The scale of the buildout
Microsoft reiterated on 29 July 2026 that it expects roughly $190bn of capital expenditure in calendar 2026, while also telling investors that an accounting change reduces the reported figure to about $175bn. For scale, the same company spent $15.8bn in the quarter ended 31 December 2024 and $29.9bn in the fourth quarter of 2025. One firm's annual budget is now larger than the entire fibre-optic overbuild that bankrupted the telecom sector, and it is one of several.
The borrowing has scaled with it. Investment-grade AI-related issuance ran at about $10bn in 2024 and about $109bn in 2025; Amazon, Alphabet, Meta and Oracle alone sold about $194bn of bonds by 7 July 2026, with JPMorgan estimating $279bn of investment-grade issuance for the full year. US AI-infrastructure high yield added $26.6bn in the first four months of 2026. Data-centre securitisation ran at roughly $27bn in 2025 with $30–40bn projected for 2026.
The more interesting number is the one that is not on any balance sheet. Nikkei went through the filings of Alphabet, Microsoft, Amazon, Meta and Oracle and found about $1.65tn of off-balance-sheet obligations — more than the roughly $1.35tn of debt they actually carry, and around eight times the level of four years ago (Nikkei). Meta's share is about $420bn, nearly triple its recorded debt. Oracle's is $273.3bn as of the end of May 2026, up more than thirtyfold in four years. Goldman puts sector-wide leasing commitments at about $1.5tn against roughly $200bn five years ago, of which roughly $1tn is invisible in headline statements.
The mechanism is dull and entirely legitimate, which is precisely why it works. A lease becomes a balance-sheet liability when the lease commences. A signed twenty-year commitment on a building that is currently a field in Ohio lives in a footnote, as do undelivered GPU purchase contracts. So the period of maximum commitment is also the period of minimum disclosure, and during it the leverage ratios look like those of the capital-light software businesses these companies used to be. We wrote this up on 19 August.
The revenue gap and accounting illusions
The question of whether the revenue exists was asked early and clearly. David Cahn of Sequoia laid it out in AI's $600B Question, which simply asked where the money to service the infrastructure spending was going to come from.
Two years on, the arithmetic at the single most important customer runs as follows. OpenAI's annualised revenue is more than $40bn as of August 2026. Its projected cash burn through 2028–29 is $115bn. Its total multi-year compute and infrastructure commitments are about $1tn. The revenue is roughly a third of the burn and about four per cent of the commitments. Every structure described in this entry exists to bridge that gap.
Meanwhile, the reported profitability of the buyers is partly a choice about depreciation schedules. Depreciation spreads an asset's cost over the period management expects it to be useful, and management gets to pick the period. Microsoft says its computing equipment typically lasts two to six years; Google, Oracle and Microsoft have all indicated servers can run to six. Meta extended its server useful lives to 5.5 years in early 2025, which reduced depreciation expense by $2.9bn. The works in reverse too: cutting Amazon's assumed server life from six years to five raised depreciation expense by $700m and lowered operating income by the same amount.
The difficulty is that a number of analysts argue AI hardware becomes economically obsolete in about two to three years, because new GPU generations arrive fast and the performance jumps are large. So the accounting life is roughly double the useful life. If they are right, current earnings across the hyperscaler complex are borrowing from future write-downs, and the borrowing is being reported as profit.
Circular financing and vendor guarantees
When the customers cannot support the debt, the vendor does. Nvidia's reported commitments to OpenAI escalated through 2026: a letter of intent to invest up to $100bn, announced 22 September 2025, tied to the deployment of 10 gigawatts of Nvidia systems; then a reported up to $250bn financing guarantee in July 2026 to backstop OpenAI's leasing of computing power from a US data-centre project; and a separately reported discussion of up to $350bn of financing for OpenAI chip purchases. Bloomberg framed the package as part of more than $750bn of AI arrangements under negotiation.
