Nine major US tech giants were revealed to be hiding 3 trillion US dollars of “off-balance sheet bills”. Is the “subprime mortgage crisis” alert in the AI era sounding?
The Zhitong Finance App learned that as global investors focus on the capital expenditure of tech giants, which can easily reach hundreds of billions of dollars, a larger financial undercurrent is quietly accumulating. An analysis of notes to nine top tech companies' latest regulatory documents revealed a shocking fact: these companies have about $3 trillion in off-balance sheet commitments, most of which are AI-related — five times their total capital expenditure announced in the past year (around $600 billion), and about three times the total amount of existing lease liabilities and long-term loans.
The truth about “promises outside the balance sheet”: the 3 trillion that disappeared from the profit sheet
The so-called “off-balance sheet commitments” refer to large expenses that have not been officially included in the company's balance sheet but may be paid in the future, including data center leasing agreements and chip purchase contracts that have not yet been implemented, and other long-term agreements signed to lock in AI computing power. The nine companies analyzed this time include Alphabet (GOOGL.US), Amazon (AMZN.US), Microsoft (MSFT.US), Meta (META.US), Oracle (ORCL.US), Nvidia (NVDA.US), Broadcom (AVGO.US), SpaceX (SPCX.US), and AMD (AMD.US).
These hidden obligations are mainly divided into two major sectors: about $1.2 trillion in lease commitments that have not yet been implemented, and about $1.9 trillion in procurement commitments for chips and data center infrastructure. This scale has increased about four times in just one year.
Take Meta's “Hyperion” super data center project in Louisiana, as an example. The project covers an area equivalent to 1,700 soccer fields. Although Meta is responsible for building and operating the data center, it does not appear on Meta's balance sheet. Most of the equity in the project is held by funds managed by Blue Owl Capital, which raised approximately $27 billion in construction capital through the issuance of bonds. As a minority shareholder and future tenant, Meta plans to rent Hyperion starting in 2029 for an initial period of 4 years with the option to extend to 20 years. Once Meta fails to fulfill its entire 20-year lease, it will need to reimburse the bond investors for their losses. By the end of June, Meta's total unimplemented lease commitments had reached $347 billion.
In addition to data center leasing, long-term procurement agreements for AI chips also pose a huge off-balance sheet burden. Due to the tight supply of AI hardware, companies usually need to sign procurement contracts with suppliers several years in advance.
Alphabet's off-sheet promises have grown the most astonishing. As of June 30, the company's procurement commitments and contractual obligations had jumped from $332 billion three months earlier to $811 billion. Alphabet said these commitments mainly involved “technical infrastructure and inventory” and “agreements to ensure data center energy supply,” with some energy purchase agreements even continuing until 2054. The company did not explain in detail why this figure grew by nearly $500 billion in just one quarter.
SPV architecture: a “legal tool” for off-balance sheet financing
The core mechanism for this type of off-balance sheet financing is a special purpose vehicle (SPV) — a technology company and a private lender establish an independent legal entity that owns assets such as land, buildings, and electricity supply; SPVs are responsible for loans, and technology companies sign long-term leases to lock in production capacity. Institutional investors such as PIMCO, BlackRock, Apollo, Blue Owl Capital, and banks such as J.P. Morgan Chase provide debt and equity financing, and these loans do not appear on the parent company's balance sheet.
Unlike traditional capital expenses, these obligations can be temporarily not recognized as liabilities until assets are actually delivered or services actually commence, in accordance with current accounting standards. They do not appear on the main balance sheet; they are only hidden in financial notes in the form of small footnotes. This kind of operation is legal and compliant, but it makes it difficult for investors to see the true level of a company's total leverage from traditional financial indicators.
This accounting treatment is legal: long-term leases and procurement obligations only need to be disclosed in notes to financial statements before the facility starts operating or chips are delivered, and then fully appear on the balance sheet. This time difference allows companies to maintain low debt ratios and complete credit ratings during the most capital-intensive construction phase.
“Big Short” Bury Warns: Risk of Triple Compression
“Big Short” investor Michael Berry issued a stern warning about this. He pointed out that hyperscale cloud service providers use excessively long depreciation periods for AI chips and servers, but when the actual economic cycle is closer to two to three years, they set the service life to five to six years. He estimates that this approach could reduce depreciation costs by $176 billion between 2026 and 2028.
Burry also warned of the risk of triple “compression”: falling demand for AI, falling profits as costs catch up, and tightening financing. He recently revealed short positions between Oracle and Nebius, and pointed out that revolving financing between hyperscale cloud service providers, AI labs, and chip makers may artificially drive demand and revenue.
Deeper hidden dangers: equity investment guarantees and “residual value guarantees”
In addition to leasing and procurement commitments, off-balance sheet risks also spread to the more complex field of financial engineering. In terms of equity investment commitments, Nvidia promised to invest 27 billion US dollars in equity investments from April 2026 to the end of the fiscal year in January 2027.
The “Residual Value Guarantee” (RVG) has become a new risk carrier. This structured arrangement allows large companies to use their higher credit ratings to help customers reduce financing costs. Specifically, a special purpose entity (SPV) borrows to buy chips, and the loan is supported by the contractual cash flow of companies that will use the chip in the future. Should there eventually be a funding gap, the guarantor will make up for the difference.
Broadcom has extended this logic to the field of chip financing — providing guarantees for a $35 billion debt transaction, with institutions such as Apollo Global Management and Kuroishi investing in the purchase of custom AI chips and then leasing them to Anthropic for use. Nvidia also said it may provide a residual value support mechanism of up to 25% for relevant financing opportunities.
