Section XII: XIICapitalism’s Self-Consumption of Capitalism

Where the shells and the interceptors were not built, the data centers were: the build’s single-year spending against a defense request Congress has not funded, which is the allocation the war is now pricing. It rhymes with the Genoese fairs, the incumbent’s money lent to its own last reach, though the fairs financed a war and this finances a market that must pay it back. In October 2025 the US company 1X Technologies opened pre-orders for NEO, a $20,000 robot the company described as the first consumer-ready home robot, one that would fold your laundry and cook for you and eliminate repetitive household tasks. The launch tweet drew tens of millions of views and over 50,000 X posts within 24 hours. The company was simultaneously attempting to raise up to $1 billion at a $10 billion valuation, 12x its January 2024 figure.⁷⁷ OpenAI’s own Startup Fund had led the Series A in 2023. The robot was framed as “embodied AI,” the physical manifestation of OpenAI’s general-intelligence project.

The Wall Street Journal’s Joanna Stern tested it and wrote up the experience. She reported that 100% of NEO’s actions during her demo were tele-operated rather than autonomous. A human 1X employee in another room, wearing a VR headset, was piloting the robot through its motions. The robot took two minutes to fold a shirt. It teetered on its heels trying to close a dishwasher. It could not crack a walnut. It lifted a fraction of the 154 pounds the company had claimed it could carry. Its battery lasted four hours.

The company calls the human-driven experience “Expert Mode.” The CEO has described it as a “social contract” with early adopters who, in exchange for the privilege of paying $20,000, contribute the audio and video of their daily lives as training data for the autonomous system that may eventually exist. The robot’s cameras and microphones stream the interior of the home to the operator. The privacy policy presumes the surveillance and works to make it tolerable, or at least priced at a reasonable exchange rate for maybe not having to do the dishes.

Marques Brownlee called the launch “a hype reel for a thing that they’re hoping to be able to make someday.” The YouTube channel Fireship characterized NEO as “a cloud connected device that can watch and listen to your family all day,” an Amazon Alexa with the optional feature of occasionally moving stuff around slowly, potentially damaging your property in the process, with “tech support from a remote human wearing a VR headset.”

A circularity started to emerge. In December 2025, 1X struck a deal with its lead investor EQT to ship up to 10,000 NEO robots to EQT’s portfolio companies between 2026 and 2030. The robot marketed for the home pivoted to the factory. The investor became the customer for the product. The investor’s portfolio companies will pay for machines piloted by humans that 1X employs to perform the labor of actual human employees.⁷⁸

OpenAI is correct: the 1X NEO is the general promise of AI in a body. It is the embodiment of what the AI build-out is: a promise of a generally intelligent agent that can navigate the world, perform varied labor tasks, and replace humans in the labor process. What was actually delivered, packed inside the soft 3D-knit exterior, is human labor teleoperating from another room, waving its arms around wildly while the robotic replacement fails to function, delivering a sub-par, expensive, and incomplete substitute for the human. The “autonomous home labor” that drives the valuations and the hype is data being collected to train a future autonomous AI that may or may not eventually exist. Nothing below depends on the machine failing. The argument runs the same whether the autonomous robot arrives next year or never, and the sections that follow take each case in turn.

The non-physical chatbot-based AI build-out is strikingly similar. The premise the hype is intended to sell, living labor replaced by constant capital as the source of returns, is what makes the trajectory for capitalism terminal in a specific sense Karl Marx identified more than 150 years ago. There are a few paths, and they are all terminal, and what decides which terminal path it takes is capability: whether the machine can do the labor and rebuild itself without a human hand on it. What the hype decides is how much capital is committed before that question is answered definitively one way or the other, and how long the definitive answer can be deferred.

To Marx, capitalism is generalized commodity production, things made to be sold, through wage labor, humans making them for a wage, with the extraction of surplus value as the engine and the self-expansion of capital as its mode of existence. The defining feature of capitalism is not that markets exist, or that private property exists, or that exchange occurs. Those existed in ancient slave societies and in feudal Europe. What distinguishes capitalism is the dominant social relation of production: privately owned commodity production by wage labor, in which the worker’s labor creates more value than the worker’s wage returns, and the private owner takes the difference as profit.