Those headline figures are not cleanly reconcilable with what has since been documented at deal level, and we would rather say so than pretend. What we have reported is the Ohio transaction: Nvidia agreed to guarantee up to $105bn of OpenAI's lease payments on a Pike County campus being built by SoftBank-backed SB Energy, plus $1.5bn of equity into SB Energy itself, with the backing phasing in between 2028 and 2030, running twenty years, and covering only the first 4.25GW of a roughly 8GW site (Bloomberg). Subsequent reporting narrowed it further. The commitment is a backstop on a portion of the value of the completed data centre, structured to require a series of mitigation steps, and SB Energy would first have to try to re-lease the site before Nvidia owes anything (WSJ). Our estimate of Nvidia's direct exposure went down on that reporting, not up.
The function is unchanged whatever the number is. Lenders will not fund a campus of this size against a five-year-old company's credit. They will fund it against Nvidia's. Same building, same tenant, radically cheaper money, and the money buys Nvidia chips. The same shape appears in the equity stakes and capacity deals: $6.3bn of cloud capacity bought from in September 2025, a further $2bn of CoreWeave equity in January 2026 at $87.20 a share, $5bn into xAI in December 2024, $2bn into Nebius in March 2026. On top of that sit memoranda with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilise more than $500bn of third-party compute financing, of which Jensen Huang has said Nvidia could backstop up to $125bn. Huang says this is not circular financing.
It is not an American peculiarity either. We covered the Chinese version on 20 August: humanoid robot makers selling machines to government-backed training centres they themselves part-fund, then buying the training data back. Nearly 370 Chinese humanoid startups have been founded in two years and more than 50 have listed or are preparing to; Unitree rose more than 600% after listing on Shanghai's Star Market, reaching a $50bn valuation. Every leg can be priced at arm's length and every contract can be real. The net cash arriving from outside the loop is still much smaller than the revenue line implies.
The regulatory blind spots
In a no-action letter dated 29 July 2026, responding to a request from Latham & Watkins dated 23 July, the SEC agreed that a large subset of data-centre securitisations are not asset-backed securities. The reasoning is that a data centre is not a "financial asset" like a mortgage or a car loan. The consequence is that the 5% credit risk retention requirement, the post-2008 rule forcing whoever assembles a deal to keep a slice of it, does not apply, along with parts of the standardised disclosure regime (Telegraph).
The technical argument is genuinely defensible. A building with a tenant is not a pool of consumer loans, and the loan-level disclosure template does not fit. Practitioners were clear about the effect anyway. "It gives them the opportunity over time to push down the required equity in the deal," Orion Mountainspring of Orrick told CNBC. Cambridge's Raghavendra Rau summarised the posture as the SEC replacing regulatory constraints with a reliance on market discipline. Two caveats we keep attached: this is staff guidance, not a rule, and staff guidance can be withdrawn.
The insurance market has made a similar decision by simply declining the business. Meta and BlackRock's one-gigawatt El Paso campus, internally called Sopaipilla, cost about $14bn and carries up to $427mn of all-risk property cover during construction and $450mn once operating, for a premium of roughly $5mn a year rising 2% annually, plus $645mn of terrorism cover, $218mn of rent-abatement cover and general liability capped at $50mn per event and $50mn in aggregate (FT). The project is not insured against total loss. That is roughly three per cent of asset value covered, because insurers will not build that much aggregate exposure to a single site. The gap does not vanish. It sits with whoever owns the debt. Whether the lease abates on a casualty, the clause that decides whether this is a footnote or a hole, is not public.
The physical-permitting version is xAI in Memphis. The turbine count at the site went from 18 in August 2025 to 27 in spring 2026, 46 in May 2026 and 69 by late July 2026. The units were framed as "mobile" and "temporary" and therefore outside Clean Air Act permitting for fixed sources, an interpretation the EPA concluded violated federal air rules. Mississippi's Department of Environmental Quality barred new turbine additions after 15 July 2026 and required removal under an agreed order beginning 18 August 2026. A permanent replacement plant of 41 stationary combustion turbines, about 1.2GW, had already been approved by the Mississippi Environmental Quality Permit Board on 10 March 2026. The NAACP, the Southern Environmental Law Center and Earthjustice sued. The surrounding community is around 40% Black.
