$1.65 Trillion in AI Debt Is Hiding in the Footnotes

Alphabet, Microsoft, Amazon, Meta, and Oracle are carrying more off-balance-sheet AI debt than they report as formal debt. The structures are legal. That's the problem.

Key takeaways

• A Nikkei Asia study found five U.S. tech giants hold approximately $1.65 trillion in off-balance-sheet AI infrastructure obligations, exceeding the $1.35 trillion they report as on-balance-sheet debt.
• The financing structures use special-purpose vehicles and lease arrangements that are GAAP-compliant, but they use the same mechanics that concealed Enron's liabilities before its collapse; when data centers go live, their leases roll onto balance sheets all at once.
• Oracle's off-balance-sheet exposure has grown roughly 30x in four years; S&P has already downgraded its credit rating citing AI commitments, and Moody's estimated $662 billion in pre-commencement lease obligations sitting off the books across the five companies as of February 2026.

Five of the largest companies on earth are running a financing operation that their headline balance sheets don't show. Per a Nikkei Asia study, Alphabet, Microsoft, Amazon, Meta, and Oracle have accumulated approximately $1.65 trillion in off-balance-sheet obligations tied to AI infrastructure, data center leases, GPU supply contracts, and special-purpose vehicles, a figure that has grown roughly eightfold since 2022. That $1.65 trillion exceeds the $1.35 trillion these five companies officially report as debt. The footnotes are now larger than the cover page.

The Structures, by Company

Meta's off-balance-sheet exposure sits at approximately $420 billion, nearly triple its transparent debt. A meaningful portion is tied to the Hyperion data center in Louisiana, financed through a joint venture with Blue Owl Capital that took on $27 billion in debt with Meta as the sole tenant. Meta argues it need not record the obligation.

Oracle is the most extreme case. Its off-balance-sheet AI debt stands at approximately $273.3 billion, up more than 2,900% in four years per the Nikkei Asia study. S&P downgraded Oracle's credit rating to BBB- on July 9, 2026, explicitly citing leverage from AI commitments including roughly $260 billion in future lease obligations. Those obligations will eventually land on Oracle's books when the data centers go live.

A February 2026 Moody's Ratings report estimated $662 billion in data center lease commitments signed but not yet commenced across the five hyperscalers. Under GAAP's lease commencement standard, those obligations sit off the books until the facilities open.

The Enron parallel is structural, not rhetorical. Bloomberg Law quoted analyst Gil Luria on the comparison: "Enron's crime wasn't having special purpose vehicles. Enron's crime was hiding them." The difference here is that the disclosures exist, buried in supplemental tables and footnotes almost no investor reads. That makes the risk legally contained but practically invisible to anyone relying on headline numbers.

An anonymous industry source cited by Nikkei put the chip problem plainly: "In a typical real estate lease, the building retains its value, but semiconductors, the core of AI data centers, depreciate rapidly due to the fast pace of technological advancement."

The Mechanism Nobody Is Pricing

That quote identifies the structural timing mismatch at the center of this. AI servers have replacement cycles far shorter than the bonds funding them, which carry durations of five to twenty years. When the next GPU generation renders current hardware obsolete, the collateral backing these structures evaporates, while the debt remains.

Financial regulators have raised concerns about this style of financing and its potential to heighten systemic market risk. The AI capex buildout has been accounting for a significant share of U.S. GDP growth, which raises the stakes if the demand side doesn't materialize.

The risk isn't primarily landing on hyperscaler equity holders. They have contractual options to walk away from SPV structures at a price. The risk lands on the bond investors, insurers, and pension funds who funded the SPVs and hold no such option. The C-suites get the upside. The write-downs go elsewhere.

This is also the accounting contrast that makes Bitcoin's architecture look less like ideology and more like engineering. Every unit of Bitcoin's UTXO set is on a public chain, unambiguous, not subject to auditor judgment calls about consolidation thresholds. The five companies named in the Nikkei study chose a different approach: obligations that "technically comply" with GAAP while remaining functionally invisible to most investors. That is a choice, not a constraint.