Marx’s general formula for capital is M-C-M’. M is money advanced by a capitalist. C is the commodities it buys: labor-power, and the means of production, machines and materials. M’ is the larger sum the finished commodity sells for. The difference, M’ minus M, Marx’s ΔM, is surplus value, and the system must run on its continuous expansion, which is observable as the capitalist’s constant chase for a positive return on investment: profit / capital advanced.

Marx specified a deeper structural point in Volume III chapter 15: “the real barrier of capitalist production is capital itself.” Marx meant that the drive to expand value runs into the limits set by the relations that produce it. The extension is this work’s own: an accumulation process that promised to run with no humans in it would be either impossible or not capitalism, and AI capex is that barrier made visible.

Combined 2026 capital spending at the four largest technology hyperscalers, Microsoft, Google, Amazon, and Meta, is guided near $725 billion, up about 77% from 2025’s record of roughly $410 billion, per first-quarter earnings reports compiled by the Financial Times. Microsoft alone has guided to $190 billion for calendar 2026, about 130% year-over-year growth, with roughly $25 billion of this attributed to rising memory and AI component costs, costs that are rising because the same hyperscalers are bidding for the same components at the same time. Increased demand, low supply, increasing costs.

Capital expenditure itself has reached levels until recently unthinkable: capex as a share of sales runs near 86% at Oracle, 54% at Meta, 47% at Microsoft, 46% at Alphabet, and 25% at Amazon. Amazon’s free cash flow went negative in their Q2 filings, and Alphabet’s followed. The AI build-out can no longer be funded from operating cash flow, and not from AI’s own cash flow alone: it now consumes more than the whole company’s cash flow across all product lines leaves after existing expenses.⁷⁹

Microsoft’s AI revenue run-rate is roughly $37 billion, approximately 70% of it, on one analyst’s inference, supplied by a single unprofitable customer, OpenAI.⁸⁰ Amazon’s is roughly $15 billion at an 80% concentration, meaning 80% of it is one client, Anthropic, itself financed by Amazon’s own equity investment.

On the variable-capital side, the labor the tools are supposed to make more productive, the adoption of AI is sold as a productivity miracle. In reality AI is in wide use and thin use, adopted by two firms in three and used by the executives who answer for them about 90 minutes a week, and the productivity gain is vaporous, 0.29% over three years across all the firms surveyed on their executives’ own count, with 56% of CEOs telling PwC they saw neither a revenue nor a cost benefit from it in the past year.⁸¹ A provincial audit tested the tools where the gain was promised: Ontario’s auditor general reviewed the province’s own tests, two simulated consultations run through the 20 scribe vendors it had approved for its physicians, and found every one’s notes inaccurate in at least one, nine inventing referrals or tests, 12 recording the wrong drug, 17 missing the mental-health details, with accuracy weighted 4% of the score that approved them.⁸²

Marx anticipated this dynamic in Volume III of Capital, chapters 13 through 15.

Imagine a machine as a battery for labor. Building it took 1,000 hours of human labor (the number is for demonstration only). Those 1,000 hours are stored in the machine the way charge is stored in a battery. When a worker uses the machine, it wears down just as a battery discharges, and the wearing down releases the stored labor into the products it makes. By the time it has worn out, all 1,000 hours have passed into the products. The machine added nothing of its own: it merely released what was already in it. What it did for the worker was multiply output: more product per hour, with fewer hands at the point of production. The multiplier is in output, never in value. The labor-hour adds the value it always added, spread across more product created.

A battery does not charge itself. The charge came from labor, and the discharge needs labor: no machine builds itself, and none runs so far without some form of a human hand on it. That is why living labor is what generates new value. Labor-power reproduces itself through the wage: the worker eats, sleeps, raises the next worker, and returns to the point of production. No machine has done that yet.

