Section XIII: XIIIThe Shunt: The Closed Circuit and Its Claims

The buildout’s demand generation driving such growth is largely its own capital coming back to it, the shunt, stated in accounts. The Google->Anthropic->Broadcom->Google loop shows the mechanism. Google funds Broadcom to develop its tensor processing units (TPUs), which Broadcom delivers to Google, who installs them in Google data centers. Anthropic, itself partially funded by Google’s $43 billion in equity investments, rents TPU capacity in Google data centers with Google money. Anthropic’s spending, financed in part by Google itself, is then booked as Google Cloud revenue growth. The equity is financing on the laboratory’s books and the spend is revenue on the cloud’s when it is recognized, which is the shunt in the accountant’s own terms. This is Google’s own investment coming back to Google, with the payment recognized as “third-party revenue” even while the equity stake in the other companies is simultaneously recognized as a Google asset.¹⁰¹

The Information’s documentation shows that 50% of all remaining performance obligations at Microsoft, Amazon, and Google are owed between them by OpenAI and Anthropic, two unprofitable companies funded by the hyperscalers themselves. A closed circuit was created in which the buyer supplies the money the seller receives: investment dollars from other profitable lines are put into a loop that loses money on a future promise, and the loss is booked as growth. The chips are real, the labor in the fabs is real, and the demand for them is the capitalist paying himself, which is why the loop is self-eating before it is financial: a productive circuit run on the buyer’s own investment rather than on a buyer, a configuration Section XV names. Finance is what the self-eating circuit builds on top, and the close of this section names what the two are together.

The central bankers’ own bank has described the financing behind it. The Bank for International Settlements finds the build-out moving from operating cash flow to debt: private credit funds lent more than $40 billion to AI companies in 2025 against $3 billion in 2010 and hold over $200 billion of such loans, hyperscaler bond issuance passed $100 billion in 2025, and leverage is being moved off the balance sheet into vehicles backed by bank funding lines. It names the circular financing inside the AI ecosystem among its concerns, rates the risk to financial stability as moderate at the date it wrote, and makes the boom’s survival depend on the AI firms meeting the earnings expected of them, observing that lenders price these loans like any other while equity prices imply outsized returns, so that one side of the market may have the risk wrong.¹⁰²

The TPUs themselves are real commodities made at Broadcom and TSMC. Workers labor in chip fabs to manufacture the units Google buys, and that is M-C-M’ in the classical sense, with real surplus value extracted from real labor: expensive chips sold to companies, produced by laborers and machines that cost less on an ongoing basis than the value received from selling the chips they produce by private owners to private owners. But Volume III catches structural problems even at such an upstream level.

The organic composition of capital is at an extreme here: a TSMC advanced-node fab runs over $20 billion in constant capital inputs against ongoing variable capital that is a fraction of that, and TSMC’s visible 40 to 50% operating margins are also sustained by monopoly positioning led by technological moats at the leading chip nodes, rather than by the underlying labor exploitation dynamics. Monopoly rents are masking what would otherwise be an equally compressed rate of profit, and China’s rush to create their own version of leading chip manufacturing node technology (just like the frontier AI, just like memory chips) to break the monopoly positioning and therefore the price premium is the pressure this work expects on the rate of profit for chip manufacturing, and it would show in TSMC’s gross margin and its leading-node wafer price by 2028. Simply: the leading-node technology monopoly is generating a price premium the fab’s labor itself would not be, and will not be generating under rapidly arriving competitive pricing.¹⁰³

Such upstream margins are coupled to the loop: a substantial portion of TSMC’s leading-node demand runs through the same hyperscaler capital flowing through the loop described above: the increased demand with limited supply causing the price increases and supply expansion. The loop also bids up the price of its own means of production: in August the chip seller’s largest customers were told that servers shipped early next year would cost more than 15% more in many cases, memory prices having soared on the build-out’s own orders, so each dollar of capacity costs more than the last and needs more borrowing, the rising cost of constant capital in its most literal form.¹⁰⁴ The seller’s finance chief said as much on the same call: the memory shortage is largely a result of the build-out itself, one symptom of the same surge that drives its growth, and its cost took the seller’s own gross margin from 75% toward a trough of 71 to 72%. A new memory fab takes three years or more to build, so the scarcity the build-out creates outlasts the orders that created it.¹⁰⁵ The ongoing realization of the chip’s value at the Broadcom-Google sale, driving Nvidia and Broadcom and TSMC valuations and profitability, depends on demand for chips OR for AI compute from outside a closed loop that is itself unprofitable, a demand the loop’s own accounts do not yet show as a return on the capital. If the closed loop stops buying chips because it is no longer “profitable” to do so via circular financing, then demand for chips goes down, supply goes up, and price therefore drops along with profits.

