The Future Already Leased

Part two of a diptych on how China and America finance capacity nobody is paid to count, and who ends up holding it. Part one is The Price of Abundance.

On 10 September 2025 the richest man alive was, for part of a morning, an eighty-one-year-old who had spent his working life selling database software. Oracle rose as much as 43 percent that Wednesday and closed up 36 percent, its steepest day since 1992. Larry Ellison's fortune grew by $101 billion, the largest single-day increase Bloomberg had recorded.

He had sold nothing. A newspaper had reported that OpenAI would pay Oracle something like $300 billion over five years beginning in 2027. A company then booking around $13 billion of annual revenue had undertaken to pay out roughly the annual economic output of Finland, from money it did not have, for computing capacity that did not exist in buildings nobody had finished.

The market believed the report for two days. By Friday Ellison was $34 billion poorer, although nothing about the reported agreement had changed. Over the following months, as Oracle's financing burden became clearer, the shares fell about 64 percent from their peak. S&P lowered the company's rating to BBB– on 9 July, one notch above junk, while its remaining performance obligations grew 363 percent in a year to about $638 billion. Oracle's nominal backlog now exceeds the market value of its equity. The promise remained in place while its price collapsed.

Under American accounting, a lessee generally recognises an asset and liability only when the lease commences, when the underlying asset becomes available for use. The term, the rent and two decades of payments agreed before that moment remain in the notes rather than on the balance sheet. A data centre's first existence is textual, a promise a court would enforce, sufficient to make bankers lend and utilities build. No ministry coordinates it and no appropriations bill contains it. The decisions are concentrated in about five executive suites.

I.

Microsoft disclosed $329.1 billion of leases yet to commence at 30 June, against $196.6 billion three months earlier, with commencement dates running from fiscal 2027 to fiscal 2033. Meta reported $279 billion at the same date, half again what it carried in March, on terms extending as long as thirty years. Those two disclosures alone, on a single accounting basis, come to more than $600 billion of rent on buildings that do not operate. Adding Amazon, Alphabet and Oracle takes the figure past a trillion.

The category matters more than the total, because the totals in circulation are assembled from incompatible parts. Moody's counts every lease commitment across six companies and reaches $1.2 trillion; Bloomberg, adding purchase commitments, reaches $2.4 trillion for four; Alphabet's own $902 billion folds energy, equipment and leases into one line. The uncommenced-lease figure is narrower and harder to argue with. One category, one basis, buildings that do not yet exist.

The defence is scale. These are among the strongest balance sheets in commercial history, and a trillion dollars spread across fifteen or nineteen years looks manageable beside what they earn. That defence looked sound eighteen months ago and has become less comfortable as the commitments have multiplied.

On current forecasts, capital spending across the five rises by $534 billion between 2025 and 2027 against a $340 billion increase in operating cash flow. The buildout therefore absorbs $1.57 of additional capital for every additional dollar of operating cash generated over the interval. Alphabet and Amazon have each tipped into negative free cash flow, and Meta is expected to follow. Microsoft, Alphabet and Meta retain the capacity to be wrong and survive comfortably.

Oracle lost much of that freedom a year ago. Capital spending rose 162 percent to $55.7 billion in fiscal 2026 against $32 billion of operating cash flow, producing a free cash flow deficit of $23.7 billion. S&P now expects that deficit to approach $42 billion in fiscal 2027 on capital spending of $90 to $95 billion, half again its own forecast of a few months earlier. The company issued $5 billion of mandatory convertible preferred in February and plans $20 billion of equity this year.

The cash still arrives in torrents and leaves again in the same quarter, for the same purpose. What the notes disclose is discretion already spent, recorded with the antiseptic thoroughness American accounting reserves for the genuinely astonishing.

II.