The worker adds value through the labor itself, assembly, discernment, expertise, above the cost of the inputs: the machine’s transferred hours, the raw materials, and the wage. That surplus, the ΔM, is where profit lives, and in Marx’s framework it can come from nowhere else, because only living labor produces value beyond its own cost of reproduction. Marx allows two ways a machine raises the value produced in a clock hour, and both run through the worker: intensity, a faster pace that packs more labor into the hour, and complexity, labor the machine requires to be more skilled, which counts as multiplied simple labor. The machine’s own contribution stays what it was, the transfer of its own cost.

Profit and surplus value both name what is left after costs, and they are kept in two ledgers. Surplus value is kept in hours, at the factory door: of an eight-hour day, three hours produce the value of the wage and five are unpaid, and the five exist in the product the moment the shift ends whether or not the product ever sells. Profit is kept in dollars, after the sale, measured against everything the owner advanced, the machine included, and shared with the banker, the landlord, and the state before competition redistributes what remains across firms in proportion to their capital rather than their labor. The five hours are the source. Profit is the source after it has been sold, measured against the machine, and shared out, and each of those steps can fail or be faked. Surplus value that is never sold is the realization failure. Profit with no surplus value behind it is the fictitious circuit. This work is organized around those two gaps, and the owner’s ledger is built so that neither shows.

The objection follows on schedule: pay the worker the full value of their labor and there is no profit, no business, no society. Labor has no value; it is what value is made of. What the worker sells is labor-power, a day’s capacity to work, and the wage pays its full value, the cost of keeping the worker and raising the next one. The exchange is fair on its own terms, and the surplus arises because a day of labor-power produces more than a day of it costs. And every society extracts a surplus product from the working day, what its producers make beyond what they consume. Marx listed the deductions himself, in the Critique of the Gotha Programme, before a single individual is paid: replacement of the means of production used up, expansion, a reserve against calamity, administration, schools, health, and funds for those who cannot work, all of it returning to the worker as a member of society what was deducted from the worker as an individual. The surplus is universal. Profit is the form in which one class appropriates it as private property, sets its size, and decides its use. A society without surplus is no society. A society without profit is every society before this one and, on the evidence of the cooperative down the street, several inside it.

A second objection follows: if the same labor-hour makes more product, more product is sold, and profit goes up.

Before the machine, a labor-hour makes 10 units of a product. Each finished product carries 1/10th of an hour of labor-added value. After it, the same hour makes 20, so each carries 1/20th. Twice the commodities, half the labor-added value in each, the same total value per hour, because value is socially necessary labor time and doubling productivity halves the labor time each unit needs.

Selling 20 units at 1/20th returns what selling 10 at 1/10th did. The objection is true in exactly one situation, and briefly: the first adopter. While competitors still make 10 an hour, the price stays at 1/10th, and the first adopter sells 20 at 1/10th. This is twice the money for the same hour. That is extra surplus value, real money in their hands and no new value anywhere, a gap between their cost and the social price. This is exactly what closes the moment the machine generalizes and competition pulls the price down to 1/20th universally.

The capitalist experiences this as “the machine making more profit.” The belief is correct for them and false for the system, which is why every capitalist is forced to chase an advantage that vanishes once all of them hold it.

Offshoring was the route abroad. The system’s own books have one route at home to more surplus value per hour, and it runs through the worker. Labor-power is a commodity like any other, and its value is the labor time it takes to produce what the worker needs to show up tomorrow and do more labor: food, rent, clothes, transport, the next worker raised, a bundle whose size is set by history and by struggle. When those goods get cheaper to make, the worker gets cheaper to keep. Their day does not get shorter: the part of their day required to pay for keeping them going does. If it once took four of their eight hours to produce the value of their own subsistence, and the basket of goods in that subsistence now takes 1/2 the labor to make, then two hours cover them and six go to the owner. The worker’s standard can stay exactly where it was. What changed is the labor time inside it, and the owner keeps the difference. Marx calls it “relative surplus value.” It is the one route that raises surplus value per hour on the system’s own books, it runs through the worker, and it comes from the worker’s share of their own hour falling. More units sold has nothing to do with it.

