Five Tech Giants’ Off-Balance-Sheet AI Debt Surges 90% in Half a Year to $3.2 Trillion, Hidden Obligations Reach 80% of Total Assets

On November 4, 2026, according to tracking by Nikkei Asia, five U.S. tech giants—Alphabet (Google’s parent company), Microsoft, Amazon, Meta and Oracle—saw their off-balance-sheet AI-related obligations surge from $1.65 trillion to about $3.2 trillion in just six months, an increase of nearly 90%. Although these figures are recorded in footnotes to U.S. Securities and Exchange Commission (SEC) disclosures in a compliant manner, they do not appear on the main balance sheets that investors read first. The total off-balance-sheet AI obligations of the five companies are already equivalent to about 80% of their total assets.

To understand the structural logic behind these figures, it is first necessary to clarify what “off-balance-sheet obligations” actually mean. Unlike traditional loan debt, the core spending in the AI computing arms race—long-term leases for data centers, multi-year pre-purchase agreements for chips such as Nvidia GPUs, and cloud capacity take-or-pay contracts bundled with computing power—can, under the framework permitted by accounting standards, be excluded from the liabilities section of the balance sheet. These commitments are equally real and equally require future cash payments; they are simply buried in the footnotes of financial reports under the names “operating lease obligations” or “future purchase commitments.”

Meta is the most concentrated exposure among the five companies. According to aiweekly’s interpretation of the data, Meta alone has about $420 billion in hidden off-balance-sheet AI obligations, while its disclosed long-term debt on the books is about $58.7 billion—more than seven times the latter. Meta’s Hyperion data center project in Louisiana is a typical case of this structure: Meta and Blue Owl Capital jointly injected capital into an independent structure, which took on the debt and built the center; Meta, as the sole tenant, committed to payments, but argued that it did not need to consolidate the related debt because it did not hold a controlling interest. This logic holds within the boundaries of accounting rules, but for investors who only look at the main balance sheet, Meta’s true financial obligations are systematically underestimated.

Oracle’s case more directly demonstrates the real-world consequences of such hidden obligations. According to reports by edgarstat and other media, in July 2026, S&P downgraded Oracle’s credit rating to BBB-, and explicitly cited Oracle’s future lease obligations of about $260 billion in its rating report—a signal that off-balance-sheet AI commitments are being incorporated into the credit analysis frameworks of mainstream rating agencies. BBB- is the lowest investment-grade tier; one notch below is “junk,” meaning Oracle’s financing costs will rise materially.

Such financial structures are not without historical precedent. Several independent analytical institutions have drawn parallels between the current situation and two historical episodes: first, Enron in 2001, which moved liabilities off its books through special purpose entities (SPVs) and ultimately went bankrupt when the truth of its finances collapsed; second, the global telecommunications bubble at the turn of the century, when major carriers used long-term capacity take-or-pay contracts to make mutual bets on bandwidth demand and ultimately left behind a large number of defaulted obligations when demand failed to materialize as expected. The common feature of both episodes is that off-balance-sheet obligations were treated as “strategic positioning” rather than “financial risk” during the boom, until liquidity pressure forced them back onto the balance sheet in an uncontrollable manner.

There is a structural difference between the current situation and the telecom bubble: the five companies themselves are among the strongest cash-flow companies in the world, and there is no short-term solvency crisis. But this does not mean the problem of information asymmetry is irrelevant. The real risk lies at two levels: first, when the return timeline for AI infrastructure investment is stretched out—whether because AI commercialization lands more slowly than expected or because a major shift occurs in the technology path—these fixed obligations still must be paid according to contract, which will compress free cash flow and affect valuations. Second, for small and medium-sized enterprise users and downstream cloud service providers whose follow-on logic depends on “the healthy balance sheets of the big tech companies,” overestimating the degree of these companies’ true financial flexibility may affect judgments about supplier stability.

From the perspective of the competitive landscape, this structural expansion itself constitutes a competitive barrier: Alphabet, Microsoft and Amazon have locked in computing power supply in advance by bundling multi-year computing contracts, which in effect also raises the ceiling on competitors’ ability to expand capacity on short notice. But this double-edged sword also means that these five companies have bet much of their strategic flexibility for the coming years on the premise of sustained high demand for AI computing power, and the cost of transformation will rise accordingly.

For enterprise users and developers, this means that when purchasing multi-year computing contracts, they need to incorporate suppliers’ true burden of obligations into risk assessments. The commitments of counterparties to long-term agreements may see changes in performance priority and pricing flexibility when they themselves face liquidity adjustment pressure.