NextFin News - The world's most valuable companies are building the artificial-intelligence era on a foundation of promises they have not yet had to keep. Amazon, Microsoft, Alphabet, Meta and Oracle are using guarantees, residual-value backstops and special-purpose vehicles to hold roughly $300bn of AI-related exposure off their balance sheets, converting what would otherwise be debt into future rent. The arrangement lets the hyperscalers race to build data centers without showing the full cost on their books - but it also plants the bill for that buildout with private credit funds, insurers and the index funds that own both the tech giants and their lenders. The broader obligation is larger still: by the end of 2025 the five companies had signed $969bn of future data center lease commitments, of which $662bn had not yet commenced and therefore did not appear as a liability at all.
The Scale of the Off-Balance-Sheet Pile
The scale is hard to overstate. Moody's Ratings calculated that the $662bn of not-yet-commenced leases equals roughly 113% of the five hyperscalers' combined adjusted on-balance-sheet debt. In other words, the infrastructure these firms have promised to pay for now exceeds the debt they officially report. A separate investigation put the five companies' total hidden obligations - leases, GPU contracts and off-book ventures - at $1.65tn, about eight times the level four years earlier.
The mechanism is straightforward. A special-purpose vehicle - typically a joint venture between a hyperscaler and a private credit sponsor - raises the money and builds the data center. The hyperscaler takes a minority equity stake, signs a long-term lease or capacity offtake agreement, and often adds a residual-value guarantee: if it walks away, it covers the gap between the facility's market value and a pre-agreed floor. The debt stays in the SPV. The hyperscaler records rent, not borrowings.
Moody's analysts David Gonzales and Alastair Drake stress that this is not a hidden liability in the accounting sense. The obligation has not yet been recognized because the related services have not been delivered, Gonzales said. The commitments will migrate onto balance sheets as the leases commence. But the timing gap - between when the risk is economically incurred and when accounting rules force it into view - is exactly where the opacity lives. Moody's has warned that disclosures may not show the full picture, because the accounting treatment of short initial lease terms combined with residual-value guarantees can cause reported liabilities to materially understate companies' true obligations.
The Guarantee Has Replaced the Bond
For decades, hyperscalers funded their own expansion. Amazon, Alphabet, Meta and Microsoft spent a combined $381bn on capital expenditures in 2025, almost entirely from operating cash flow. That model broke under the AI buildout. Spending is forecast to rise more than 60% in 2026, to roughly $700bn for the four companies alone, while AI revenues - about $60bn in 2025 - remain a fraction of the capital being deployed. The gap cannot be funded from cash flow, and issuing that much corporate debt would crush investment-grade credit ratings.
So the guarantee replaces the bond. By pledging to make lenders whole if a project fails, the hyperscaler gives private credit the credit quality it needs without formally borrowing. The Bank for International Settlements calls this "shadow borrowing": obligations economically akin to debt that live outside corporate balance sheets. Economically the hyperscaler is on the hook; legally and accountingly, it often is not.
Meta's Hyperion campus in Louisiana illustrates the template. The roughly $30bn project is financed through a joint venture in which Blue Owl Capital holds the majority of the equity and Meta holds 20% while retaining full operational control. Morgan Stanley arranged more than $27bn of debt and about $2.5bn of equity into the special purpose vehicle; PIMCO anchored the lending. The debt matures in 2049 and was priced in the 144A private-placement format, with completion due in 2029. Meta records only its equity stake and a lease obligation; the $27bn of underlying loans never touches Meta's balance sheet. Reports on the deal described the SPV debt as rated one notch below Meta's own AA- credit, at a spread of roughly 225 basis points over Treasuries - the market's price for moving tens of billions off balance sheet.
Why It Has Not Broken Yet
The buildout has survived scrutiny so far because demand for compute has grown faster than anyone modeled. Nvidia's GPUs have remained scarce, hourly rental rates have risen, and every hyperscaler has guided to higher capital spending rather than lower. In that environment, a residual-value guarantee looks like free money: the guaranteed asset is appreciating, so the guarantee is never triggered.
That is the cyclical leg of the trade, and it is mean-reverting by construction. AI server equipment typically has a useful life of four to six years, versus 10 to 15 years for traditional data center infrastructure. Hyperscalers have responded by signing shorter initial lease terms with renewal options - which is precisely what forces the guarantees into place, because landlords will not build billion-dollar facilities on a three-year promise. The structure works only as long as each renewal actually happens. The moment a hyperscaler declines to renew, the guarantee crystallizes into a cash payment.
The Structural Shift: Compute Becomes a Securitized Asset Class
Beneath the cycle is a durable change in financial architecture. Compute is being transformed into a bankable, financeable asset class - with Nvidia playing the role that Fannie Mae and Freddie Mac once played in mortgages. In August 2026, Nvidia announced partnerships with private equity funds to mobilize more than $500bn of third-party capital for AI cluster buildouts, backing each deal with a residual-value guarantee of up to 25%. Jensen Huang emphasized that the support was limited and that institutional investors would price each project. But the effect is the same: Nvidia becomes the arbiter of what counts as bankable compute, the liquidity provider of last resort, and the guarantor standing behind the paper.
This is not a cyclical fluctuation. Once capital markets learn to underwrite GPU-collateralized loans, once insurers and pension funds hold the tranches, and once rating agencies build models around hyperscaler guarantees, the financing template persists even if the capex cycle turns. The debt intensity of the sector has already shifted structurally: the five hyperscalers issued more than $100bn of debt in 2025, more than four times their average annual pace of $28bn from 2020 to 2024. Goldman Sachs expects roughly $250bn of direct hyperscaler supply in 2026 alone, and project-finance and data center transactions of about $300bn on top of that in 2027. Global AI-related debt issuance is on track to reach nearly $570bn in 2026, more than double 2025, according to Morgan Stanley.
