The administration’s account is supply contraction: deportations have shrunk the workforce, the break-even hiring rate has fallen toward zero, and negative months are consistent with health. But a genuine supply contraction with intact demand bids wages up, because scarce labor is expensive labor. Wages are decelerating below inflation while the labor force shrinks, which means labor demand is falling at least as fast as labor supply. That falls short of proof of realization failure while supplying its necessary condition, and it is flatly inconsistent with the account under which the report was celebrated.

The damage the labor market is absorbing has a name the balance sheet does not carry. Intellectual property is what a firm can write down and sell. Process knowledge is what its workers know and cannot write down, which machine jams and who fixes it, which customer will take the defective batch, which room not to enter alone, and it is the difference between profit and loss in every shop that has ever run. A firm that dismisses the worker for a chatbot that cannot do the job destroys that knowledge in the dismissal, and rehiring does not restore it, because the worker has retired, retrained, or left the workforce by the time the chatbot is switched off. The cost is a productivity loss that outlasts the bubble in the sectors that did it, and it is why the bubble’s critics want it popped early rather than late.¹⁴⁵ Where the replacement is real it is also temporary. It runs on tokens sold below their cost, so the saving is borrowed from the seller’s investors and shrinks as the price is metered toward the cost, and the firms that made it are already reversing it: Gartner expects half of those that cut service staff in the technology’s name to rehire for the same work under other titles by 2027, and Forrester half of all such cuts to be quietly undone, the jobs returning offshore or at lower wages, which is the wage bill degraded and not the labor removed.¹⁴⁶

The instrument itself deserves one paragraph of suspicion, stated precisely. The headline rate’s definition has not changed since 1994. Its blindness is architectural rather than statistical, exit is invisible by design, and any month the labor force shrinks faster than the count of the unemployed, the rate improves while the condition worsens, which is July exactly. What has degraded is the apparatus around the definition: survey response rates have fallen for a decade, and the birth-death imputation flatters payrolls at precisely the turning points where the business formation it extrapolates has stopped, the mechanism that kept 2008’s real-time prints mild until the benchmark revisions erased them, and the mechanism behind the hundred thousand jobs just erased from this spring. Against the flatterable series stand the gauges that cannot be adjusted into comfort: employment-to-population at 58.9%, the lowest since 2014. A hires rate scraping decade lows in a market the coverage itself calls low-hire, low-fire. A quarter of the unemployed out of work more than six months. Low unemployment, low hiring, and long duration describe few hires, few fires, and exits through nonparticipation rather than a strong labor market. Even the month’s celebrated fall in teenage unemployment resolved, in the same day’s flow data, into the young ceasing to look rather than starting to work.

The same design principle governs the leverage statistics. The visible margin book, a record 900 billion dollars, the number the retail scare stories carry, is relatively smaller than the brokers’ loans of 1929, and it is the decoy: the borrowing migrated to where no single tape can count it, the basis trade, the off-balance-sheet trillions, portfolio margin, the options complex’s embedded gearing, the private books that mark to nothing. In 1929 the leverage sat visibly on one ticker a Babson could total. The present structure’s innovation is leverage made uncountable by architecture. And the architecture has a lineage: each invisible layer is an escape from the 1933 settlement, the credit moved off the books the SEC was built to read, the savings channel moved outside the deposit-insurance perimeter, the margin rules escaped by instrument design, the era’s largest enterprises organized never to enter the public-disclosure regime at all. The reforms sworn in 1929’s memory were left in place and built around, and that is why the instruments fail.

The composition completes the reading. Private payrolls rose 30,000; construction contributed 22,000 and health care 22,000; strip out the buildout and the substantially state-reimbursed health sector and the market-facing private economy shed workers outright. The construction gain is itself a monoculture. Nonresidential specialty trades expand on data-center demand, the industry’s own economists attribute the July gain to the buildout directly, while residential construction logs its seventeenth consecutive month of year-over-year decline and private nonresidential spending outside the data-center segment has shrunk for seven consecutive months, the buildout crowding out the commercial investment around it. Read through the reproduction schemes of Volume II, the table is the closed circuit rendered as sectoral employment: Department I expands wherever the capex flows, Department II sheds labor everywhere wages would have to be spent, retail down 19,000, leisure and hospitality down 40,000, and the financial intermediation layer, down 121,000 jobs since its May 2025 peak, has been shedding its own workforce for 14 months. This is the Tugan-Baranovsky configuration, accumulation sustained by Department I producing for Department I, and the classical objection to Tugan holds with full force: the spiral has no independent stability. It runs exactly as long as the credit bridging it runs. The construction workers hired in July are producing real surplus value whose realization is a claim on AI revenues that do not yet exist.¹⁴⁷

The purchasing class has also now been surveyed directly. The National Bureau of Economic Research working paper by Nicholas Bloom and coauthors, “Firm Data on AI,” surveying nearly 6,000 executives in the United States, Britain, Germany, and Australia between November 2025 and January 2026, three years into the deployment, found 89% reporting no impact on labor productivity from the technology. Set the number beside the celebrated report above. The spend appears in the capex line and the GDP print. The productivity the spend was to purchase is missing from a survey of the executives who decide the purchases. This is the third condition of Section XIV, the tool in a human’s hand, whose multiplier nine surveyed firms in ten, users among them, have not found. Other surveys find gains among adopters, uneven, and a European firm study puts the adopters’ productivity level 4% higher, none of it measured against the capital advanced, which is this work’s further question.¹⁴⁸ The New York Fed’s August survey of its own district has the same shape at the firm: 61% of service firms and 51% of manufacturers using it, three quarters of the first and nine tenths of the second calling their spending minimal to modest, and among adopters a median 17% and 7% of workers using it.¹⁴⁹ Put together, the circuit fails at both of its buyers at once. The enterprises cut the wage bill in the machine’s name, whether or not a machine does the work, and the cut removes the knowledge the machines cannot replace and the income their own customers live on, without the productivity that would pay for either, which is the loop of the next section’s layoffs. The households split along the K, the bottom’s demand eroded by the same cuts and the war’s prices, the top’s riding on the valuations the build-out sets. Every source of the circuit’s revenue is therefore its own financing, spending carried by its own marks, or income its labor strategy removes, and the shunt cannot close even where the product works.