Demand is real. AWS grew 37 percent in the second quarter, its fastest since the end of 2021, which is the wrong direction for anyone expecting saturation. Google Cloud grew 63 percent in the first. Sundar Pichai said Google was compute-constrained and that cloud revenue would have been higher had the company been able to meet demand. At Amazon, Andy Jassy's position was that most new supply had already been spoken for, which is a comfortable thing to say about capacity nobody outside the company can verify. Microsoft added $450 billion of market value on 30 July, the largest single-day gain in the history of the stock market, in the same week it disclosed the $132 billion. The market is not currently punishing the buildout. It is paying for it.

The customers are real too, and more fragile than the aggregate suggests. OpenAI was reported at roughly $25 billion of annualised revenue in March, up from $6 billion two years earlier, against $13.1 billion of revenue in 2025 and roughly $34 billion of expenses. First-quarter revenue came in near $5.7 billion, below internal targets, with competition from Google, Anthropic and a widening field of open-weight Chinese models cited as the reason. Around a billion people use these systems monthly and roughly sixty million pay. Whatever constrains this business, it is not obscurity, and close to a billion people have already tried the product on hardware they owned at a price of zero, which places the consumer market at its broadest point now rather than at some date in the future.

The larger case rests on agents. An agentic system consumes far more tokens than a chat exchange and, if run continuously, allows demand to scale with the number of tasks in an economy rather than the number of people typing into a box. That is almost certainly the assumption that the buildout is actually underwritten by. It also produces the most price-sensitive demand there is. A consumer pays $20 a month semi-irrationally; an enterprise measures the cost of an agent against a salary line and changes provider for fifteen percent. The demand that scales is the demand that shops — high volume, thin margins, ruthless substitution — which is a description of a commodity utility, and what none of these companies are priced as.

The rejoinder is Jevons. If each unit of intelligence gets cheaper, more of it gets bought, and total spending can rise while unit prices fall. That is likely correct, and it protects the buildings. A shell can be re-racked, a grid connection serves whatever is plugged into it, and land near cheap power appreciates. What it does not protect is the schedule. Every Nvidia generation supplies more performance per watt, so the intelligence produced by a given rack loses value on a roughly annual clock while the rent beneath it is fixed for nineteen years. In a quarter when its chief executive said demand exceeded supply, Google cut compute prices by 8 percent across all regions.

Oracle faces a second uncertainty. Anthropic's run rate reportedly passed $47 billion in May, against roughly $10 billion for all of 2025, on enterprise and coding work from a much smaller user base. Frontier compute has a market. Its division among providers remains unsettled, and S&P puts close to half Oracle's backlog on one of them.

III.

A falling price matters less when the thing being paid for holds its value. Nothing inside these buildings holds it at the same rate. Land and grid access may last for decades, while the shell depreciates over twenty-five or thirty years. Servers are assumed to last five or six. The accelerators that supply the purpose of the facility can be commercially superseded by the next product launch, and Nvidia now ships a new architecture roughly every year.

In January 2025 Meta extended the estimated useful life of its servers, deferring about $2.9 billion of depreciation. Amazon moved the other way in the same month, shortening the life of some servers and networking equipment from six years to five and citing the increased pace of development in artificial intelligence, at a cost of roughly $0.7 billion of operating income with another $0.6 billion attached to equipment marked for early retirement. The same asset class met the same technological shock in the same month and produced opposite audited estimates. Amazon had lengthened those lives only a year earlier.

The estimates now move in the direction the buildout requires. Microsoft extended the depreciable life of its data-centre buildings from fifteen years to twenty-five in the same filing that disclosed the $329.1 billion, and reclassified enough future leases from finance to operating treatment to move reported calendar-2026 capital spending from roughly $190 billion to roughly $175 billion, while stating that its underlying investment expectation had not changed. Oracle, carrying the thinnest cover for a mistake, moved from five years to six.

Michael Burry offered the aggressive version in November 2025, calculating that five- and six-year schedules against economic lives of two or three years would understate depreciation by roughly $176 billion across 2026 to 2028. He was short Nvidia, although his inputs were public and audited. Meta's own filings show the strain from the other direction: depreciation on servers and network assets rose 83 percent in two years, after an extension that deferred nearly $3 billion of it.

