On August 13, AMD raised $4.75 billion in its largest dollar bond sale ever, adding another log to the AI-fueled debt fire. AMD is not alone. NVIDIA sold $25 billion in bonds in June and drew roughly $85 billion in orders. Alphabet issued $25 billion in early August and saw $115 billion of demand. Amazon has signaled plans to raise at least $25 billion. According to LSEG data cited by Reuters, Alphabet, Amazon, Meta and peers have sold nearly $220 billion in bonds so far this year, more than double the $108 billion they issued in all of 2025.
The official rationale is straightforward: training clusters keep growing, AI agent deployments are lifting inference demand, and high-end GPUs remain scarce. In a boom, borrowing to build is normal. What is less normal is what sits underneath, and what never shows up on balance sheets.
The $2 Trillion Layer That Never Touched a Balance Sheet
Goldman Sachs estimates that major cloud operators hold roughly $1.5 trillion in off-balance-sheet lease commitments, about $1 trillion of which has not even started yet. Under US GAAP these are disclosed only in footnotes, not recognized as lease liabilities or right-of-use assets. Morgan Stanley separately counts $982 billion in procurement commitments for chips, equipment, power and compute across Alphabet, Microsoft, Amazon, NVIDIA and Oracle as of the first quarter. Together, the two estimates approach $2 trillion.
Meta Built the Template
The structure traces back to Meta's Hyperion data center in Louisiana. Meta and asset manager Blue Owl set up a joint venture called Beignet, which develops and owns the facility. Meta holds only a 20% stake but guarantees a lease of at least 20 years. That guarantee let Beignet issue a record $27 billion in amortizing bonds. Meta effectively carries the repayment burden, yet the debt does not appear on Meta's books, letting it scale AI infrastructure without inflating reported leverage. Rivals noticed, and similar structures are spreading.
Credit Markets Are Starting to Flinch
By traditional metrics the cloud giants still look healthy: Morgan Stanley puts their average net leverage at 0.5x, versus 0.8x for tech and 1.8x for non-financial corporates, and they hold more cash than debt. Optimists argue equity holders absorb losses first if data center returns disappoint.
But the repricing has begun. US 30-year real yields sit near 18-year highs at roughly 3%, with UK and German 10-year real yields near decade peaks, and analysts point to hyperscaler borrowing as a key driver. Beignet's bonds now yield around 6.95%, up from 5.65% at issue last autumn. Credit default swap spreads for Oracle, NVIDIA, Meta and even Alphabet have hit record highs.
What This Signal Actually Means
Every technology boom has a financing signature. This one just shifted: first equity, then debt, and now off-balance-sheet vehicles that hide leverage in footnotes. That progression is a textbook late-cycle pattern, the same one that marked railroad overbuilding and telecom fiber manias. The bullish case is real, agent-driven inference demand is growing, but the financing case is increasingly fragile: if returns on AI data centers disappoint, the hidden commitments would hit credit markets before anyone reads about it in a headline.
For context on the capital climate, our analysis of the AI agent funding winter shows the same tightening from the venture side.
What to Watch Next
- Footnote disclosures: check quarterly filings for lease and procurement commitments that never hit the balance sheet.
- Real yields and CDS spreads: the market's honest verdict on AI debt, updated daily.
- New Beignet-style structures: if off-balance-sheet vehicles keep multiplying, the hidden layer grows faster than reported debt.
- API price dynamics: today's cheap model prices may be partially subsidized by cheap debt, a durability question every developer should ask.
The AI buildout is no longer just a technology story. It is a credit story, and the credit market has started to notice.
