How the Machines Were Paid For

1. Why this is worth a few pages.

The novel spends a chapter on a retired machinist and a carburetor. It does so because his pension turns out to hold a thin slice of the loans that built the buildings where the artificial-intelligence industry keeps its machines. Readers have asked whether that chain is real. It is. This note lays out the real structure as it stood in 2026, with the figures the companies and their lenders disclosed, so that a reader can judge for themselves how much the book compressed and how much it simply reported.

The short version is that the machines are not owned by the companies whose names are on the buildings, the buildings are not owned by the companies whose machines are in them, and the money that paid for both is owned by people who have never heard of any of them.

2. The players.

There are five parties in any large data-center deal, and most of the confusion comes from assuming there are two.

3. The building: leases that have not started yet.

The clearest measure of how far the hyperscalers have moved from building to renting is a line in their own filings called leases not yet commenced. These are signed commitments for buildings that are not finished, or not started, and so are not yet on the balance sheet as liabilities.

As of June 30, 2026, Microsoft reported $329.1 billion of such leases, with start dates scheduled from fiscal 2027 through fiscal 2033. A year earlier the figure had been $92.7 billion. Against that pipeline the company carried about $88.5 billion in recognized lease liabilities. Across the five largest companies, a Reuters analysis of LSEG data put the not-yet-started total at roughly $1.09 trillion, about four times what appears on their books as recorded lease obligations. One of the five added another $68 billion in a single month that summer.

Microsoft is the largest renter, not the only one. Meta carried about $279 billion of such leases and Oracle about $260 billion, the latter nearly seven times what it had recognized. Alphabet reported $85.2 billion, a smaller number that hides a different structure. Google increasingly has other companies build and own the buildings that hold its own chips, and then guarantees those companies' leases so they can borrow cheaply. Its maximum exposure under such guarantees reached $43.8 billion by mid-2026, up from $6.5 billion nine months earlier, with another $24.1 billion awaiting final terms. Of all that, about $815 million sat on the balance sheet as a liability. The landlords in these deals are often former cryptocurrency miners with power contracts and empty halls. In one arrangement a cloud operator leases the buildings, fills them with Google's chips, rents the computing to an AI lab, and Google stands behind the lease. The tenant, the landlord, the chip owner, and the guarantor are four different companies, and the building is in a town that has never heard of three of them.

The filings do not name the landlords. What they do say, in the careful language of risk factors, is that the arrangements depend on third parties for financing, utilities, and permitting, and that if demand does not arrive the companies could be left paying for long-lived capacity that is hard to shed. Read from the other side, the same sentence describes the landlord's position. The landlord has borrowed to build on the strength of a tenant's signature, and the tenant's signature is the only thing standing between the landlord and its lenders.

4. The machines: a company created to own the chips.

The chips are financed separately from the building, and in a structure that most people outside the industry have never seen. The operator creates a special-purpose vehicle, a company with no employees whose only asset is the chips. It sells the chips to that company outright, a true sale, so that if the operator fails, the chips are not part of its estate. The special-purpose company then borrows against the chips, and the lender takes a first claim on the chips and on the operator's contract with its hyperscaler customer, which is assigned to the lender as well.

That last part is the heart of it. The lender is not really lending against hardware. Hardware depreciates, and how fast is contested, with estimates running anywhere from two to six years. The lender is lending against the hyperscaler's promise to keep paying for the hardware's output. A loan is sized so that it amortizes inside the contract, on pessimistic assumptions about how long the chips stay useful. Lenders underwrite the customer's credit, not the operator's.

The scale is no longer small. One neocloud closed an $8.5 billion chip-backed facility on March 31, 2026, the first such financing to receive an investment-grade rating, priced at a modest spread over the overnight rate and maturing in 2032. The same company's total debt was reported in the tens of billions. Individual hyperscaler commitments to neoclouds were disclosed at $23 billion, $19.4 billion, and $9.7 billion, each covering tens or hundreds of thousands of chips through the end of the decade. The chipmaker itself entered the lending business, backstopping as much as $105 billion of data-center financing for a single customer in August 2026, which is the circularity the trade press began to worry about out loud.

Meanwhile the price of renting a top-tier chip by the hour fell from around eight dollars in 2023 and 2024 to between three and three-fifty by late 2025, a decline of seventy to ninety percent depending on how you count. Any operator whose revenue was not locked into a contract was, by then, earning less on its chips than it owed on them.