A short history
The closest precedent is the of 1999–2001, and the resemblance is structural rather than poetic. Roughly 80 million miles of fibre were laid in the United States by 2001–02, of which somewhere between 85% and 95% was never lit. Lucent carried more than $15bn of vendor financing at its peak against operating cash flow of around $300m, the vendor manufacturing the demand that justified the vendor. Global Crossing and WorldCom both went bankrupt in 2002, WorldCom with more than $11bn of fraudulent accounting behind it, and Lucent no longer exists. The fibre, it should be said, was eventually used.
Why it matters to this crash
The equity market has already split the AI trade in two, and the fault line is the balance sheet rather than the technology. Over the twenty trading days to 20 August 2026, Microsoft was up 24.1% and Amazon up 8.6%, while Core Scientific was down 20.8%, Talen 14.6%, Vistra 14.4% and Meta 12.9%. Over five days: CoreWeave −15.6%, Nebius −13.6%, Talen −11.7%, Core Scientific −10.2%, Applied Digital −9.3%. That is not a sector losing faith in AI. That is the market sorting the complex by who funds capex from operating cash flow and who funds it with project debt, leases and power contracts. Meta is the one hyperscaler in the losing column, and it is the one that has leaned hardest on off-balance-sheet structures.
The credit market made the same point with a single number. Blackstone-backed QTS marketed about $3.9bn of five-year notes for a Microsoft-linked data centre, Project Odyssey, at initial talk of roughly 7.63%, on paper expected to be rated Baa3 by Moody's and BBB− by Fitch (Bloomberg). At the time the investment-grade index was 81 basis points and high yield was 270. An investment-grade rating clearing at a junk yield is the ratings agencies pricing the tenant while the buyers price the building. The deal drew about $10bn of indications and was upsized by roughly $1bn, so the market is not refusing to fund this. It is refusing to fund it cheaply. Where the rating and the clearing price disagree by this much, believe the price.
is the other mismatch. The financing vehicle for Meta's Hyperion campus in Louisiana carries roughly $27bn of debt, rated A+, priced at 6.58% and amortising out to 2049, with PIMCO reported to have taken about $18bn and Blue Owl-led investors holding around 80% of the equity. Those splits come from secondary aggregation rather than a filing we have read, so treat them as indicative. The structure is the point: a twenty-three-year amortisation against GPUs most operators depreciate over about five. Somebody re-equips that hall four times, at prices nobody can forecast, for the rent to keep arriving. That is a bet on the site — power interconnection, land, cooling — outliving several generations of silicon, and it is being placed with insurance and pension money, through vehicles whose buyers do not it daily.
And it is crowding the sovereign market. The 30-year Treasury yield hit 5.31% on 17 August 2026, the highest since 2007, after a $25bn 30-year auction cleared at 5.216%, the highest since 2001 (Bloomberg). Nomura estimates the roughly $200bn borrowed by the biggest tech companies equals about 25% of Treasury's net issuance of notes and bonds to private investors, five times the 2025 share. "Whoever's issuing, be it a government or a hyperscaler, is now competing with more borrowers," Nuveen's Tony Rodriguez told Bloomberg. "And therefore yields have to be higher."
What would make this dangerous
A vendor guarantee actually being called. Nvidia's Ohio backstop requires SB Energy to attempt a re-lease first. The first time a purpose-built gigawatt-scale GPU hall is offered to a tenant other than the one it was designed for, we find out what is worth. Every backstop in the complex rests on that untested number.
A QTS-style deal failing to clear. Project Odyssey cleared at 7.63% with $10bn of indications. If the next one has to widen again and still comes up short of the book, the marginal lender has stopped, and the buildout is funded from operating cash flow only.
A casualty at an underinsured campus. $450mn of cover against a $14bn asset, with the lease's abatement clause unpublished. One fire, one storm, and the answer becomes public in the worst possible way.
Useful lives being shortened. If move server lives back toward the two-to-three-year economic estimate, the earnings that justify the capex fall by billions a year at each company, all at once, with no cash changing hands.
continuing to reprice. Hyperscaler ultra-long issuance is competing with governments for the same limited pool of duration buyers, at 30-year yields already at 2007 levels. Japan's 10-year touched 2.93%, and Deutsche Bank's Shoki Omori has noted that 3% is "a critical defence line for fiscal credibility" because it is the rate assumed in the government's own budget. Every valuation in the AI complex discounts against that curve.