While global investors focus on the hundreds of billions of dollars in capital expenditure announced by tech behemoths, a far larger financial undercurrent is quietly building. An analysis of the latest regulatory filings from nine leading technology companies has revealed a staggering reality: these firms hold approximately $3 trillion in off-balance-sheet commitments, most of which are AI-related. This figure is five times the roughly $600 billion in capital expenditures they announced over the past year, and nearly three times their combined existing lease liabilities and long-term borrowings.
These so-called "off-balance-sheet commitments" refer to substantial future payment obligations that have not yet been formally recorded on corporate balance sheets. They include data center lease agreements that have not yet commenced, chip procurement contracts, and other long-term agreements signed to secure AI computing power. The nine companies analyzed include Alphabet (GOOGL.US), Amazon (AMZN.US), Microsoft (MSFT.US), Meta Platforms, Inc. (META.US), Oracle (ORCL.US), NVIDIA (NVDA.US), Broadcom (AVGO.US), SpaceX, and AMD (AMD.US). These hidden obligations are divided into two main categories: approximately $1.2 trillion in lease commitments that have not yet taken effect, and roughly $1.9 trillion in procurement commitments for chips and data center infrastructure. This scale has ballooned about fourfold in just one year.
Take Meta Platforms, Inc.'s "Hyperion" super data center project in Louisiana, which spans an area roughly equivalent to 1,700 football fields, as an example. Although Meta Platforms, Inc. is responsible for building and operating the data center, it does not appear on the company's balance sheet. The majority of the project's equity is held by funds managed by Blue Owl Capital, which raised approximately $27 billion in construction funding through bond issuance. Meta Platforms, Inc., as a minority shareholder and future tenant, plans to lease Hyperion starting in 2029, with an initial lease term of four years and an option to extend up to 20 years. If Meta Platforms, Inc. fails to fulfill the entire 20-year lease, it would need to compensate bond investors for their losses. As of the end of June, Meta Platforms, Inc.'s total lease commitments not yet in effect had reached $347 billion.
Beyond data center leases, long-term procurement agreements for AI chips also constitute a massive off-balance-sheet burden. Due to tight AI hardware supply, companies typically need to sign procurement contracts with suppliers years in advance. Alphabet's off-balance-sheet commitments have grown the most dramatically. As of June 30, the company's procurement commitments and contractual obligations surged from $332 billion three months earlier to $811 billion. Alphabet stated that these commitments primarily relate to "technology infrastructure and inventory" and "agreements to secure energy supply for data centers," with some energy procurement agreements extending as far as 2054. The company did not provide a detailed explanation for why this figure grew by nearly $500 billion in just one quarter.
Where the hidden leverage lives
The core mechanism behind this off-balance-sheet financing is the special purpose vehicle (SPV). Technology companies establish a separate legal entity with private lenders, which owns assets such as land, buildings, and power supply. The SPV is responsible for borrowing, while the tech company signs long-term leases to lock in capacity. Institutional investors including PIMCO, BlackRock, Apollo, and Blue Owl Capital, as well as banks like JPMorgan, provide debt and equity financing, and these loans do not appear on the parent company's balance sheet. Unlike traditional capital expenditures, these obligations, under current accounting standards, can temporarily avoid being recognized as liabilities until the assets are actually delivered or services truly commence. They do not show up on the main balance sheet but are instead hidden in the footnotes of financial reports in small print. This practice is legal and compliant, yet it makes it difficult for investors to discern a company's true total leverage from traditional financial metrics. This accounting treatment is legitimate: long-term leases and procurement obligations only need to be disclosed in financial statement footnotes until the facilities begin operations or chips are delivered, at which point they will appear in full on the balance sheet. This time lag allows companies to maintain low debt ratios and intact credit ratings during the most capital-intensive construction phase.
"Big Short" investor issues triple-compression warning
Investor Michael Burry, known for "The Big Short," has issued a stern warning on this front. He points out that hyperscale cloud providers are using excessively long depreciation periods for AI chips and servers, setting useful lives of five to six years when the actual economic cycle is closer to two to three years. He estimates this practice could suppress depreciation expenses by $176 billion between 2026 and 2028. Burry also warns of a triple "compression" risk: declining AI demand, falling earnings as expenses catch up, and tightening financing. He recently disclosed short positions in Oracle and Nebius, noting that circular financing among hyperscale cloud providers, AI labs, and chipmakers could artificially inflate demand and revenue.
Deeper risks: equity investments and residual value guarantees
Beyond leases and procurement commitments, off-balance-sheet risks are spreading into more complex financial engineering. In terms of equity investment commitments, NVIDIA has pledged to invest $27 billion in equity investments by the end of its fiscal year between April 2026 and January 2027. "Residual value guarantees" (RVGs) have emerged as a new risk vehicle. This structural arrangement allows large corporations to leverage their higher credit ratings to help customers lower financing costs. Specifically, an SPV borrows to purchase chips, with the loan supported by contract cash flows from the company that will use the chips. If a funding gap ultimately exists, the guarantor makes up the difference. Broadcom has extended this logic into chip financing, guaranteeing a $35 billion debt deal in which Apollo Global Management and Blackstone purchase custom AI chips and lease them to Anthropic. NVIDIA has also indicated it may provide up to a 25% residual value support mechanism for related financing opportunities. Investors are already paying attention to the approximately $70 billion in "shadow liabilities" that do not appear on the balance sheets of AI companies. Mariya Entina, a portfolio manager at DoubleLine, puts it bluntly: "This is like exploiting a loophole in the system... we are entering an era of financial engineering."