The loop shows in the buyers’ own accounts. On Street estimates three-quarters of the AI revenue at the two largest clouds is two unprofitable labs, and the second-quarter reports show what carrying them costs the cash line. The reports delivered revenue growth, Microsoft and Amazon near 20%, Meta near 30%, and Meta’s free cash flow fell 91%, with Amazon’s and Alphabet’s negative as Section XII recorded. The buildout’s revenue is substantially the buildout’s own spending returning through its two principal tenants.

In the valuation scholar Aswath Damodaran’s formulation the same week, a healthy business runs its expenditures at 10-20% of revenues, not 80%. The forward estimates have institutionalized the same structure: the Street’s 2027 model attributes 48% of Google Cloud’s entire segment revenue, not its AI revenue, the whole segment, to the two frontier AI laboratories, more than $100 billion of projected sales contingent on two unprofitable companies continuing to spend money given to them by the patrons.

The circularity now runs through the forward estimates. The income statements have begun to consist of belief directly: $53.4 billion of Amazon’s $80.9 billion in pre-tax income last quarter, in the 10-Q for the quarter, roughly 2/3rds, was an unrealized mark on its Anthropic stake, and Alphabet’s $98 billion of other income, unrealized gains on its equity securities for the most part with a SpaceX position carried near $94 billion against a $1 billion 2015 cost and most of it still locked up, was 71% of its $139 billion of pretax income and turned $41 billion of operating income into $112 billion of net income: fictitious-capital appreciation, booked as earnings. Take the two companies out of the S&P 500 and its 47% second-quarter earnings growth falls to 29% on FactSet’s July 31 count.¹⁰⁶ The comparison measures their weight and not the marks alone, but the weight is carried by income no customer paid: reported profit can rise when an investment is marked up, with no sale made.

The year’s profits came from the two reaches this work names. Domestic corporate profits rose $738 billion in the year to mid-2026, 17.7%, to $4.4 trillion, and outside finance, which added a quarter on the stock market’s rise, nearly two thirds of the increase came from two sectors: technology, the build-out’s, with 39%, and energy, riding the war’s prices, with 24%. The profit boom is the war and the machine, and a turn in either takes the rate of profit with it.¹⁰⁷ The build-out is not paid for by its own fruits. It is paid for by the rents of the last technology, search and the advertising that sells to the consumer, which is why the one builder without such a rent is the one whose default swaps the market now prices hardest, and why the consumer the build-out squeezes is the consumer whose attention pays for it.¹⁰⁸

The loop can now be read in a single laboratory’s accounts. Anthropic’s prospectus shows 47% of its 2025 revenue sold through the four cloud companies that also sell it its compute, nearly a quarter from two customers, and about $518 billion of compute and infrastructure commitments, substantially non-cancelable, against $4.6 billion of revenue, more than a hundred times the year’s sales: the other side of the leases the builders have signed and not yet begun. Its 2026 is reported to be far larger, second-quarter revenue above $11.5 billion with a positive adjusted operating income, and the dispute now is over what the adjustment leaves out.¹⁰⁹

Half a dozen investors have now confirmed to the Financial Times a $2 trillion October target for the Anthropic public listing, more than double the May round. At that print, the stakes held by Amazon and Google revalue again, the revaluation books as their income under the same standard, the income compresses the published multiples, and the multiples justify the next round of the capex whose spending returns as the laboratory’s revenue.

Amazon’s own filings show the structure at the scale of one relationship: a single Anthropic relationship, anchored by a 10-year, $100 billion commitment to run on Amazon’s chips, appears in their quarterly disclosures as AWS revenue, as part of the AI run-rate, as part of the chips run-rate, and, separately, as the $53.4 billion equity gain. The same dollar, several labels, the two run-rates nested inside the one revenue line and the gain marked on the stake that funded it.¹¹⁰

The neocloud layer books the circuit’s dollars a third time. CoreWeave lost 626 million dollars in the second quarter while growing revenue 112%, carried about 35 billion dollars of debt at the end of June, leases aside, and lists as its customers Nvidia, OpenAI, Microsoft and Google on OpenAI’s behalf, Anthropic, and Meta, which is the buildout’s own spending arriving as a customer list.¹¹¹ Its operating business ran near breakeven, and the quarter’s loss was the interest on the debt that built the machines, $640 million against a loss of $626 million. S&P revised its outlook on the company to positive in April while affirming its B+ rating, on the strength of the capacity and contracts it had added and a deepening relationship with Nvidia, with the weaknesses in its accounting controls still an open condition. The rating room has a precedent. In the autumn of 2001 S&P’s Enron analyst went on television with the company already on credit watch to say its ability to keep an investment-grade rating was excellent in the long term and that the off-balance-sheet partnerships were in the past. In March of that year Fortune’s reporter had asked how the company made its money, and the analysts she called had answered that the earnings were a black box and the company one you simply had to trust, while the agencies held it at investment grade until four days before the bankruptcy.¹¹² The rating is a room like the others, and it reproduces the form.