Physical utilisation can run for years while economic value at the frontier collapses. Project finance already separates some of these clocks, using long debt against land and power and shorter finance against equipment. What it has not produced is a common way to allocate the loss when one generation of machinery devalues several layers at once. Answering that question deal by deal is possible. Answering it across the system would slow the building down.

IV.

Richland Parish, Louisiana, has about twenty thousand residents and now hosts the most expensive piece of private infrastructure in American history. Meta's Hyperion campus was announced at $10 billion in December 2024 and is now projected above $200 billion. In October 2025 funds managed by Blue Owl took 80 percent of a joint venture to develop and own it, with Meta keeping 20, contributing land and assets under construction, and taking a $3 billion distribution at closing. A vehicle called Beignet Investor issued $27.3 billion of senior secured notes due 2049, arranged by Morgan Stanley and anchored by PIMCO and BlackRock, rated A+ by S&P, one notch below Meta itself. Outside capital owns most of the infrastructure, whose reason for existing is Meta's promise to use it.

The structural terms of that promise are interesting. Meta's initial lease runs four years, with options to extend, against a bond maturing in 2049. The gap is bridged by a residual value guarantee covering the first sixteen years, under which Meta would make a capped cash payment based on the then-current value of the campus if it walked away. Twenty-five years of debt service rests on four years of contracted rent and a guarantee that expires nine years before the notes do. The clocks have been separated. At the far end they are attached to nothing. Louisiana's Public Service Commission approved three 754-megawatt gas plants for the campus two months before the financing structure became public; the generating assets have utility lifespans and the customer arrangement has a four-year initial term.

Oracle has taken the opposite position. It borrowed against a promise it received. If the reported OpenAI agreement accounts for nearly half its backlog, Oracle has levered its balance sheet against a single counterparty whose losses remain enormous, and its own annual report now discloses that some customers may be highly leveraged and unable to meet their contractual obligations. It resembles a creditor of somebody else's forecast, carrying a concentration few regulated banks would be permitted to hold without substantial capital behind it.

Meta shifts the financing outward. Its promise becomes the asset against which private-credit funds and their lenders advance money, each standing at a different distance from one Meta forecast and paid for the legal risk passed down to it. Meta retains more optionality and sells certainty, although the promise remains enforceable and retreat would carry a cost. Its exposure is chiefly to its own forecast. Oracle's is to somebody else's.

The pattern recurs in the other half of the diptych. CXMT's return depends on restraint it does not control, with three incumbent rivals choosing from quarter to quarter how much memory to produce. Zhongji's return rests on a lead it has already demonstrated. Oracle rides a promise made by a weaker counterparty while Meta supplies the promise. In each pair the firm that becomes creditor to somebody else's conviction occupies the more precarious position, whatever the balance sheet says on the day.

A corporate commitment of this kind performs work often associated with the state. A municipality makes a power station financeable by guaranteeing its output; Meta's future rent turns graded earth and an electrical drawing into something a pension fund in Rotterdam will underwrite. The resemblance weakens in failure. A corporation cannot tax or compel a bank to refinance a dying project for reasons of state. It can default or restructure, leaving damages and an empty building to be divided among creditors.

China puts public capital in front of the company and carries technical risk until a manufacturer becomes investable on its own. America puts corporate credit in front of the project and allows an unbuilt facility to raise private money because somebody powerful has promised to use it. The Chinese state creates an investable company. The American corporation creates an investable asset. Both are planning. Only one says so.

V.

On 27 July the method was proposed in a purer form. Nvidia is reported to be negotiating a guarantee of as much as $250 billion of OpenAI's lease payments on a ten-gigawatt campus in Pike County, Ohio, built on a decommissioned uranium-enrichment site by SoftBank's energy subsidiary in partnership with the Department of Energy, with a further $350 billion discussed for the chips inside it. The talks are early and may collapse. The stated reason for the guarantee is that OpenAI has no investment-grade credit rating and a campus of that size cannot be financed without one, so the largest chipmaker in the world proposes to lend its balance sheet to the tenant in order that the tenant can rent buildings in which to install the chipmaker's product. The previous record for a commitment of this kind was Google's roughly $44 billion of data-centre rent guarantees. Nvidia's shares fell about 5 percent on the report.

