Five of the world’s largest technology companies report combined debt of roughly $1.35 trillion. Look past the official ledgers, and another $1.65 trillion in obligations tied directly to artificial intelligence infrastructure comes into view. That hidden pile has swelled eightfold in four years. It now exceeds the visible borrowings. And it sits largely off the books through structures that echo tactics from corporate scandals two decades ago.
Off-Balance-Sheet Structures Mask AI’s True Cost
The numbers come from a detailed examination of regulatory filings by Nikkei Asia. Alphabet, Amazon, Meta, Microsoft and Oracle have signed long-term leases for data centers not yet running. They hold forward contracts for massive deliveries of graphics processing units and servers. Some rely on joint ventures funded by private credit that keep associated borrowing away from consolidated accounts. Meta alone carries about $420 billion in such commitments, triple its reported debt. Oracle’s off-balance-sheet figure has jumped thirtyfold since 2022.
These arrangements let the companies substitute large upfront capital expenditures with future operating expenses. The debt services those leases and offtake agreements. Yet it stays with special purpose vehicles or variable interest entities that the hyperscalers do not fully consolidate. Accounting rules permit this when the parent does not hold primary beneficiary status or direct the entity’s activities. The result? Cleaner balance sheets that flatter leverage ratios and return metrics. Investors see less than half the picture.
But the exposure remains real. The companies still design the facilities, operate them, occupy most of the space and often guarantee performance or residual value. “The accounting treatment itself is in fashion,” technical accounting consultant Tom Selling told Bloomberg Tax. “But what if one of these companies was a house of cards and was propping itself up with this accounting treatment? To me, that’s the risk.”
Analysts at the Bank for International Settlements reached similar conclusions months earlier. In its March 2026 Quarterly Review, the BIS described these deals as “shadow borrowing.” Hyperscalers partner with private credit funds and institutional investors. A dedicated vehicle raises debt through private placements. The tech company takes a minority stake, signs multi-year leases and may extend credit support. Debt stays off the hyperscaler’s books. Cash flows from leases repay it. BIS Quarterly Review authors Egemen Eren, Ingomar Krohn and Karamfil Todorov noted that such structures channel private credit into AI data centers while creating new channels for financial shocks. Refinancing problems at the vehicle level or sudden drops in private credit appetite could force guarantees to kick in.
Corporate bond markets have already felt the strain. Hyperscalers issued more than $100 billion in long-term debt in 2025 alone. Credit default swap spreads widened, particularly for those with lower credit ratings. Uncertainty over whether massive AI projects will deliver sufficient returns drove the move. Yet bond issuance covers only part of the tab. Off-balance-sheet vehicles handle the rest. Total visible and hidden obligations for the five companies approach $3 trillion.
This pattern revives memories of Enron. That energy trader used special purpose entities to hide debt and inflate profits until its 2001 collapse. Rules tightened afterward. Variable interest entity consolidation standards grew stricter. Still, judgment calls remain. Companies must assess whether they control the entity or bear the majority of expected losses and returns. Tech firms argue they do not meet the tests for many data center vehicles. Auditors and regulators accept those assessments for now. But the structures look remarkably similar.
Meta, for example, formed a joint venture with private credit firm Blue Owl Capital for a Louisiana data center project. The venture could cost up to $250 billion. Meta’s maximum exposure sits at $46 billion while the vehicle carries $27 billion in debt. That debt does not appear on Meta’s balance sheet. Alphabet maintains variable interest entities for leases and credit backstops on data centers and power infrastructure. It discloses them but does not consolidate. Microsoft offers scant detail on its variable interest entities beyond stating it does not consolidate them. Oracle reports $260 billion in future lease commitments and a $3.3 billion backstop on one arrangement.
These commitments activate once facilities come online. Demand for AI services must materialize to justify the expense. Many leases run for 10 to 15 years or longer. Power purchase agreements lock in electricity costs regardless of utilization. If returns disappoint, writedowns could hit suddenly. Equity dilution from share sales already under way at some firms adds pressure on valuations.
Four of the five companies prepared to release second-quarter earnings in late July 2026. Analysts planned to scour footnotes and management commentary for fresh clues. Yet the off-balance-sheet totals receive limited attention in headlines. Standard credit models still focus on reported debt. That leaves a blind spot.
Private credit funds and insurers now hold large slices of this exposure. Banks provide revolving credit lines to the vehicles. Interconnections multiply. A downturn in AI investment sentiment could ripple through non-bank lenders and back to the hyperscalers through guarantees. The BIS warned of procyclical swings in private credit appetite that might amplify stress.
Tech executives maintain confidence. Capital spending continues at record pace. Data center construction shows no sign of slowing. Nvidia’s order book stays full. Yet the financial architecture supporting the boom relies on optimistic assumptions about future AI revenue. Power constraints, chip obsolescence after three to four years and unproven long-term software margins complicate the math.
Recent coverage underscores the growing unease. The Next Web highlighted how the same accounting devices that felled Enron now bankroll the AI data-center surge. Discussions on X echoed the figures within hours of the Nikkei report, with users noting Meta’s hidden obligations alone rival the balance-sheet debt of many large corporations.
Regulators watch closely. The Securities and Exchange Commission has pressed for clearer disclosures on control judgments. Auditors flag the subjectivity involved. Still, current rules give companies latitude. As long as AI hype sustains investor appetite, the structures persist. Should demand falter, those hidden obligations will move onto balance sheets at the worst possible moment.
The scale surprises even seasoned observers. Four years ago the off-balance-sheet total stood at roughly $200 billion. Today’s $1.65 trillion reflects an industry sprinting to secure compute capacity before competitors. Contracts for GPUs often precede actual delivery by years. Lease commitments cover facilities still under construction. The obligations exist. They simply wait for accounting triggers.
Comparisons to utilities emerge in private conversations. Hyperscalers increasingly resemble power companies with fixed, long-duration assets and regulated-like returns on compute. Yet they trade at technology multiples. The mismatch grows as capital intensity rises.
Investors face a choice. Accept management’s view that these commitments represent prudent capacity reservations with strong upside. Or view them as aggressive bets masked by accounting that could unravel if AI adoption slows. The coming earnings season will test which narrative holds. Footnotes deserve more than a glance.
WebProNews is an iEntry Publication