Everyone warns about the AI crash. Everyone keeps buying.

Rohit Sipahimalani, chief investment officer of Singapore's state investor Temasek, named the risk plainly this week: the primary danger to global markets, he said, is the unwinding of the AI trade. He was not forecasting an imminent reversal. The momentum in AI earnings, he allowed, is what has kept the S&P 500 near record highs even as Treasury yields climbed, masking weakness in the rest of the market ^[1].

We don't see that as imminent, but will you have bumps in 2027? Yeah, possibly yes. — Rohit Sipahimalani

Then, in the same on-record conversation, he said Temasek plans to raise its exposure to publicly traded AI assets to between 70 and 75 percent of its portfolio ^[1]. The reason he gave was agility.

One of the things we recognize is that AI is such a fast-changing environment that things could change quite easily, and you have to be able to pivot. — Rohit Sipahimalani

Sipahimalani was not the outlier that week. He was the median. On September 30, the Bank of England's Financial Policy Committee warned that valuations resting on aggressive productivity assumptions could correct severely enough to spill into sovereign debt markets; $450 billion of AI-related debt had already been issued, with debt-financed capital spending projected to reach $4.1 trillion by 2030 ^[2]. The IMF's Kristalina Georgieva located the risk in the United States, in what she called circular financing and high expectations ^[3]. The Reserve Bank of Australia warned that lenders and investors face losses if returns disappoint, through exposures built off-balance-sheet ^[2]. The Federal Reserve under Kevin Warsh has a formal review underway into the boom's effect on financial stability ^[4]. Career Treasury analysts drafted a July report comparing the market to the dot-com era ^[5]. Roughly 45 percent of professional fund managers surveyed by Bank of America name an AI collapse their top tail risk ^[6]. Now set the buying ledger beside the warning ledger. In the same late-September stretch when the Bank of England and the Reserve Bank of Australia were escalating, hedge funds bought US technology, media, and telecom shares in ten of eleven sessions, at a pace in the 97th percentile of the past five years ^[7]. The hyperscalers whose spending drives the rally watched their free cash flow turn negative and borrowed into it, even as 10-year Treasury yields reached their highest level since 2002 ^[8]. The sophisticated money did not get out. It rotated inside the trade, toward the companies with balance sheets strong enough to withstand a cooling of the AI market ^[9]. What nobody in the trade did was exit. The named risk was managed from inside the position. And the cash is real, which is the one thing keeping the warning camp honest. Nvidia reported a $96.22 billion quarter, more than double a year earlier ^[10]. Jensen Huang put the shift in a sentence.

AI has reached its inflection point. — Jensen Huang

Alphabet sits on a $515 billion cloud backlog ^[10]. Allianz's Mohamed El-Erian has called the trade a rational bubble: real potential, rationally priced, with failures arriving as isolated credit accidents that do not threaten the system ^[11]. The bears set their own tripwire, and the market crossed it without stopping. Ruchir Sharma warned in September that a decisive breach of 5 percent on the 10-year Treasury yield would begin a tight-money era that makes AI mega-projects harder to fund ^[12].

Rising public borrowing costs will squeeze other borrowers sooner, and hit the bubbly AI markets harder. — Ruchir Sharma

By early October the yield sat near 5.27 percent, its highest since 2002, and on October 6 the S&P 500, Nasdaq, and Dow still closed at record highs ^[13]. JPMorgan's Jason Hunter, who sees the 1999 pattern, concedes the tape has not turned ^[14].

In isolation, the broad indexes are still bullishly trending and have not yet given any clear warning signs that a downturn is imminent. — Jason Hunter

And the unwind the warners fear has already run once, at national scale, and been absorbed. South Korea's KOSPI surged to records in June on AI demand, then fell 35 percent, with the Korea Exchange suspending trading five times in the first half of the year ^[15]. The index entered a bear market on July 8, and Samsung fell even as it reported record quarterly operating profit of nearly 90 trillion won ^[16]. The trigger was a single supplier's capital-spending decision: SK Hynix slowing its HBM4 expansion because production forecasts for Nvidia's next chip were trending down, which knocked the KOSPI nearly 10 percent in a day, tripped a circuit breaker, and spilled onto Wall Street ^[17]. Within weeks it was re-priced: Samsung rebounded on buyback reports, investors pivoted to Micron's earnings as the demand barometer ^[17], and by October SK Hynix had raised $26.5 billion in a New York listing while the S&P 500, Nasdaq, and Dow set records ^[18][13]. The market's most recent live demonstration of the unwind ran straight through to new highs. A year ago this week, on October 8, 2025, the Bank of England and the IMF issued essentially the same warnings, the Bank comparing AI firms' earnings yields to levels not seen since the dot-com peak and Georgieva making the internet comparison herself ^[19]. Jamie Dimon of JPMorgan gave the drop six months to two years. The market rose through all of it. The cleanest measure of where things stand is the Reserve Bank of Australia's own analysis, which cuts both ways. If the trade fails, lenders and investors absorb losses through exposures that are opaque and off-balance-sheet ^[2]. If it succeeds, the buildout itself strains national capacity and feeds inflation ^[15]. Either ending breaks something. The record closes price only one of them.