Wall Street's financing machine: fueling the boom or seeding the next bust?
The Bank for International Settlements recently warned that this financing model is reminiscent of 19th-century railway speculation and the internet bubble, with current amounts already far exceeding those periods. The Financial Stability Board found that AI projects accounted for more than one-third of all private lending in 2025, up from 17% in the previous five years. Allianz Research pegs investment intensity at 34% of revenue, more than double the 15% peak during the internet bubble era, which stood at $205 billion. Sequoia Capital estimates the annual gap between investment and revenue growth is approximately $600 billion. Morgan Stanley analysts warned as early as April: "As these off-balance-sheet commitments become more frequent, larger, and more complex, it becomes increasingly difficult for investors to assess a company's underlying total leverage." This $3 trillion off-balance-sheet arms race is pushing the financial risks of tech giants into unprecedented uncharted waters.
Wall Street investment banks are responding in kind with multi-trillion-dollar financing plans. JPMorgan launched a $1.5 trillion "Security and Resilience Initiative" in October 2025, and Morgan Stanley followed with a similarly sized "U.S. Innovation Infrastructure Initiative." Bank of America has also joined the fray, announcing $250 billion in commitments over 18 months. Taken together, a troubling picture is emerging: tech giants are hiding true leverage through off-balance-sheet vehicles, while Wall Street investment banks fuel that leverage through financial engineering. This brings to mind 2008, when banks hid high-risk assets off-balance-sheet through structured investment vehicles (SIVs), while credit rating agencies labeled those assets as AAA. In essence, Wall Street banks are acting as matchmakers and architects for AI infrastructure financing, channeling global capital into AI infrastructure through bond issuance, syndicated loans, and private credit. The question is: when Wall Street's "financing machine" becomes deeply intertwined with tech giants' "off-balance-sheet empire," who bears the ultimate credit risk?
Three lessons from the subprime crisis echoing in the AI era
The first lesson is that the invisibility of off-balance-sheet leverage masks real risk. In the subprime crisis, the problem with SIVs was that risk was hidden, and investors could not see the true level of leverage until asset prices collapsed and SIVs were forced to bring assets back onto bank balance sheets. In today's AI infrastructure, the $3 trillion in off-balance-sheet commitments is similarly in a "semi-transparent" state. When AI demand remains strong, these commitments are just numbers in financial footnotes; but if demand weakens, leading to rising data center vacancy rates or accelerated hardware depreciation, these off-balance-sheet liabilities will quickly convert into on-balance-sheet losses. As Michael Burry, the inspiration for "The Big Short," has warned, AI hardware is depreciating at roughly 50% per year, while tech companies are "creating the illusion of immediate profit inflation" by extending depreciation periods.
The second lesson is the systemic failure of credit ratings and risk pricing. Before the subprime crisis, rating agencies rated large volumes of subprime mortgage securities as AAA because models assumed house prices would never fall nationally. Today, AI infrastructure financing faces a similar risk-pricing failure: circular financing is artificially inflating demand. The loop of NVIDIA investing in AI companies, those companies buying NVIDIA GPUs, NVIDIA's revenue growing, leading to more investment, bears a striking resemblance to the subprime cycle of "banks issuing loans, packaging them into CDOs, selling them, and freeing up capital to issue more loans." Residual value guarantees assume AI chips will retain sufficient residual value after lease expiration, but if the iteration speed of AI chips far exceeds expectations, these guarantees will become real liabilities. Credit spreads are narrowing, and investors' risk premium pricing on AI debt may be severely inadequate. As the DoubleLine portfolio manager put it, "We are entering an era of financial engineering."
The third lesson is the contagion risk of "too big to fail." In the subprime crisis, the collapse of institutions like AIG triggered a systemic crisis because risk was highly interconnected across financial institutions through a complex derivatives network. Today, the risk network is similarly highly interconnected: tech companies are linked to each other through off-balance-sheet vehicles, Wall Street banks are deeply involved through financing arrangements, and pension funds, insurance companies, and mutual funds become the ultimate risk bearers by purchasing AI bonds. If AI demand experiences a substantial slowdown, risk will rapidly transmit along the chain of "tech company to SPV to bondholder to financial institution." As one market observer noted, "If demand for AI products suddenly weakens, or if profitability takes longer than investors expect, the resulting adjustment could send economically significant shockwaves around the world." Moreover, the systemic importance of current AI participants goes without saying. Amazon, Microsoft, Google, and Apple are deeply embedded in the global economy, with their stocks widely held in 401(k) retirement plans and index funds. The consequences of their failure could far exceed those of the 2008 financial crisis. This is no longer a question of "whether to rescue" but "whether we can afford to rescue." During the subprime crisis, the U.S. government stabilized the financial system with the $700 billion Troubled Asset Relief Program (TARP). Today, the off-balance-sheet commitments of just nine tech companies stand at $3 trillion, and combined with Wall Street banks' multi-trillion-dollar financing plans, the total potential risk exposure of AI infrastructure could reach $5 to $10 trillion. If these risks were to materialize simultaneously, no single rescue plan would be sufficient to cover them.
Comments