What to Watch

The falsifiable case against the bubble framing is real: if AI generates sufficient revenue to service these obligations before leases commence and roll onto balance sheets, the SPV structure was prudent financing, not a ticking clock. The bull case requires AI revenues to scale dramatically from current levels to cover what the industry has committed to spend. That is the number to watch, not the debt figure in isolation. Oracle's credit trajectory and the pace of data center lease commencements are the near-term tells.

Update, July 24, 2026

The bond market is now pricing what the footnotes already said. Bond investors are demanding yields north of 7% on Meta's $12 billion El Paso data center paper, a marked jump from the terms Meta secured just months ago. The repricing is the real story -- what it costs to raise that $12 billion, not the headline amount.

Capacity Global's analysis calls El Paso "the clearest evidence yet that lenders are starting to differentiate between AI infrastructure deals rather than treating them as a uniform asset class carrying a hyperscaler halo."

The supply backdrop making that repricing inevitable has been building all year. Global debt issuance tied to AI is on track to more than double to nearly $570 billion in 2026 per a Morgan Stanley forecast, and for bond investors the AI buildout is no longer just a stock story -- it is becoming the largest single position in the investment-grade credit market. AI-related issuers had already sold nearly $236 billion of debt globally by May 31, roughly four times the pace of 2025.

Year to date, four of the five hyperscalers have underperformed the broader market on a spread basis, and Man Group's H2 2026 credit outlook argues that despite spread widening, levels still do not adequately compensate investors for the risks given the supply overhang and growing concerns around competition, buildout delays, and regulatory intervention.

A structural mismatch between 25-year debt and five-year assets, repeated across dozens of SPVs now embedded in the sector's capital stack, is a sector-wide exposure.

A key concern is the shift of financing toward less transparent private credit and off-balance-sheet structures, estimated at $800 billion for data centers -- a move that makes potential losses harder to track, as public balance sheets no longer tell the full story of who bears the risk if AI returns are delayed or fall short. The credit market is starting to say out loud what the Nikkei footnote study already quantified.

Update, August 10, 2026

The contagion now has a second vector. While the hyperscaler footnote story has been about customers hiding obligations, Nvidia itself stepped into the same architecture from the supply side. Nvidia's five-year credit default swap surged to a record 82 basis points on Monday, July 27, its largest single-day intraday gain since the contract began actively trading in November 2025, according to ICE Data Services. The trigger was the scale of Nvidia's deal-making: Nvidia takes equity stakes in or provides debt guarantees to customers such as OpenAI and CoreWeave, which then use the capital to purchase Nvidia's own chips. Bloomberg reported that Nvidia had already announced more than $540 billion of such deals in 2026 alone, excluding a potential new OpenAI arrangement.

Separately, Nvidia is in talks to guarantee as much as $250 billion so OpenAI can lease a 10-gigawatt data center campus that a SoftBank subsidiary is developing in southern Ohio. The structural problem is the circular financing loop now embedded in Nvidia's own deal-making: Nvidia writes a check or backstops debt, and the recipient turns around and spends heavily on Nvidia. OpenAI does not hold an investment-grade credit rating, and without that rating the kind of long-term lease and construction debt a campus that size requires is effectively inaccessible from conventional bond markets -- Nvidia is the workaround.

Now Nvidia has moved to syndicate that risk. Nvidia formally announced it has signed memorandums of understanding to mobilize a staggering $500 billion in third-party capital for AI infrastructure, teaming up with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR.

The coalition will "create dedicated pools of capital at significant scale at attractive rates for Nvidia customers," according to a statement Monday. Nvidia CEO Jensen Huang said he approached only the six firms for the commitment, and none turned him down. The framing from Nvidia is bullish. The read from the structure is different. The decision to expand its lending and spread the risk shows there is a limit to how much the world's most valuable company will singlehandedly finance the AI boom.

The $500 billion figure is a mobilization target over time, not a committed pool, and Nvidia's release is explicit that the partnerships remain subject to execution of final agreements.

The move could reignite fears about the circular nature of AI financing -- in which a supplier like Nvidia provides financing or investment capital to its major customers. The fear is that if one major company runs into trouble, it could have a ripple effect through the AI ecosystem.

