The $3 Trillion AI Debt Hiding Off-Balance-Sheet

Overview

The visible capital expenditure driving the artificial intelligence buildout across America’s largest technology companies stands at approximately $600 billion annually. However, a deeper analysis of securities filings reveals a far more significant figure lurking beneath the surface: roughly $3 trillion in off-balance-sheet AI infrastructure commitments by nine major firms.

This hidden obligation is roughly triple the total amount these same companies owe under existing leases and long-term borrowings. While headline numbers suggest a manageable pace of investment, this gap between reported figures and actual commitments represents where the true financial risk resides for shareholders and creditors alike.

Most of this $3 trillion never appears on corporate balance sheets due to specific accounting treatments for two distinct types of promises. The larger portion, totaling about $1.9 trillion, consists of purchase commitments—long-term contracts to buy chips and infrastructure that are not recorded as liabilities until the goods physically arrive.

The remaining $1.2 trillion involves signed leases that have not yet commenced. Under current accounting rules, a lease only becomes a recognized asset and liability once operations begin. Consequently, a data-center lease scheduled to start in 2029 registers no financial impact today, yet these non-started lease obligations have quadrupled in just the past year.

Structural Leverage and Off-Balance-Sheet Strategies

Companies like Meta and Oracle have increasingly utilized structured partnerships to shift AI debt to third parties, effectively hiding leverage from public shareholders. This strategy allows them to expand capacity without inflating their reported debt ratios, though it transfers financial risk elsewhere.

The Meta-Blue Owl Model

Meta’s recent Hyperion campus project in Louisiana illustrates how a multi-billion-dollar data center can leave almost no trace on the owner’s balance sheet. In a joint venture with private-credit giant Blue Owl, Meta owns only a 20% stake while contributing land and construction-in-progress assets. Blue Owl retains 80% ownership, funding the project through bonds sold to investors like PIMCO.

Rather than purchasing the facility, Meta signs an operating lease to use the completed campus. Because this is classified as an operating lease, Meta recognizes no asset and no leasing liability, keeping the roughly $27 billion development cost entirely off its books.

"The debt is real. It sits with PIMCO's bondholders and Blue Owl's funds, off Meta's books. The party holding the downside of the growth story is private credit, not the shareholders reading the income statement."

This structure is mutually beneficial but carries hidden risks. To secure a cheaper lease rate, Meta provided a capped residual-value guarantee for the first sixteen years. However, Meta retains the option to vacate the campus without paying the difference on outstanding obligations if demand weakens. While this shields Meta’s balance sheet, it exposes third-party lenders to the downside of the AI boom.

Oracle’s On-Balance-Sheet Approach

In contrast, Oracle chose a different path, financing a larger portion of its AI infrastructure directly through its own books. This approach offers greater transparency but comes with visible financial strain. Data from early October 2026 shows Oracle carrying approximately $236 billion in total debt against only $67 billion in book equity.

This results in a gross debt-to-equity ratio nearing 3.5 times, coupled with negative free cash flow of roughly $29 billion. While Oracle’s strategy makes its leverage apparent, it highlights the potential long-term sustainability issues associated with heavily indebted balance sheets compared to the off-balance-sheet models favored by peers.

Circular Financing and Demand Risks

The sustainability of these massive commitments depends heavily on continuous, robust demand for AI infrastructure. However, investor Michael Burry has raised alarms regarding circular financing risks, suggesting that a significant portion of hyperscaler spending may be artificially inflating each other’s revenue streams.

The Nvidia Connection

Burry argued in August 2026 that approximately $879 billion of hyperscaler purchase commitments circle through Nvidia. His analysis suggests that the same capital appears as revenue for one company and then re-emerges as a customer order funded by another. For instance, Oracle has committed roughly $300 billion in purchases tied to OpenAI, while OpenAI has committed about $90 billion to AMD, all while holding direct equity stakes in Nvidia.

While Nvidia’s chief executive has dismissed the circular-financing label as "ridiculous," and Burry is known for his skeptical stance, the underlying structure is visible in public filings. The companies financing the AI buildout are routinely each other’s largest customers, meaning a single demand disruption could echo simultaneously across several income statements.

The Cost to Shareholders

The primary cost to shareholders is the discrepancy between apparent and actual leverage. On reported metrics, the five biggest hyperscalers appear lightly borrowed. FactSet notes that while incremental debt has climbed from about 9% of capex two years ago to nearly a third recently, gross leverage remains around 1x EBITDA for four of the five firms.

This comfort is fragile and depends on off-balance-sheet commitments remaining unrecorded. A narrower tally of only unrecognized lease-related commitments across these five hyperscalers amounts to roughly $820 billion, a separate but significant figure that underscores the extent of hidden obligations. If demand for AI services falters, these non-cancelable leases and chip orders could rapidly transform from assets into heavy liabilities, forcing substantial write-downs.