Investors are already watching these “shadow liabilities” of about $70 billion, which do not appear on AI companies' balance sheets. Mariya Entina, portfolio manager at DoubleLine (DoubleLine), put it bluntly: “It's like exploiting a gap in the system... We are entering an era of financial engineering.”
Regulatory Alerts
Wall Street's “financing machines” are powering the AI era, but they may also be planting the seeds for the next financial crisis.
Similar to 19th century railway speculation
The Bank for International Settlements recently warned that this type of financing model is reminiscent of railway speculation and internet bubbles in the 19th century, and that current amounts far exceed the scale of those periods. The Financial Stability Council found that AI projects accounted for more than one-third of all personal loans in 2025, up from 17% in the previous five years.
Allianz research set investment intensity at 34% of revenue, which is more than double the 15% peak during the internet bubble (that is, 205 billion US dollars). Sequoia Capital estimates that the gap between investment and revenue growth is around $600 billion a year.
As early as April, analysts at Morgan Stanley warned: “As these off-balance sheet commitments become more frequent, larger, and more complex, it is becoming more difficult for investors to assess a company's potential total leverage ratio.” This $3 trillion off-balance sheet arms race is pushing the financial risks of tech giants into uncharted waters that have never been seen before.
Wall Street banks are involved
Wall Street investment banks are responding simultaneously with trillions of dollars in financing plans. J.P. Morgan Chase launched the $1.5 trillion “Security and Resilience Initiative” in October 2025, and Morgan Stanley followed suit by announcing the “American Innovation Infrastructure Initiative” of the same scale. Bank of America also joined the fray, announcing an investment of 250 billion US dollars over 18 months.
Looking at these two things together, an unsettling picture is emerging: tech giants use off-balance sheet tools to hide real leverage, and Wall Street investment banks fuel these levers through financial engineering. This reminds me of 2008 — when banks hid high-risk assets from the balance sheet through structured investment tools, and credit rating agencies labeled them AAA.
Essentially, Wall Street investment banks act as matchmakers and architects for AI infrastructure financing, introducing global capital into AI infrastructure through tools such as bond issuance, syndicated loans, and private equity credit. The question is: when Wall Street's “financing machine” is deeply coupled with the tech giants' “off-the-shelf empire,” who bears the ultimate credit risk?
Are the three lessons of the subprime mortgage crisis being repeated in the AI era?
Lesson 1: The “invisibility” of off-balance sheet leverage conceals real risk
In the subprime mortgage crisis, the problem with SIV was that risk was hidden — investors couldn't see the true level of leverage until asset prices crashed, and SIV was forced to re-add assets to bank balance sheets.
In today's AI infrastructure, the $3 trillion off-balance sheet promise is also in a “semi-transparent” state. As demand for AI continues to be strong, these promises are just numbers in the financial footnote; once weak demand causes data center vacancy rates to rise or hardware depreciates at an accelerated pace, these off-balance sheet liabilities will quickly turn into in-statement losses. As Michael Berry, the “big short” prototype, warned, AI hardware is depreciating at a rate of about 50% per year, while technology companies are “creating the illusion of instant profit expansion” by extending the depreciation period.
Lesson 2: The “systemic failure” of credit ratings and risk pricing
Before the subprime mortgage crisis, rating agencies rated a large number of subprime mortgage securities as AAA because the model assumed that housing prices would never fall nationwide.
Today, AI infrastructure financing is facing similar risk pricing failures:
The revolving finance model is artificially boosting demand. Nvidia invests in AI companies → AI companies buy Nvidia GPUs → Nvidia's revenue growth → more investment — this closed loop is similar to the cycle of “banks issue loans → package and sell in CDO → release capital to issue more loans” in the subprime mortgage crisis.
The “residual value guarantee” assumes that the AI chip still has sufficient residual value after the lease expires — but if the AI chip is iterated much faster than expected, these guarantees will become real liabilities.
Credit spreads are narrowing, and investors may be seriously underpricing the risk premium on AI debt. As Dual Tier Capital's portfolio manager said, “We are entering an era of financial engineering.”
Lesson 3: Risk of infection “too big to fail”
In the subprime mortgage crisis, the collapse of AIG and other institutions triggered a systemic crisis because risks were highly interconnected among financial institutions through complex derivatives networks.
Today, risk networks are also highly interconnected: technology companies are interconnected through off-balance sheet tools, Wall Street investment banks are deeply involved through financing arrangements, and pensions, insurance companies, and mutual funds are the ultimate risk-takers by purchasing AI bonds. Once AI demand substantially slows down, risk will rapidly spread along the “technology company → SPV → bondholder → financial institution” chain. As one market watcher said, “If demand for artificial intelligence products suddenly weakens, or profit times exceed investors' expectations, the ensuing adjustments could cause far-reaching economic shocks.”
Furthermore, there is no need to say much about the current systemic importance of AI participants. Tech giants such as Amazon, Microsoft, Google, and Apple are deeply rooted in the global economy, and their shares are widely held in 401 (k) retirement plans and index funds. The consequences of their collapse are likely to extend far beyond the 2008 financial crisis.
This is no longer a question of “do you want to save”, but rather a question of “can't be saved”. At the time of the subprime mortgage crisis, the US government stabilized the financial system through the $700 billion Problem Asset Relief Program (TARP). Today, the nine technology companies' off-balance promises are as high as $3 trillion. Coupled with the Wall Street Investment Bank's multi-trillion dollar financing plan, the potential risk exposure for the entire AI infrastructure could be as high as 5 to 10 trillion US dollars. When these risks are simultaneously exposed, it is difficult for any single relief plan to cover.