← Back to feed

Five of the largest US technology companies have accumulated $1.65 trillion in off-balance-sheet AI infrastructure obligations, according to a Nikkei Asia analysis of SEC footnotes and supplemental disclosures. The figure has grown eightfold over roughly five years as the companies locked in long-term compute, power, and facility capacity that does not appear in their reported debt totals.

Meta carries the largest hidden load at $420 billion — three times its formal reported debt. Oracle, Amazon, Google, and Microsoft make up the remainder. None of this is technically concealed: SEC rules require disclosure of operating leases, purchase commitments, and take-or-pay contracts in financial statement footnotes. But the obligations are buried well below the headline figures analysts and investors typically track.

What the Obligations Cover

The $1.65 trillion figure spans three categories:

Data centre leases: Long-term commitments to occupy facilities owned by third parties — Equinix, Digital Realty, NTT, and the new AI-specialist landlords like Aligned, Stack, and Vantage. Under current accounting rules, these appear as right-of-use assets and lease liabilities, but only the current-year portion surfaces in standard debt covenants.

GPU and custom silicon contracts: Multi-year supply agreements with NVIDIA, TSMC, and Broadcom for Blackwell, Grace Hopper successor chips, and custom ASICs. Meta’s MTIA deal, Google’s Broadcom TPU agreement through 2031, and Microsoft’s NVIDIA commitments are all structured as purchase commitments rather than debt.

Power agreements: 20-year PPAs with utilities and independent power producers, some at fixed rates that may exceed market pricing if energy costs fall. These are disclosed as contractual obligations but not included in net debt calculations.

The Structural Risk

The obligations are sustainable as long as AI revenue continues to scale. The problem identified by multiple analysts in recent weeks — including BIS, Goldman Sachs, and now the Nikkei data — is what happens if the revenue curve flattens before the obligation schedule does.

A typical GPU supply contract or data centre lease has a 7-12 year term with limited exit clauses. A company that signed a commitment in 2024 projecting 40% annual revenue growth from AI services has limited ability to renegotiate if that growth rate declines to 15-20%. The capital is locked.

The $1.65 trillion figure also excludes off-balance-sheet obligations held by non-consolidated entities. Several AI compute ventures — Apollo and Blackstone’s CoreWeave partnership, Anthropic’s TeraWulf lease, and similar structures — sit inside vehicles that are not fully consolidated on tech giant balance sheets. The true contingent exposure is larger.

Why It Has Not Moved Markets

Tech stocks have largely absorbed the concerns because Q1 and Q2 2026 earnings were strong across the board. Google Cloud hit $20 billion in a quarter. Azure AI exceeded $37 billion in annual run rate. The revenue is appearing, and analysts are willing to give the capital commitments the benefit of the doubt.

The risk scenario is not collapse — it is a margin squeeze if AI revenue growth moderates while infrastructure cost obligations remain fixed. At $1.65 trillion spread across five companies, the margin impact of even modest utilisation shortfalls would be visible in earnings. The obligations are only invisible until the quarter they are not.