5. The paper: how the loans became a pension.

A loan against one building, or one company's chips, is an illiquid thing that one lender holds. It becomes a market when enough loans are pooled and sliced. Originators gather the loans, group them by risk, and sell tranches to asset managers and pension funds. The safest tranche gets the highest grade and the lowest yield. The riskiest gets the opposite.

By 2025 the first data-center securitization carrying two agency ratings was sold, at $1.1 billion. By 2026 data centers made up 61 percent of a $79 billion market in digital-infrastructure securitizations. The largest single projects were financed off the hyperscalers' balance sheets through the same special-purpose technique, including a $30 billion structure for one campus and a $38 billion syndicated loan for two others.

The point of the structure is that the debt does not appear on the famous company's books. The company looks less leveraged than the system it sits on top of. The leverage has not gone anywhere. It has moved to the landlord, the operator, the special-purpose company, and from there into the tranches, and from the tranches into the boring accounts of people who were told they held boring things.

6. Who owns what.

7. The cascade, and the one thing that does not cascade.

The lawyers who structure these deals describe the failure mode plainly. A developer default triggers the lender's enforcement and a change of control over the building. A neocloud default breaks the lease payments to the data-center vehicle, which undermines the cash flows behind the securitized tranches. Cross-default clauses in the loan agreements mean that distress at one node can fire several obligations at once, so that what looks like one company's trouble becomes a system-wide enforcement event.

There is one thing the structure is designed to protect, and it is worth understanding because it is counterintuitive. The machines keep running. A landlord can default, its lenders can take the building, and the tenant's operations inside do not stop, because the tenant's contract is the collateral and the last thing a lender wants is for the collateral to stop paying. The engineer on the night shift inside that building would not learn that the walls had changed hands from anything on his screens. He would learn it from the news, if at all.

That is the asymmetry the novel is built around. The money under the machines is fragile and the machines are not, and the people whose savings sit at the bottom of the structure are the only ones in it who cannot see the top.

8. What the novel compressed.

The book gives these parties fictional names and folds some of them together. The landlord in the story is a composite of a wholesale developer and a neocloud: it owns buildings in several states, it leases one of them to the cloud company that runs the digital dollar's Iowa region, and it owns the chips inside through a company it created for the purpose, so that a lender could own the company. Its loans are in the pile, and its default is the first fact in a pile of opinions. None of that is invented structure. Only the names are.

The cloud company that is the landlord's tenant is never named in the book. That is deliberate. The real companies that run regions inside other people's buildings have done nothing in this story but rent space and run machines, which is what the structure asks of them. The failure in the book belongs to no one, and a real name would have suggested otherwise.

Sources.

• Microsoft Corporation, annual report for the fiscal year ended June 30, 2026, as reported in "Microsoft's $329.1bn lease stack is capacity that has not started" (BTW Media) and "Microsoft Reports Over $130 Billion in New Data Center Leases" (Bloomberg, July 29, 2026).
• Reuters, "AI data-centre race builds $1 trillion lease burden for Big Tech", LSEG analysis of five companies' filings.
• Bloomberg, "Google Piles On Off-Balance-Sheet Risk to Fuel Its AI Chip Push" (August 2026), on Alphabet's uncommenced leases, lease guarantees, and the Fluidstack, TeraWulf, Cipher, and Hut 8 arrangements.
• Fortune, "Moody's flags $662 billion risk at the heart of the data center build-out by just 5 companies" (February 25, 2026).
• Introl, "Microsoft's $60B Neocloud Bet", on disclosed hyperscaler commitments to Nscale, Nebius, Iren, Lambda, and CoreWeave.
• Peony, "Neocloud Capital Raise (2026): SPV Off-Take Structures and the GPU-Backed Capital Stack", on true-sale SPVs, offtake assignment, amortization against depreciation, and the CoreWeave investment-grade facility.
• Quinn Emanuel, "Emerging Litigation Risks in Financing AI Data Centers Boom", on project finance, off-balance-sheet SPVs, securitization share, chip rental-rate declines, and cascading default.
• Capacity, "CoreWeave's debt hits $35bn".
• Data Center Dynamics, "Microsoft acquires 111 acres in West Des Moines, Iowa, for data center campus", and Wisconsin Public Radio, "Microsoft built 5 data center campuses in this Iowa city", on the real Iowa footprint that the novel's unnamed region borrows its geography from.

Figures are as the sources reported them and will have moved by the time you read this. None of the companies named in the sources appears in the novel, and nothing in the novel should be read as describing any of them.