The AI Scare Has Moved From Stocks to Debt

On October 15, the United States reported strong retail sales and its biggest rise in producer prices in four years — the kind of data that normally lifts the dollar. The dollar fell anyway, 0.3% in its sharpest single-day drop in over a month, down 1% against the Swiss franc, in its first weekly decline in five. Traders attributed the move not to the data but to a sell-off in AI-related debt, on doubt about whether the companies that own the data centers can keep servicing what they owe, and money fled into Treasuries, handing the 10-year yield its largest weekly fall in over a year ^[1]. The Malaysian ringgit strengthened on the same AI-selloff headlines ^[2], and Korea's KOSPI had already answered them two days earlier with a 5.46% fall, in the same wave that pulled the 10-year yield down from its 5.365% peak ^[3]. That is the opposite of how the market handled the last AI scare. In December 2025, when Broadcom fell more than 11% despite strong profits and Oracle slid after a $15 billion increase in capital spending, the selloff was an equity event. Bond yields rose, the 10-year climbing to 4.18% on inflation fears ^[4]. The market then worried AI was overheating the economy; the market now worries the companies building it cannot repay their loans. The direction bonds move on an AI headline is the simplest way to see where the risk lives, and the direction has reversed. The debt was always there. Even while the December selloff was equity-led, the scaffolding was taking shape: Meta, Google, and Amazon were routing billions in data-center debt through special-purpose vehicles, shell companies set up to hold the borrowing off the parent's balance sheet, and lenders were covering up to 150% of construction costs. CoreWeave carried $14 billion in debt against $34 billion in coming lease payments, its AI chips posted as loan collateral, and the cost of insuring that debt — credit default swaps, the market's price on default — rose nearly 60% in two months. The Bank of England was already reviewing the lending ^[5]. By this fall the pile had grown too large to stay in footnotes. The giant cloud companies and their main suppliers owed more than $4.4 trillion across on- and off-balance-sheet obligations, part of a buildout Chase Bank estimates at $5.5 trillion by 2030, a sum that exceeds the cash these companies actually generate ^[6]. Goldman Sachs counted $1 trillion in uncommenced data-center lease commitments — off-balance-sheet obligations visible only in financial footnotes — up from $725 billion in under a month ^[7]. And the price of borrowing was turning: Meta had to offer yields above 7% on its July financing for a Texas data center, roughly 0.4 percentage points more than nine months earlier, on a project that carried mixed credit ratings, S&P at A+, Fitch and KBRA at AA- ^[8]. The risk did not sit where the public assumes it sits. Andromeda Capital described the AI credit cycle as a "heads I win, tails you lose" arrangement: credit investors accepting record-tight spreads and capped upside while the real exposure shifted onto life insurers, pensions, and endowments ^[9][10]. The officials followed the market's cue. On September 25, Kansas City Fed president Jeffrey Schmid asked the question once reserved for the banking system.

Where we have to start to really synthesize what's happening in the AI and the data center build-out is are we moving to a too-big-to-fail AI ecosystem. — Jeffrey Schmid

Five days later, the Bank of England and the Reserve Bank of Australia warned that AI debt issuance had already reached $450 billion and could grow to $4.1 trillion in debt-financed spending by 2030 — more than some national governments borrow — and that a correction could spill into sovereign debt markets because government finances increasingly lean on AI-driven growth ^[11]. Earlier in September, the Federal Reserve had named the buildout in its own reasoning: its rate hike on the 16th, its first since 2023, explained that the AI infrastructure buildout was a primary driver of core-goods inflation, alongside the war in Iran ^[12]. And two days before the dollar's drop, a private-credit chief executive put the bargain bluntly: in this buildout the equity holders collect the upside while the credit holders absorb the downside. He estimated defaults were creeping toward 5%, higher than the industry advertises, called the loan amendments "can kicking," and flagged that there is almost no deal flow outside data centers ^[13]. The hedge funds that sit between the debt and the shares have spent the year whipsawing rather than exiting. They sold in the second quarter ^[14]. In July, after a momentum crash, prime brokers — the banks that finance these funds — force-liquidated a fund leveraged roughly four times, handing its entire $45 billion equity book to Citadel in a single block trade ^[15]. By mid-September they were back in force, buying US tech at a pace in the 97th percentile of the past five years ^[16], and on October 7 Goldman Sachs' own volatility trader flipped from recommending AI upside to taking the opposite side, on the grounds the trade had grown too crowded ^[17]. One IMF number explains why any of this matters beyond the funds themselves: hedge funds now hold 9% of US Treasury debt, more than double their 2022 share, so any collective unwind would transmit straight through the bond market ^[3]. None of this is to say the demand was fake. Nvidia reported quarterly revenue of $96.22 billion, more than double a year earlier ^[18], and Anthropic reached $65 billion in annualized revenue and turned profitable ^[19]. The buildout is real. What has changed is where the argument must now be won. The strongest published defense of the buildout is no longer that the debt is small or that demand alone settles it. JPMorgan Asset Management strategist Stephanie Aliaga put it in credit terms.

We think that the market is very capable of absorbing this new issuance. And if anything, it might just enable this AI boom to be sustainable. — Stephanie Aliaga

Her arithmetic: the six largest cloud companies carry less leverage than the broad investment-grade index and could add roughly $1.5 trillion in debt before reaching market averages, though TD Asset Management cautions that the capacity to absorb may not be infinite ^[6]. That is the tell. The bull case now rests on the same ground the skeptics occupy — the bond market's capacity to carry the debt, not the size of the debt itself — with the opposite conclusion. Both sides are making their case on the liability side of the buildout. The debate has not been settled. It has moved.