For 40 years the system found that route abroad, the route the liquidation section recorded. Cheaper labor-power in the periphery lowered variable capital directly, and the cheaper goods shipped home lowered the value of labor-power at home in the way just described, which is relative surplus value delivered by geography. Offshoring also, like AI, disciplined the existing labor force through the threat of displacement by offshoring and automation, increasing the industrial reserve army of unemployed.

This held up the capitalist rate of profit, and it is the same move that hollowed the US productive base the AI build-out now runs on domestically. The offshoring route is now exhausted: the periphery’s wages rose, the manufacturing base cannot be brought back at similar costs on any timeline the rate of profit can wait for, and there is no further “outside” to move to. The AI build-out is the attempt to do at home, by replacing labor, what offshoring did by relocating it, and it arrives after the geographic route has been spent.

That is the mechanism of the “cost-of-living crisis,” read from the worker’s side. The route that cheapened the bundle for 40 years has closed, and the bundle is getting dearer: rising energy costs are in everything and the wars make it worse, and the goods that were cheap because they were made abroad now carry a tariff and rising wage demands abroad. What it costs to “keep” the worker is rising, whether the rise is labor time or the price of goods alone, and the wage has not followed for a long time. Wage growth below inflation is the wage being held under the value of labor-power, which is the counteracting tendency Marx listed and set aside because it cannot run for long: a worker paid less than it costs to keep them is kept going on credit until they are not kept at all. What it costs to keep them has itself risen, because the value of labor-power carries what Marx called a historical and moral element, and the history has changed what a worker must have to sell labor-power at all: a car where work is dispersed and transit is not, a phone and a connection where the job is found, scheduled and paid through them, insurance where care is priced privately, childcare where a household needs two wages, and on the platforms the car, the phone and the fuel themselves, the worker now supplying means of production the employer once owned. The average American family spent 80% of its budget on food, clothing and shelter in 1901 and half in 2002, which reads as liberation until the modern necessities are counted: in 2024 housing, transportation, food and health care took 71% of the average household’s spending, and the lowest-income fifth spent more than four fifths of its budget on basic needs, the share of 1901.⁸³ And the build-out has made the machine itself dearer: the chip seller’s own filing put slower consumer PC sales down to elevated memory and system prices.⁸⁴

The core’s workers are not a bought class waiting to be paid again. Their needs rose with the productive forces, the car, the connection, the care, and Marx’s point in Wage Labour and Capital holds: wants are social and measured against what the society can produce, so the little house becomes a hut when a palace rises beside it. It is the blocking of those needs by relations that cannot meet them, not the loss of an imperial dividend, that makes the core’s politics combustible.⁸⁵ Plekhanov answered the reformists of his day with the history of the class that came before the proletariat: the reforms the bourgeoisie won, far from blunting the contradiction between its aspirations and the old order, gave a fresh impetus to the growth of its forces and sharpened the contradiction until it was no longer a matter of reform but of revolution, and the victories that produced the reforms brought revolution closer, provoking reaction in the conservatives and in the innovators a thirst for new conquests. The core’s workers have lived the first half of that sequence and the reaction. Their concessions were won by the 1970s, the counteroffensive since 1980 has raised the rate of surplus value by more than half, and the needs the concessions formed have not shrunk with the wages that paid for them.⁸⁶

The AI build-out arrives as capital’s answer to the closed route, and the answer is incomplete on its own terms. As a chatbot, it cannot cheapen the bundle’s core directly, because it makes tokens and the core is calories, kilowatts, square feet, and hours of care, two thirds of the economy on Epoch’s count of what cannot be done at a screen. The services around that core are where its productivity was first promised, and nine surveyed firms in ten, users among them, report they cannot find it. To date, it cannot lower variable capital the way the imports did. The robot is the build-out’s claim on the core, and what has shipped so far is the worker in the headset: the wage relation, teleoperated. In September the largest social network’s agent was found placing its users’ phone calls through a call center, the callers not told.⁸⁷

Volume III formalizes the law under all of this as the rate of profit: surplus value over total capital advanced, s / (c + v), where c is constant capital (machinery, materials, infrastructure) and v is variable capital (wages). Surplus value comes only from the living labor that v buys, and that labor is the numerator’s ceiling: a workforce that lived on air would still yield no more surplus than the hours it worked.

