SPV (special purpose vehicle) off-balance-sheet structures are a core financing tool in the AI data-center buildout

They allow hyperscalers to expand capacity rapidly while keeping the bulk of associated debt off their consolidated balance sheets, converting what would be large capital expenditures into multi-year operating leases or offtake commitments. This creates meaningful “shadow” leverage and interconnected risks.

How the Structures Typically Work

A dedicated vehicle (SPV, joint venture, or variable interest entity) is formed, often with a private-credit or infrastructure partner (e.g., Blue Owl, Apollo, Blackstone, Pimco). The vehicle owns or develops data center assets and raises most of the debt (plus some equity). The hyperscaler typically holds a minority equity stake (around 20%), signs long-term operating leases or capacity commitments, and may provide residual-value or borrowing guarantees. If the hyperscaler is not the primary beneficiary, the SPV’s debt stays off its balance sheet. Lease payments appear as operating expenses; many commitments are disclosed in footnotes. The Bank for International Settlements calls this “shadow borrowing”, obligations economically equivalent to debt but outside corporate balance sheets, channeling private credit into AI infrastructure and linking hyperscalers with non-bank investors.

Scale of the Exposures (as of mid-to-late 2026)

As of mid-to-late 2026, off-balance-sheet exposures for major tech firms are substantial relative to on-balance-sheet debt. Moody’s estimated around $662 billion in data-center lease commitments for the five largest U.S. hyperscalers, comparable to or exceeding their adjusted debt. A Nikkei analysis placed combined off-balance-sheet AI obligations for Alphabet, Amazon, Meta, Microsoft, and Oracle at about $1.65 trillion, versus roughly $1.35 trillion in reported debt. Meta’s share was particularly high, around $420 billion. Broader tallies, including uncommenced leases and purchase commitments across a wider group of tech firms, have reached approximately $3 trillion. Specific examples include Meta’s Hyperion SPV in Louisiana, involving tens of billions in private-credit debt, with Meta as the primary tenant and residual-value support. Similar structures are used by Oracle and Google-linked vehicles.

Primary Risks

The primary risks associated with hyperscaler data center financing include opacity and understated leverage, as standard balance-sheet metrics and leverage ratios often fail to capture true economic commitments.

Footnote disclosures exist but receive less scrutiny than headline debt figures, with rating agencies adjusting to varying degrees, potentially causing retail and some institutional investors to miss the full picture. Debt-to-equity and similar ratios would appear materially higher if these obligations were consolidated or treated equivalently.

Incomplete risk transfer and contingent liabilities are also significant, as hyperscalers are frequently the sole or dominant tenant, with lease payments (often containing take-or-pay features) continuing regardless of utilization.

Residual-value guarantees or similar supports, sometimes not recorded as liabilities if management judges payment “not probable,” can crystallize precisely when asset values fall, such as due to lower utilization or rapid GPU obsolescence, meaning the economic risk does not fully leave the hyperscaler. Private-credit and non-bank transmission channels pose another risk, as SPVs link investment-grade hyperscalers to private-credit funds, insurers, and pension capital.

A wave of impairments or refinancing stress in these vehicles can transmit losses into less-transparent, less-regulated channels, with the BIS highlighting this interconnectedness as a potential amplifier of shocks. Asset-liability and depreciation mismatches are a concern, given AI hardware’s short competitive economic life (often 2-4 years) compared to longer lease and debt terms.

Aggressive depreciation assumptions on the parent side or residual-value assumptions in the SPV can leave collateral underwater if technology advances or demand softens, while simultaneous stress across multiple projects could pressure secondary-market values for GPUs and facilities.

Construction, power, and refinancing risks include delays in power delivery, construction overruns, or rising interest rates that can strain SPV cash flows before facilities generate revenue. Concentration among a limited set of private-credit providers and projects raises the chance of correlated problems. Finally, pro-cyclicality and litigation potential are risks, as guarantees and supports are inexpensive in a boom but become relevant in a downturn, exactly when counterparties are weakest. Emerging concerns include challenges to non-consolidation accounting, true-sale characterizations of asset transfers, adequacy of disclosures, and structured-finance disputes if credit enhancements fail to perform as expected.

Hyperscalers maintain strong operating cash flows and high credit ratings, with structures compliant with accounting standards for non-consolidation. Risk is shared with private capital, and assets retain residual value, enabling faster capacity growth without pressuring leverage metrics.

SPV off-balance-sheet financing rationally addresses AI infrastructure’s capital intensity but understates leverage, creates contingent exposures in stress scenarios, and links tech balance sheets to private-credit/insurance markets. Combined with circular equity loops and AI capex scale, these structures amplify potential correction severity if utilization disappoints. Key monitoring points include growth of uncommenced lease commitments, residual-value guarantees, private-credit fund performance, power/construction execution, and rating-agency or accounting shifts. While not inherently improper, their scale and opacity contribute to financial-stability risks in the AI investment cycle.