Both the IMF and the Bank for International Settlements have flagged AI circular financing as a systemic downside risk. The Benzinga timeline of Nvidia's circular financing deals documents how this pattern compounds: Nvidia anchors the SPV, the SPV buys Nvidia chips, the customer leases them back, and the chip revenue that justifies Nvidia's valuation flows through structures that don't appear on anyone's headline balance sheet. Much of the resulting obligation sits outside standard debt disclosure. Michael Burry's three-word summary -- "Around and around we go" -- describes the loop precisely. Now six of the largest private capital pools on earth are being pulled inside it.

Societe Generale's Head of U.S. Equity Strategy, Manish Kabra, distilled the shift bluntly: "For hyperscale computing companies, it's CDS, not EPS, that matters now."

Nvidia shares fell 4.99% on Monday to close at $196.51, stripping roughly $250 billion from the chipmaker's market value and ceding the title of world's most valuable company to Apple. The next forcing function remains Nvidia's Q2 earnings call in late August -- watch whether Jensen Huang addresses guarantee exposure and whether any of the six consortium partners put capital cost disclosures into public filings. The signal is not default risk. It is skepticism about how far the AI ecosystem can stretch balance sheets to keep the boom going. Nvidia gets its first chance to answer on the record on August 26.

Update, August 21, 2026

Michael Burry has now put his name and his money directly on the Enron comparison. On August 14, Burry disclosed on Substack that as of August 11 he views Wall Street's $500 billion AI infrastructure financing deal as a "Wall Street gimmick" reminiscent of Enron's operations. The specific mechanics he flagged go beyond the general circular-financing critique already circulating: "that $500 billion Nvidia Wall Street stunt involves Nvidia taking 25% stakes and providing residual value guarantees on purchase of its chips," he wrote on X, "all filtered through Private Equity's Private Credit schemes."

Burry is not just talking. He has doubled his short position against Nvidia, pointing to the $500 billion financing plan and arguing that complex credit structures are prolonging market momentum.

He doubled down on his criticism in his August 11 Substack post, this time comparing the current period to 2005, right before the housing market crash. His short book has also widened scope: Burry has intensified his bearish stance across the AI and semiconductor sectors, expanding short positions on Nvidia, Micron, and the SOXX ETF, while initiating a new short on Caterpillar.

His short position in Oracle increased at $145 -- notable given Oracle is the most extreme off-balance-sheet case in the original Nikkei study.

The institutional response is also worth tracking. Fundstrat's Tom Lee countered that gross financial obligations routinely exceed actual cash at risk and that top tech firms have flexible exit options on future commitments. But the regulatory lane is opening independently of that debate: Cleveland Fed President Beth Hammack warned that debt-fueled AI investment could stoke inflation, and her remarks suggest awareness is spreading within the Fed that AI infrastructure investment could affect the broader macroeconomy, not just equity market bubbles.

Burry's own framing in his newsletter: "Structuring credit is a natural part of the system. Structuring unnatural credits to prolong momentum late in the bull phase is where the worry comes in."

Update, September 20, 2026

The Oracle footnote risk the article flagged is now showing up as a line item in the loan market. About $18 billion of loans tied to Oracle's Project Jupiter campus in New Mexico moved into stressed territory, with the paper quoted at 89 to 91 cents on the dollar.

The project had secured those loans from a bank consortium late last year to begin construction, but efforts to distribute the debt to a broader investor base are stalling, leaving banks including Santander and Jefferies holding more Oracle-linked project debt on their balance sheets than originally planned.

The debt carries a private investment-grade rating largely because the SPV structure is synthetically boosting the project's overall credit profile -- the same mechanism the original article identified as legally compliant but practically misleading. Project Jupiter was originally planned to run on 2.2-gigawatt gas turbines, but state land authorities rejected the natural gas pipeline application; Oracle pivoted to Bloom Energy fuel cells, and those related permits have drawn similar regulatory scrutiny.

Barclays, in its latest credit report on Oracle, flagged that "the timeline for its data center projects, particularly for those that have faced setbacks such as New Mexico/Jupiter" remains an area of discomfort.

The political risk is compounding the credit risk. New Mexico's Democratic gubernatorial nominee Deb Haaland has pledged to sign an executive order on day one in office to immediately halt new large-scale data center projects while state lawmakers craft guardrails.

A recent poll found 65% of New Mexico respondents opposed large-scale data center development.