As capitalism develops, capitalists substitute more and better machinery for labor, raising c against v. The denominator grows while the numerator’s ceiling shrinks, and the rate of profit falls unless the rate of exploitation rises fast enough to cover the gap. This is the question the last 40 years answered with offshoring: is the savings on wages enough to offset the cost of the machines that replaced them, and for how long? A reader who does not start from Marx is owed the reason he is used here, and it is practical. The mainstream theory pays each input its marginal product and has no account of why a technology that raises every adopting firm’s margin should shrink the pool those margins are paid from, and the Kaleckian account explains how the build-out’s spending becomes the chip maker’s profit this year and not what the machines do to the rate of profit once they are running. Marx’s does both, because it locates profit in the living labor a firm employs and competition in each firm’s effort to employ less of it, so that what enriches the firm that automates first starves the system that automates together, which is the build-out’s shape. The labor theory of value is contested as a theory of prices, and most economists think it superseded, but as a regularity it has held up better than its critics expected: across input-output tables in several countries, sectoral prices track the labor embodied in them closely, whether or not that is more than a matter of industry size, which is still argued.⁸⁸

The rate has been held up the way Marx said it would be, by raising exploitation faster than the machines could lower it. Corporate profits reached 19.4% of national income in 2026, the highest since the 1940s, and even on the Tax Foundation’s net measure, which deducts depreciation and puts labor’s share higher, that share at 68.3% is the lowest since 1948. Converted to Marx’s categories, the rate of surplus value has risen 57% since 1980 and 17% since the pandemic, to near its post-war peak. The counteracting factor in production is the realization problem in circulation: the share that props up the rate of profit is the share the wage no longer buys back.⁸⁹ The argument over labor’s share is a century old. Struve and Bernstein held that reform was blunting the contradiction between labor and capital, and Plekhanov answered that despite the reforms the relative share of the working class in the social income had fallen in every advanced capitalist country, and that social development proceeds through the aggravation of contradictions and not their blunting. The case made this year for a stable labor share on a net measure is the blunting thesis in the form of national accounts, and on its own series the share is the lowest since 1948.⁹⁰

Put the year’s two prints together and the recovery of the rate has two props, and each undermines the other. The first is the rate of surplus value, wages held under prices while output rises, which is the household squeeze seen from the other side. The second is the build-out’s own spending, which arrives at the seller as revenue and profit, the circuit Kalecki described in which investment returns as profit in the year it is spent, financed by the borrowing Section XVIII counts and by the seller’s support for its own buyers. The first prop shrinks the demand the second must eventually be validated by, and the second lasts only as long as the borrowing does, which is why the profits sit where they sit and why a turn in either reach takes the rate with it.⁹¹ Goldman’s own arithmetic says the same from the other side: the beneficiaries of the build-out’s spending are about half of the index’s earnings growth this year, the chip seller and one memory maker a third of it by themselves, while the productivity the spending was meant to buy adds four tenths of a point, and a $250 billion swing in the spending moves the index’s earnings growth by about six points either way. A markets editor put the mechanism in the language of the Levy profit equation on Bloomberg’s podcast: the safest sources of profit now are the build-out’s borrowing and the sovereign’s deficit, safe until they are not.⁹²

Surplus value can exceed the wage bill, and nothing in the law prevents it. On the national accounts rebuilt to separate productive from unproductive labor, the series the economists Anwar Shaikh and Ahmet Tonak began and Dimitris Paitaridis and Lefteris Tsoulfidis carried to 2007, the American rate of surplus value stood at the highest levels on record, on rising productivity and flat or falling real wages for productive workers.⁹³ A rising rate of exploitation is the counteracting tendency that has run the longest, and it has two limits. The first limit is arithmetic: surplus labor is the working day less the hours that reproduce the worker, so each further cut in those hours adds less. Cutting necessary labor from four hours to two adds two hours of surplus, and cutting it from one hour to a half adds half an hour. At the limit the whole day is surplus, and the rate of profit can be no higher than living labor over constant capital, a ceiling that falls as the build-out raises constant capital against the labor employed, whatever the wage. The second limit is realization: the more of the product is surplus, the less of it wages can buy back, and the rest must be bought by investment, by the rich, by the state’s deficit, or abroad. The same series shows where the surplus went. The general rate of profit recovered from the early 1980s on that rise, and the net rate of profit stayed far below its level of the 1960s, because the growth of unproductive activity consumed what the higher rate of exploitation produced.