The rhyme with Hefei is worth stating alongside the difference. Both arrangements place a stronger balance sheet in front of a borrower the private market will not fund on its own credit. The guidance fund's money has already gone and cannot be revoked by a change of mood, and the institution holding the loss answers to a hierarchy that can decide to wait. A guarantee is contingent, negotiated and worth what the guarantor is worth on the day it is called, and Nvidia's balance sheet is itself a function of the accelerator demand the guarantee exists to sustain. That makes it a cheaper instrument than sunk public equity and a weaker one, and it is the instrument available because no American institution has the authority to put the money in directly.

Vendor financing is old and often legitimate. Railways and telecoms were both helped into being by suppliers lending to customers. The current version becomes peculiar where participants gain from accepting a higher number. Nvidia has committed more than $40 billion of equity in 2026, $30 billion of it to OpenAI, and taken stakes in CoreWeave, Nebius, IREN and others. Amazon invests in a company whose principal expense is cloud computing. A seller booking revenue from a company it has funded has little reason to resist the valuation. The useful test is how much revenue enters from outside the circle, and that is the figure nobody publishes.

Nobody chose the total. Every firm concluded, rationally at the level of the firm, that underbuilding presented an existential risk if a rival took the capacity first. Nobody is paid to decide on the industry's behalf that enough has already been ordered. China gets duplication when provinces converge on a political priority, and its formal apparatus can block a listing or starve a sector of credit, although the first half of this diptych showed how often it acts after a price has collapsed. America gets duplication when corporations converge on the same fear, and the American version is harder to arrest because every participant can explain his own decision.

VI.

In the late 1990s American carriers laid something over 80 million miles of fibre on the strength of WorldCom's claim that internet traffic was doubling every hundred days. In reality it was roughly doubling each year. Carriers sustained the appearance of demand by buying capacity from one another through reciprocal rights of use, while Lucent lent customers the money to buy Lucent equipment. The SEC later found that, without these arrangements, Global Crossing's adjusted EBITDA for the second quarter of 2001 would have been a $43 million loss rather than a $472 million profit.

The disanalogy is severe. Global Crossing's cable was dark the day it was laid, waiting for applications that had yet to be invented. AI capacity is sold before it exists, to customers with revenue, by providers that say they cannot build quickly enough. Capacity in search of a use defined the earlier boom and is absent from this one.

Yet the fibre forecast was right about direction. Traffic grew enormously, at annual doubling rather than every hundred days. Estimates of how much fibre remained unlit afterwards run between 85 and 95 percent and none is firm, but by 2004 wholesale long-haul prices had fallen by more than half in a year, which is the figure that mattered. The traffic eventually arrived and the cables carried it. The people who paid for them were destroyed anyway. Calling that a failure of timing is too generous, since timing implies patience would have fixed it.

The capital structure had named a date. A nineteen-year lease and $130 billion of debt encode revenue of a certain size arriving on a schedule. Demand that turns up three years late may never reach the creditors who financed it. It reaches whoever buys the campus out of administration. Infrastructure survives an error in timing because it remains in the ground. A capital structure dies when the date inside it passes.

VII.

The first half of this diptych closed on a household holding the only position in its chain that reprices publicly and in real time. America produces a queue rather than a single terminus, and the queue does not run in the order the contracts imply.

Microsoft, Alphabet, Amazon and Meta absorb a bad forecast in guidance. Cash from businesses unrelated to any campus meets the shortfall, while the correction appears as a slower capital-spending ramp or compressed margins. A developer absorbs it in re-lettability, a harsher currency. A campus built to one tenant's specifications must find another occupant at a rent sufficient to service the debt raised to build it, and most purpose-built single-tenant industrial property historically has not.