Oracle is forecasting up to $95 billion in capital expenditure for fiscal 2027 and counting on up to $25 billion in customer repayments -- a repayment assumption that now hinges on infrastructure that is behind schedule, facing permit fights, and whose financing banks cannot offload.

Update, September 23, 2026

The crossover point the footnote study implied is now a matter of arithmetic, not theory. Epoch AI's data analysis finds that the five largest hyperscalers are increasing cash capex faster than operating cash flow, with aggregate spending on track to overtake inflows around Q3 2026.

Operating cash flow is growing roughly 23% per year while cash capex is expanding at 70% per year, with the two lines crossing near zero aggregate free cash flow right now.

Oracle has already crossed the threshold, Amazon is crossing around now, Alphabet follows around Q1 2027, Meta around Q3 2027, and Microsoft around Q3 2028.

The numbers behind that crossover are getting harder to look away from. J.P. Morgan Asset Management's analysis puts it plainly: AI capex has gone from 33% of hyperscaler cash flow from operations in 2023 to an estimated 93% in 2026.

J.P. Morgan the bank now estimates the five largest hyperscalers will spend $697 billion in 2026, a figure it has raised by $173 billion since January.

PIMCO estimates combined hyperscaler capex will consume roughly 94% of operating cash flow by 2026 to 2027. The gap between what the business generates and what the buildout requires is no longer a rounding error.

The funding shift that gap demands is already visible in the capital structure. FactSet's analysis documents that in the last 18 months the group moved from almost fully self-funded capex to raising external capital at scale, with incremental annual debt rising from 9% of capex in FY24 to 32% by mid-2026; Alphabet priced an $84.75 billion equity raise in June 2026 and Oracle plans $40 billion of combined debt and equity for FY27.

Arete Research founding partner Jim Fontanelli, presenting at a summit titled "A Market Outlook: Is AI in a Capex Bubble?", put the four largest hyperscalers' total underlying cash flow through 2028 at roughly $1.2 trillion, and noted the financing gap widens further when specialized AI cloud companies, enterprises, and governments building their own systems are added to the borrowing queue. The off-balance-sheet obligations the Nikkei study quantified are the shadow; the cash flow ceiling is the wall they are now running into.

Update, October 2, 2026

The loan market just produced its clearest stress signal yet. Syndicate banks including Santander and Jefferies are quoted on roughly $18 billion in loans tied to an Oracle-leased New Mexico data center at 89 to 91 cents on the dollar, per Reuters citing the Financial Times. The facility is Project Jupiter, a New Mexico campus Oracle is leasing to supply compute under its broader OpenAI arrangement. Loans pricing at a double-digit discount during syndication is not a routine haircut -- it is banks signaling they cannot move the paper at par, and it converts the theoretical "footnote risk" in the base story into a live clearing price problem.

The mechanism now running through the system mirrors the secondary securitization pattern that legal analysts at Quinn Emanuel flagged as an emerging litigation vector: lenders who originated data center loans are pooling those loans and selling tranches to a wider investor base through ABS and CMBS structures, the same instrument class that distributed subprime mortgage exposure in 2008. Rajat Rana, who worked on structured finance litigation after the housing crisis, told CNBC the AI data center financing cycle feels like "deja vu," with trillions of dollars moving through structures that carry almost no transparency -- a pattern he called "the largest peacetime investment project in human history, which is financed largely off balance sheet."

The Jupiter loan trouble is also the first visible test of Oracle counterparty concentration risk in the syndicated market. Banks are already grappling with counterparty risk to Oracle after the company backstopped tens of billions of dollars of project finance loans with leases, and JPMorgan projects data center securitization issuance could hit $30 to $40 billion annually across US CMBS and ABS in both 2026 and 2027 -- meaning the pipeline of paper that needs to clear the market is only growing. If syndicates cannot move Oracle-linked paper at par today, the question for every subsequent deal is what the clearing price is, and who ends up holding the residual.

Sources

• Nikkei Asia: Five U.S. tech giants' hidden debts soar to $1.65tn on opaque AI funding
• Bloomberg Law: Big Tech AI Spree Revives Accounting Devices That Toppled Enron
• Moody's Ratings report on data center lease commitments, February 2026
• S&P Global Ratings credit rating action on Oracle, July 9, 2026 (citing AI lease commitments)