The operators have begun conceding the demand curve in their choice of pitch: the executives who narrated compute scarcity in January now advertise the premium available for leasing infrastructure to others, which shows operators preferring a landlord’s rent to their own end market. None of the majors has taken a write-off, because the write-off would be a public admission that could break a belief, and Section XIX shows the means by which it is postponed. None of these companies has articulated the business model the spending is for, and when a company says nothing, markets fill in the vacuum, increasingly, with the suspicion that there may be nothing to say.

The vacuum is where the credit system enters. Marx’s Capital chapters 25 through 32 in Volume III address what happens when the falling rate of profit closes productive outlets. Capital seeks the credit system as an alternative outlet, producing what Marx called “fictitious capital,” running instead on the formula M-M’, a circuit of money expanding through financial engineering to more money without passing through a commodity production stage.

Marx’s chapter 24 names this directly: M-M’ is “the meaningless form of capital,” in which “capital appears as a mysterious and self-creating source of interest, of its own increase.” One example: a bond is a trade-able claim on future revenue streams. A share is a trade-able claim on future company dividends. These claims circulate in a market, have prices, are borrowed against, appear to grow autonomously through the credit system, and they are real enough to the holder of this asset. They are “fictitious” because their value rests on actual future commodity production and surplus value extraction/profit generation occurring somewhere in the “real productive economy” that can validate the claim’s own valuations. When the validation of future valuation fails, the fictitious capital is itself revealed exactly as fictitious speculation and is then quickly destroyed because it contains no “actual value.”

The buildout’s fictitious capital sits in three layers. The first layer is revenue recognized before cash. Nvidia’s quarter reported August 26 booked $96.2 billion of revenue and $59.7 billion of net income against $24.1 billion of operating cash flow, in part because receivables rose $22.3 billion in the quarter, on terms the filing puts at 90 days to a year for certain investment-grade customers building data centers. The sale is booked but unpaid, the money itself is merely a future claim, and the claim is an asset. The critics have the shape without the ledger: six companies losing money on AI and a seventh they lose it to, together a third of the index, with the seventh lending them the money they lose to it.¹¹³ It is the self-eating circuit described from outside the trade.

The second layer is future revenue held as backlog. Half of the remaining performance obligations at Microsoft, Amazon, and Google are owed by the two laboratories, as above, and the laboratories carry the other side of the same contracts as commitments: Anthropic’s signed this year exceed $135 billion. Nvidia’s own supply commitments doubled in a quarter to $279 billion, in the August 10-Q, which is the chipmaker promising its suppliers purchases against orders it has not been paid for. Each is a claim on revenue that does not exist yet, an asset or collateral on one book and a commitment on another.

The third layer is claims written on other claims and borrowed against. Nvidia’s off-balance-sheet guarantees reached $108.5 billion, most of it the SB Energy arrangement for OpenAI backed by Apollo, BlackRock, and Goldman. Each layer’s value rests on the layer below it, and the bottom layer is the demand the loop above supplies to itself. That is Capital Volume III’s mysterious self-creating source, three stories high, and it is what Chapter 25 says capital builds at this point.

The seller now says it aloud. On the August call its finance chief acknowledged the scale of the support it gives the laboratories and builders that buy its chips, the guarantee of up to $105 billion behind the Ohio lease among it, and said: “We know some will call this circular financing. We see it differently.” The laboratories receiving its support are expected to account for about a quarter of its business next year.¹¹⁴

The seller’s buyback shows where its rent goes and what it leaves behind. The unpaid receivables above halved its operating cash flow in the quarter to July, and five customers owed about seven tenths of all it was owed. In the same quarter it returned $26 billion to its shareholders, more than its free cash flow, and issued $25 billion of notes, and at the end of September it raised its repurchase authorization by the largest amount in history, to $235 billion. It is not an exit. The sales are real, the margin is 75% and the commitments to its suppliers stand at $279 billion. But it is a cash-out in the exact sense its accounts show: the holders are paid in cash, partly borrowed, while the seller takes its customers’ promises and its own guarantees in exchange, and its senior officers’ bonuses this year are tied to a single measure, revenue, with none for cash or return. Lenin’s chapter on parasitism named the stratum this serves, those who live by clipping coupons and take no part in any enterprise.¹¹⁵