A neocloud absorbs the error in covenants. CoreWeave carried $50.8 billion of liabilities against $4.8 billion of equity at the end of the first quarter, so a shortfall arrives as a financing event on a schedule the lenders set. In April the company walked away from its anchor position in Poolside's two-gigawatt Texas campus, days after closing an $8.5 billion facility underwritten by a long-term Meta contract. Reporting indicates the relationship had been failing since late 2025, when Poolside could not stand up its first cluster to CoreWeave's timeline; the anchor lease and a $2 billion funding round collapsed within days of each other, and which caused which is not established. Four months on there is no replacement tenant.

The public shareholder absorbs it first and fastest, in price. Oracle's investors waited for no default or missed payment. They repriced the promise and roughly two-thirds of the company's value left in ten months while the reported agreement remained. The most junior claim in the structure moved before any senior one, on expectation rather than event.

The symmetry between the two halves of this diptych breaks here but the fracture has its own analytical payload. Oracle's aggregate risk was counted publicly and violently inside a year, by a market that needed no permission and no plenum. The household in part one has no such mechanism until the fab is built and the shortage is over. America is not a system in which nobody counts the total. It is a system in which the count arrives early and lands on the holder least able to act on it. A pension fund cannot easily lend into a data centre, since the assets are poorly standardised and the technology unstable, so it owns the firm that guaranteed the rent instead, which is the same exposure with worse visibility and no way to price it separately. When Nvidia guarantees a lease it cannot recover from a tenant it also funded, the exposure has not been reduced. It has moved into the largest position most retirement portfolios hold.

Different companies and campuses will settle differently, but the weight of evidence points towards industrial success with weaker corporate returns. America gains abundant domestic compute and a deeper power system. Competition passes much of the benefit to users, while revenue grows beside depreciation and replacement spending. National success can sit beside local ruin. Northern Virginia stays full while the equity beneath a specialised site three states away goes to nothing, and the land and the grid connection hold their value while the financing above them comes apart. A useful building may be worth far less than the promise that brought it into being.

VIII.

These companies now reserve power years in advance, buy land and guarantee rent as a utility does. They are acquiring the obligations of a utility while retaining the valuation of a platform. One of those conditions will eventually give.

The choice may still be right. A chief executive who underbuilds and loses the next platform takes little consolation from having preserved a beautiful balance sheet. Demand is present and the fear is rational at the level of any firm. Yet the wager has changed what a shareholder owns. An uncommenced lease eventually becomes a facility in service, with years of payments attached. Accounting recognises the obligation at commencement. The decision was made at signature, often years earlier, by people who may have moved on before the accounts catch up.

China can defer a loss through political patience because the institution holding it answers to a hierarchy. America writes more of the residue into a contract. The building may change hands and the accelerator may age, while the agreement retains a counterparty and a court behind it. Contract provides a firmer claim than political patience, and it disperses discretion through markets, restructuring negotiations and bankruptcy courts rather than removing it. What America lacks is any institution willing to hold a loss quietly for a decade until the mood improves, and any institution empowered to decide, before the money is spent, that the total is too large.

Oracle made its wager on artificial intelligence through one overwhelmingly important customer. The demand can arrive in full while the wager fails, because a promise worth $300 billion over five years is worth something else when the same money arrives three years late. The date is a price. Money arriving in 2032 instead of 2029 loses value by the cost of the capital that waited for it, while the rent falls due on the original schedule.

Every senior claim in these structures is specified in advance. The lender has a coupon and the landlord has rent. Power suppliers have take-or-pay agreements, indifferent to whether the machines inside earned anything. Revenue appears too, in backlogs and guarantees, although it remains a promise made by a counterparty. One party receives whatever survives those prior claims. The shareholder's share is the only figure in the structure that nobody has agreed beforehand.

These buildings will fill, which is unsurprising given demand was never seriously the question. What matters is who still owns them on the day it arrives.