The bond market priced AI's risk before the stock market did
The $8 trillion selloff wasn't the moment the buildout's risk arrived — it was the stock market finally catching up to a repricing the bond market had been doing all year.
TSMC's bonds now yield less than Meta's. SK Hynix paper due in 2031 trades nine basis points above Amazon's — nine hundredths of a percentage point — and the premium investors once demanded to hold emerging-market corporate debt has shrunk from a historical 30 to 50 basis points to a handful. [1] That price printed on September 10, two months before the $8 trillion stock-market wipeout everyone treated as the event. [2] The bond market had already delivered its verdict; the stock market was still arguing about valuation. The migration began in July. When the retail trade broke — the leveraged funds and speculative crowds it had drawn in — the leverage did not disappear. Analysts flagged it moving instead onto the balance sheets of companies like Alphabet and Nvidia. [3] Those balance sheets were already stretched. By the second quarter, the hyperscalers' capital spending had crossed above their operating cash flow — the cash the businesses generate from running themselves — leaving free cash flow negative. [4] In August, Goldman Sachs counted a trillion dollars of data center lease commitments the hyperscalers had signed but not begun, obligations living in footnotes rather than on the balance sheet, up from $725 billion a month earlier. [5] Their borrowing had reached 30% of all net investment-grade debt issued this year — investment-grade meaning borrowing by companies whose credit is rated safe enough for conservative buyers — on its way toward a trillion dollars. [6] The instruments were built to stay hard to see: 144A bonds, which skip the usual public registration; special purpose vehicles; joint ventures like Meta's $50 billion Hyperion; unrated loans. Nearly $500 billion of AI-related debt was raised through such structures this year. [7] In September, Apollo's chief economist Torsten Slok named what the bond market had begun to price. [8]
What the market is repricing is hyperscaler credit fundamentals, namely a debt-financed AI capex cycle with rising leverage, negative free cash flow and uncertain payback on depreciating assets. — Torsten Slok
The crash was still two months away. By October the credit market was marking individual names. Credit-default swaps — insurance against a borrower failing to pay — widened across Nvidia, Broadcom, Oracle, AMD and Meta. Broadcom's swaps tripled while its shares had fallen only 25%; Oracle's had dropped more than 40%. The bond market was treating these as riskier credit while equity investors still debated whether the stocks were cheap. [9] In November the equity market marked down the same risk — $8 trillion erased in 45 days. [2] The same week, the European Central Bank gave the warning institutional weight: US tech giants lean on more than $4 trillion in opaque off-balance-sheet commitments, an echo of the dotcom and subprime eras, with circular financing in which chipmakers lend to their own customers. [10] Michael Burry put a number on the special-purpose share of that pile: up to $3 trillion of AI debt hidden in special purpose vehicles. [2] By then the risk had left the tickers entirely. Jefferies downgraded whole countries — South Korea to underweight, Taiwan to neutral — citing higher financing costs and weaker data center spending, in a week when AI names made up nearly half the MSCI Emerging Markets Index. [11] The ECB named France the sovereign weak link, debt at 119% of GDP. [10] Even the machines make doubtful collateral: an Indian state lender planning to finance GPU clusters flagged that the chips lose their usefulness in 18 to 24 months — faster than the loans against them pay down. [12] And Aon's chief executive warned that clustering billions of dollars of equipment in a few buildings creates a concentration insurers won't underwrite — which, he said, closes the door to the debt market itself. [13] None of this has shut the window. Goldman Sachs, publishing during the crash week, still forecasts the hyperscalers will finance 35% of next year's capital spending with roughly $440 billion of new debt, after this year's borrowing beat estimates. [14] Strategists at Morgan Stanley, Goldman and Federated Hermes kept 8,700 as their targets for the S&P 500, reading the selloff as froth removal rather than a credit event. [15] The market's own logic is blunt.
the lesson in capital markets, and we've seen this time and time again, is corporates issue when they can, not when they have to. — Goldman Sachs Private Wealth Management
The test has a date. $13.4 billion of data center developers' fixed-rate dollar debt matures by 2027, and refinancing it at October's Treasury yields adds 2.85 percentage points — at least $380 million in added interest before credit spreads widen at all. [16] Either the bond market was early, and those maturities hit a buildout already spending past its own cash flow; or the froth-removal reading holds and the borrowing rolls over on the old terms. 2027 is when it comes due.
- 1. US Tech AI Spending Shifts Global Bond Risk Hierarchy
- 2. AI Stock Selloff Wipes $8 Trillion From Markets
- 3. AI Trade Leverage Shifts to Major Tech Balance Sheets
- 4. AI Spending Surge Shifts Investors Toward Free Cash Flow
- 5. Hyperscalers Accumulate $1 Trillion in Data Center Lease Commitments
- 6. Hyperscaler Debt Issuance Rises Toward 1 Trillion Dollar Mark
- 7. Wall Street Develops Novel Financing for AI Infrastructure Boom
- 8. Apollo Global Management Warns of Rising Hyperscaler Debt Risk
- 9. AI Giants Face Widening Credit-Default Swap Spreads
- 10. European Central Bank Warns of AI Bubble and French Debt
- 11. Jefferies Downgrades South Korea and Taiwan Over AI Risks
- 12. Union Bank of India Forecasts Shift to AI Infrastructure Debt
- 13. Aon CEO Warns AI Infrastructure Risks Threaten Data Center Financing
- 14. Goldman Sachs Forecasts $440 Billion AI Debt for 2027
- 15. Wall Street Divided as S&P 500 Enters Correction
- 16. AI Data Center Developers Face $380 Million Interest Surge