Big Tech's $1.65 Trillion 'Invisible Debt': The Hidden AI Infrastructure Risk

How Does the Accounting Allow This?

The treatment is permitted under standard accounting rules (GAAP/IFRS) for operating leases and special purpose entities (SPEs) or joint ventures. Common structures include:

• A dedicated vehicle (JV or SPE) that acquires or develops data center assets, keeping the debt off the hyperscaler's balance sheet .
• "Take-or-pay" contracts and long-term lease commitments disclosed only in financial footnotes rather than recognized as balance sheet liabilities .
• The Bank for International Settlements (BIS) has labeled this practice "shadow borrowing" .

Specific Figures: Meta and Oracle

Meta secured about $60 billion in capital for data center construction, and approximately half of that — or $30 billion — was structured to sit completely off its balance sheet . Meta shows roughly $28.8 billion in on-balance-sheet debt .

Oracle is the most leveraged of the five: its current on-balance-sheet debt stands at approximately $105 billion, and it has disclosed off-balance-sheet lease commitments of roughly $20 billion . Oracle's free cash flow has turned negative .

Alignment with Moody's Report

Moody's Ratings separately found that the five major hyperscalers have signed but not yet recognized on-balance-sheet data center leases totaling approximately $662 billion, equivalent to 113% of their combined adjusted debt . The actual volume of this "phantom debt" could be significantly higher . This aligns with Nikkei's broader $1.65 trillion figure, which includes not only leases but also purchase commitments for chips and other infrastructure.

Credit Market Strain

• Goldman Sachs has observed that the investment-grade bond market may not be able to accommodate all of big tech's financing needs . Barclays separately noted that the surge in AI-related debt is "testing" the high-grade market .
• AI-related debt now accounts for roughly 30% of net new investment-grade supply in the U.S. dollar market .
• The Economist reports the five firms will spend $800 billion in 2026 alone on AI infrastructure .
• The largest hyperscalers have pledged to spend as much as $725 billion in capital expenditures in 2026, with that figure expected to cross $1 trillion by 2027 . McKinsey projects global investment in data centers could reach $7 trillion by 2030 .

Risks If AI Demand Underperforms

Industry observers and analysts have flagged multiple risks:

• Write-downs and stranded assets: If AI demand does not materialize, the massive data center buildout could leave companies with underutilized or obsolete assets .
• Cash flow strain: Off-balance-sheet structures commit companies to long-term lease and purchase payments regardless of demand, which could squeeze share buybacks and dividends .
• Hidden leverage: Much of the $1.65 trillion in off-balance-sheet obligations is legally binding but not reflected in standard leverage ratios, making risk assessment difficult for investors .
• Credit market disruption: A default or restructuring by a major player could spill over into the private credit and insurance markets that have financed much of this shadow borrowing .

Companies' Bullish Counterargument

The hyperscalers argue that the AI infrastructure buildout is a necessary long-term investment. Their key defenses include:

• Massive project backlogs: Companies note that cloud and AI service demand is growing rapidly, with long-term enterprise contracts providing revenue visibility .
• AWS's stance: Amazon Web Services has maintained that its capacity planning is driven by existing and forecast customer demand, and that it can slow down capital spending if needed .
• Fortress balance sheets: Despite increased debt, companies like Microsoft, Alphabet, and Amazon still maintain investment-grade credit ratings and strong cash positions that can absorb shocks .
• Revenue growth: Companies emphasize that AI-related revenue is growing faster than infrastructure spending, and that capacity constraints — not demand — are the current bottleneck .