The Second-Order Risk: The Same Lenders on Every Deal
The first-order risk is obvious - a hyperscaler defaults and its guarantors pay. The second-order risk is concentration. A small club of private credit sponsors - Blackstone, Blue Owl Capital, Apollo, PIMCO and BlackRock - appears on both sides of these transactions: they sponsor the SPVs, they lend the money, and their insurance-company clients own the paper. When five borrowers use one structure funded by one handful of lenders, the risk stops being idiosyncratic and becomes systemic.
The transmission channel runs through institutional balance sheets. Insurers and pension funds that bought SPV bonds on the assumption of hyperscaler support now carry exposure to the same AI demand assumptions that drove the hyperscalers' own capex. If compute demand disappoints, losses do not stay in the tech sector. They travel to the private credit funds, then to the insurers that back them, then into the retirement accounts that own the insurers. The Bank for International Settlements has flagged exactly this channel: banks fund the vehicles with credit lines, creating shock transmission paths through refinancing pressure, procyclical shifts in private credit appetite, and the activation of guarantees.
The Adversarial Case
The strongest counter-thesis comes from investors who have been right about accounting opacity before. Michael Burry, whose Scion Asset Management famously shorted the housing bubble, argued in November 2025 that hyperscalers are understating depreciation by extending the useful lives of their chips. He estimated the maneuver would understate depreciation by about $176bn from 2026 through 2028, inflating reported earnings - with Oracle's profits potentially overstated by roughly 27% and Meta's by about 21% by 2028.
"Understating depreciation by extending useful life of assets artificially boosts earnings - one of the more common frauds of the modern era," Burry wrote.
Jim Chanos, the investor who shorted Enron, has echoed the concern about AI spending and opacity.
The bears' analogy is the telecom fiber buildout of the late 1990s, when companies poured more than $500bn - mostly financed with debt - into laying fiber optic cable in the five years after the 1996 Telecommunications Act. By various estimates, 85% to 95% of the fiber laid in the 1990s remained unused after the bubble burst, earning the nickname "dark fiber." The parallel is uncomfortable but not exact. The 1990s builders had no revenue base; today's hyperscalers are among the most profitable companies in history, and AI revenue is genuinely growing.
The counter-thesis attacks the foundation of the bull case: that AI revenue will scale to meet the capital being deployed. It is backed by named authorities and by accounting logic that does not depend on the mood of the market. It deserves its weight.
What Would Prove the Bears Right
The falsifying signal is specific. If hyperscaler AI revenue - roughly $60bn in 2025 - does not grow to cover the combined annual depreciation and lease expense of the buildout within three years, the structure breaks mathematically. More precisely: if the $662bn of not-yet-commenced leases begins migrating onto balance sheets in 2027-2029 while AI revenue growth decelerates below 30% annually, free cash flow margins will compress and the guarantees will move from accounting footnote to cash outflow. A second signal: if GPU residual values fall below the strike prices in Nvidia's and the hyperscalers' residual-value guarantees, the backstops activate and the off-balance-sheet exposure becomes a real loss.
Conclusion: A Bet on Accounting, Not Just Technology
The near-term path is clear. As leases commence, more than half a trillion dollars of data center assets will begin appearing on corporate balance sheets, pressuring traditional accounting metrics. Amazon is projected to generate negative free cash flow of up to $28bn in 2026, while Alphabet and Meta are expected to see free cash flow drop by roughly 90%. Credit ratings agencies are already reacting: S&P placed a negative outlook on Oracle's BBB rating in September 2025, and Moody's has flagged the risk in Oracle's roughly $300bn of recently signed AI contracts. In January 2026, four US senators urged regulators in an open letter to investigate the AI sector's growing reliance on debt, warning that an inability to service it "could cause destabilizing losses for an interconnected set of financial institutions."
Who benefits and who is exposed is asymmetric. The winners are the private credit sponsors and insurers collecting spreads on debt that carries a hyperscaler guarantee - they earn private-market yields on what is, in effect, quasi-investment-grade tech credit. The exposed are the same institutions if the guarantee is tested at scale, plus any investor who owns the hyperscalers' equity and assumed their leverage was contained. The hyperscalers themselves occupy the middle: they preserve financial flexibility today in exchange for a contingent liability that grows every quarter.
The time-horizon split matters. In the short term - the next 12 to 18 months - the structure is stable because demand for compute remains strong and renewal decisions have not yet come due. In the medium term - 2027 to 2029, as the first wave of leases commences and depreciation hits - the accounting picture darkens and the bears' case becomes testable. In the long term, the regime change is already locked in: compute has become a securitized asset class, and the guarantee architecture will outlive this capex cycle regardless of who wins the AI race.
Scenarios:
- Base case: AI revenue compounds at 40-50% annually, renewals happen, guarantees are never triggered, and the buildout is absorbed as a manageable drag on free cash flow.
- Upside case: AI applications reach profitability faster than expected, compute becomes cash-generative ahead of schedule, and the off-balance-sheet structures are vindicated as financial innovation.
- Downside case: revenue growth decelerates below 30%, GPU prices fall on oversupply, residual-value guarantees activate, and the $662bn migrates onto balance sheets as a simultaneous hit to earnings and credit metrics.
The AI buildout was sold as a bet on technology. It is increasingly a bet on accounting - on the gap between when a promise is made and when it must be kept. That gap is where the returns are, and where the risk hides.
Explore more exclusive insights at nextfin.ai.