Herein lies the AI and robotics promise. The first capitalist to “automate fully” wins on cost while the price is still set by rivals who pay workers: they produce below the social value and sell at it, and the difference is the extra surplus value already named, a transfer from their competitors. When everyone automates, the price falls to the new cost, the advantage disappears, and the system’s rate of profit has fallen, because living labor, the only source of s, has shrunk against c.⁹⁴ And if the robot ever DOES work, the route still does not reopen. It ends, because the robot that makes the bundle with no labor in it makes everything else the same way, and where no labor is employed no surplus value is produced at any rate of exploitation. Relative surplus value has a limit, and the limit is the abolition of the surplus it was raising, along with the buyer of the bundle. When labor becomes structurally unnecessary, the source of profit collapses, and with the wage bill the demand for the commodities collapses beside it. Capitalism’s success at automating labor is the structural mechanism by which it eliminates its own profit circuit, which is its foundational engine. The industrial arm has worked for 40 years, and the humanoid and the general agent may yet. A machine that works multiplies output and adds no value, which is the falling rate for a formation that must validate its machines in profit and a plain gain for one that validates them by plan, a difference Section XXIV takes up.

Meanwhile AI expansion is running another form of this exact “first adopter” case in public. Each frontier release is the first adopter: more output per compute-hour than anyone else has, still sold at last quarter’s price. Within months the rivals match the output, and the price per token falls, first toward the value of the compute consumed and then through it, because the competitors price the token as if its electricity were its cost and leave the chip, which is nearly all of what a token costs to make, to be paid by someone else: on one operator’s ledger the capital is 97 cents of every dollar a token costs and the power two, while the budget tiers sell at $0.06 to $0.30 per million tokens against a full cost near a dollar for the cheapest open model, and the difference is subsidized.⁹⁵ The token is selling below its value and the difference is paid by whoever decides to keep pouring money to keep the seller of AI compute solvent. Eventually though, the premium evaporates, and then the value does.

What remains is the largest constant capital ever assembled, transferring its cost into a product sold for less than that cost, against a variable capital too small to carry it on the production side and a customer base that cannot find the productivity that would pay for it on the other. A token price chart falling through the full cost of compute draws the pressure on the capital as a line. The full cost of a token is the electricity it burns and then everything that had to be built and borrowed for before it could burn: its share of the chip, of the building, the cooling and the power around the chip, of the interest on the debt that paid for all of it, and of the wear that running it adds. Above the line a sale pays for all of that. Below the line a sale pays for the electricity first and only the remainder goes to the rest, so the debt is served and the capital recovered only as fast as the remainder allows. The two levers left, more tokens and harder running, work only by that margin, and the second discharges the battery faster, the maintenance and the depreciation arriving sooner for the same money. Where the price is set at the electricity, as the rivals above price it, the remainder is nothing and no volume serves the debt. And the line is worked from below as well, since the electricity is the one cost no operator can cut and its price is rising, bid up by the build-out itself and by the war. That is the interconnection: the price forced down from above by the rivals, the floor forced up from below by the power bill, the interest due whichever way the two move, and the one lever that remains consuming the machine it pulls on. What is left to read is the depreciation schedule, which is what the analysts read next.

The AI build-out assumes you can build the largest “battery” in human history, and that it can reproduce itself, maintaining and building itself or copies of itself. Trillions of dollars of constant capital in data centers, chips, electricity generation, and cooling run the system, ideally without any workers who would, in the older arrangement, be the source of value that both recharges those machine batteries through maintenance and uses them to force-multiply their own labor. The machines do all their own necessary labor and replace themselves. That is the final capability test.