The rent the build-out pays for its inputs is not going back into capacity. The memory maker’s revenue rose in its fiscal year to September from $37 billion to $133 billion while its cost of goods rose by about a seventh, its gross margin reached 87% in the last quarter, and of the $90 billion its operations produced it spent $27 billion on plant and put $35 billion into securities, while its customers paid it $12.7 billion in deposits to be sure of supply. It expects to spend no more than about $50 billion a year on expansion, out of its own caution about the memory cycle, and the only thing holding back a buyback on the chip seller’s scale is the condition attached to its public subsidy, which lapses on December 9. In Korea the two memory makers came to be more than half the stock market by June, and without them the index was falling. The scarcity is being harvested, not relieved, which is what a rent does, and the capacity that would end it is the successor’s.¹¹⁶

The incumbent now stands where Spain stood. It commands the revenue, the reserve currency’s seigniorage, the chip seller’s buybacks, the memory maker’s scarcity, the sovereign’s first claim on the world’s savings, while the organization of production has moved to the successor, which makes close to a third of the world’s manufactures. The first transition’s lesson, recorded in its own place above, was that command loses to organization, and the rents this round records are command’s form. Plekhanov’s sentence about the rising bourgeoisie applies to the successor: its strength lies not so much in its wealth as in the progress of which it is the vehicle. The incumbent’s wealth, a stock market valued at a record multiple of its output, measures its titles and not its forces.¹¹⁷

Anthropic’s August 31 agreement with Lambda is the structure in one transaction. Anthropic, the end customer, contracts for $35 billion of compute from Lambda over six years. Lambda, a cloud provider Nvidia has invested in, will deliver it on Nvidia chips it has bought, in a campus it does not own. The campus belongs to Hut 8, a data-center developer, and the tenant on Hut 8’s lease is Nvidia, which Lambda pays for access to the space. Hut 8 financed the build with $4.25 billion of notes due 2042, written against Nvidia’s lease. Lambda financed its deployment with a $926 million loan written against an investment-grade offtake, and the investment-grade party is Nvidia, because Anthropic, with $135 billion of compute commitments this year, is the one name in the chain that could not have signed such a lease. Four parties, and each layer’s financing rests on the layer beneath it. At the bottom of all of them is one balance sheet, whose own revenue is the chips the chain exists to buy.¹¹⁸

The closest topological financial precedent is the Goldman Sachs Trading Corporation (GSTC) pyramid of 1928 to 1932. GSTC issued at $104 in December 1928, reached $222.50 within two months on its own purchases of its own shares, and fell to $1.75 by 1932, in Galbraith’s account.¹¹⁹ A 99% wipeout in three years. GSTC raised money to launch Shenandoah Corporation, which itself raised more money to launch Blue Ridge Corporation, with each layer charging fees, borrowing against company assets, and holding shares of the layer below it. When asset values fell, leverage worked in reverse and the entire pyramid was wiped out.

GSTC held claims on enterprises that already existed and claims on itself, layered and leveraged. GSTC financed no new capacity and promised nothing but appreciation, and the utilities and industrials in its portfolio produced independently of the pyramid’s claims on them, before the pyramid was built and after it was destroyed. The hyperscaler pyramid is the same structure with a productive promise added, and the addition cuts the wrong way: the promise is what makes the claims sale-able, and the constant capital the claims are written on depreciates whether or not the promise is kept.

The pyramid has a season. This is the arc’s fourth autumn by the pattern section’s count, and it contains an anomaly the schema itself exposes. Autumn capital flees rigidity for liquidity, and the terminal act of this expansion is the purchase of maximal rigidity, 20-year power contracts, use-specific silicon, poured concrete. The anomaly resolves through the numbers above: AI revenue at the two largest clouds three-quarters circular, income substantially the mark-up of the counterparty, financing off the balance sheet. The buildout is the financial expansion of an autumn presented as a material one, M-M’ in the form of M-C-M’, with the self-eating circuit of the opening paragraphs as the material it points at, and the record says why simulation was the only move available. Every new material expansion in the sequence was led by a new bloc positioned to organize accumulation at greater scale. The incumbent’s autumn role is financier of the successor, never author of its own rebirth. The hyperscalers are attempting the one thing the 500-year pattern contains no instance of, and the circularity is what the attempt necessarily looks like from inside.