The AGI promise becomes the promise to keep accumulating constant capital infinitely, while abolishing variable capital along with the social relation through which capitalism has historically functioned.

The operating cost is the part of the arithmetic the sellers report as falling and the leaks report otherwise. On the critic Ed Zitron’s reading of a set of leaked financials, the line for inference at the largest laboratory is modest and the line for marketing is the size of a beverage conglomerate’s, with no agencies or billboards to show for it. Inference is being given away and booked as marketing, and renaming the expense does not lower it.⁹⁶ The railroad, once laid, ran trains for a century on maintenance. This railroad has to be relaid every two to three years, because a better model appears and the switching cost is zero, so the capital that was supposed to become a toll road is a construction site that never becomes one.

The depreciation schedules for AI companies give us an indication of whether or not self-replicating machines are coming tomorrow. BCA Research’s Peter Berezin calculates hyperscalers will carry over $2.5 trillion in AI assets by 2030. At typical 20% depreciation rates, the cautious case, that produces approximately $500 billion of depreciation a year by 2030, more than the $390 billion the four together earned in all of 2025. That sets the wager against a completed year’s profit, and the revenue that would have to cover it lies in 2030, which is the problem: the capital is being spent now, the debt service falls due on the lenders’ calendar, the chips lose their economic life inside 36 months on the bearish estimate, and the war has raised the price of waiting.⁹⁷ Michael Burry’s parallel analysis projects that normalization to realistic 2-to-3-year economic asset lives would cut Microsoft and Alphabet 2026-2028 EPS by 15% to 25%, with AWS margins facing 500 to 800 basis points of pressure, five to eight percentage points. The accumulated depreciating constant capital is larger than the surplus value that can be extracted through it under current terms: to carry it, the surplus the machines produce would have to cover a depreciation charge growing toward $500 billion a year, on top of the running costs, before any return on the capital, and that growth is the wager. The asset depreciation budget line is reading a Marxian contradiction back to an analyst class that refuses to understand. The schedules are now being lengthened from the other end. In July Microsoft extended the accounting life of its data centers from 15 years to 25, and because the longer life moves future leases from finance leases, which count as capital spending, to operating leases, which do not, its 2026 capital-spending guidance fell from about $190 billion to about $175 billion with the building plan unchanged: the build-out moving off the line the market reads while the spending holds.⁹⁸

Capital accumulation growth continues only if constant capital can be combined with variable capital in production to extract more surplus value, and the AGI buildout’s end purpose is to eliminate the variable capital today required to validate the constant capital being deployed.

Anthropic does generate real third-party enterprise revenue. More than 1,000 enterprise customers spend over $1 million dollars annually on Claude, a count that doubled in under two months this spring. Eight of the Fortune 10 companies are customers. Claude Code generated over $2.5 billion in annualized revenue within nine months of launch. The dilemma is structural. A run rate is a marketing construct, an undisclosed period multiplied to a year, and for a business billing by the token the period is whichever weeks the seller chooses. Recognized revenue is the only number that survives an audit. Even at the $45 to $50 billion annualized run rate of the May round, the revenue about equals what the company has itself committed to pay for compute, $200 billion to Google over five years and $100 billion to Amazon over 10, some $50 billion a year if spread evenly, before any return on the capital that compute runs on. A pace reported above $100 billion in September, on one chain of reporting, would double that numerator against the same commitments, and a run rate is still not a return.⁹⁹

Against the $1.1 trillion the four hyperscalers laid down from 2023 through June 2026, with $725 billion guided for this year alone, no filing attributes any earnings to the AI assets themselves, and nothing the filings show, revenue and growth and a run rate, is a return. At best the return is unread, at worst it is a shunt: a closed loop the economy’s heart now pumps blood into that returns to the heart without passing through any tissue that needs it. Our metaphorical heart monitor reads a strong pulse: 75% of the first quarter’s growth and a third of the index, while the body the pulse is meant to feed receives no benefit.¹⁰⁰