The global AI trade is going through a stress test. A sharp selloff in memory chip stocks has pushed market doubts about the sustainability of the AI narrative to the forefront. From Seoul to Taipei to Philadelphia, the semiconductor sector is under broad pressure, with the Kospi index plunging more than 5% on Tuesday, while the Philadelphia Semiconductor Index has fallen more than 7% from its high this month. The core of this selloff lies not on the demand side, but in a sudden shift in the supply-side competitive landscape and a stretched capital chain. Admittedly, similar concerns have hit AI stocks before, but as investors continued to bet on the technology's transformative potential, the rally later regained momentum. New AI models released by companies such as Meta Platforms will also increase computing demand, making memory chips even more critical. But one point is clear: risks surrounding the AI trade keep piling up, raising questions about how long investors will be willing to ignore them.

Korean Stocks Become the "Fragile Link" in the AI Trade

The plunge in South Korean stocks is the most glaring signal of this round of adjustment. During Tuesday's session, SK Hynix fell more than 10% and Samsung Electronics dropped nearly 7%, with the Kospi index having retreated by one-third from its June record high. The pace of foreign outflows far exceeds that of previous years. Data from the Korea Exchange show that foreign investors have sold a net total of more than 197 trillion won in the Kospi market so far this year, compared with net selling of only about 4 trillion won for all of last year. Of that, net selling of Samsung Electronics totaled 86.06 trillion won and SK Hynix 87.14 trillion won, with the two stocks together accounting for 87% of total net selling. Foreign ownership of Samsung Electronics has fallen from 52.4% at the start of the year to 46.38%, a low not seen since January 2008. The trigger for the slump came from China's ChangXin Memory Technologies. The company announced the official mass production of its fifth-generation DRAM technology platform and plans to launch a new-generation DDR5 RDIMM that combines a 4F² architecture with hybrid bonding technology by the end of the year, entering head-on for the first time the server memory market long dominated by Samsung, SK Hynix and Micron. Counterpoint data show that ChangXin's share of global DRAM revenue has climbed from 4% in the second quarter of 2025 to 10% in the second quarter of 2026, firmly ranking fourth globally. But that is not the whole story. Samsung's Q3 results released last week were nothing short of "epic" — operating profit of 107.4 trillion won, a year-on-year surge of 782.5%, marking the first time a single quarter has broken the 100 trillion won mark. Yet the stock fell rather than rose that day. The market is no longer focused on past earnings, but on the fact that Samsung's capital expenditure in the first half of 2026 has surged compared with the same period last year, and on whether that spending could suddenly wipe out all profits in a single quarter. The roughly 55 trillion won buyback programs of Samsung and SK Hynix combined are nearing completion, and the supporting bids are disappearing.

Capital Becomes AI's New Bottleneck

Comments made by Broadcom CEO Hock Tan on Monday at DBS Bank's Asia Forward Summit in Singapore provide a key framework for understanding the current predicament. He made clear that as the industry continues to expand infrastructure spending, capital — not chip supply — will become the biggest constraint on AI development. Hock Tan reviewed the past two years: memory shortages, insufficient wafers, power tightness — these bottlenecks appeared one after another, but each was overcome. And after the chain of "wafers, memory, power, land," the biggest constraint is capital. He poured cold water on the investors rushing in, saying many of their strategies are "more like gamblers, with almost no understanding of what they are betting on." Hock Tan is no bystander. Broadcom is in talks to arrange more than $50 billion in debt financing for OpenAI, and Apollo Global Management and Blackstone Group have approached the company hoping to become lenders. Broadcom also raised $60 billion in debt financing for Anthropic, and banks have sent investors a syndication invitation for $42 billion of Class A senior secured debt. When a chip supplier starts arranging borrowing for customers to buy its own chips, the financing model of this industrial chain has fundamentally changed. Intel CEO Lip-Bu Tan offered a consistent judgment at the same forum: "If there is any challenge right now, it is capital and the cost of capital. In the long run, people want to see return on investment." The signals from the two chip giants point in the same direction: the constraint on AI expansion has shifted from a capacity bottleneck to a funding bottleneck. "The AI machine is still running at full speed, but its fuel — capital — is becoming increasingly expensive," said Hebe Chen, an analyst at Vantage Global Prime. With Wednesday's U.S. inflation data becoming the next key risk, "a tactical shift toward defensive positioning looks increasingly wise," she added. The timing of the latest selloff is also significant, coming as companies prepare to release earnings that could offer new clues on whether massive AI investments are paying off. Companies such as ASML and TSMC will report results this week.

The Debt Market Sounds the Alarm Before the Stock Market

Signs of caution are everywhere. The boom in financing for AI is unsettling investors in the more than $10 trillion U.S. corporate bond market, triggering a repricing of risk around some large technology companies. At the same time, JPMorgan and Goldman Sachs unusually limited their exposure to a loan for AI startup Volta Infrastructure Holdings Ltd. In Asia, the failed IPO of Nvidia-backed data center operator Firmus Grid Ltd. has made investors far more selective about AI infrastructure deals that depend on future profits. According to Bloomberg data, debt financing in the technology sector totaled about $500 billion in the first nine months of this year. Goldman Sachs expects that figure to climb further to $1.2 trillion by 2027. Chris Della Fave, senior vice president at Post Oak Group, estimates that AI now accounts for 25% of all corporate bond issuance, compared with just 4% two years ago. After adjusting for inflation, the AI industry's expected borrowing in 2026 will exceed the total financing raised during the cable operators' buildout of the entire internet or during the great U.S. railroad expansion of the 19th century. The credit market has already sounded the alarm first. Paramount's CDS spread has surged to its highest level since the 2008 financial crisis; SpaceX bonds have fallen to record lows, with its 2056-maturity bonds quoted at about 85 cents and spreads approaching junk levels; Oracle's entire credit curve has set record-high bond yields. The 10-year U.S. Treasury yield touched a multi-decade high of 5.36% and is currently hovering in the 5.27% to 5.30% range. Retreat in the primary market also deserves attention. When JPMorgan and Goldman Sachs arranged a $5 billion loan for AI cloud startup Volta Infrastructure, they unusually underwrote only $1.3 billion of term loans and did not underwrite the other $3.7 billion in chip loans. Volta's debt financing yield is about 11%, already one of the highest borrowing costs in the leveraged loan market for AI financing. Even Wall Street's top investment banks, best known for risk management, are participating selectively, showing that credit market pricing of AI-related debt is shifting from "faith" to "crunching the numbers."

Fragile Structure Under Concentrated Positioning

Hedge fund positioning data further confirm the market's fragility. Goldman Sachs data show that hedge funds' net exposure to the Magnificent Seven U.S. tech stocks has climbed from 14% in June to 22%, a record high and above the previous peak of 21% in June 2024. On the surface, hedge funds are cutting exposure to the tech sector; in reality, risk is not being diversified but concentrated — from holding a basket of tech stocks to making concentrated bets on a few giants. Bridgewater founder Ray Dalio's warning on this is incisive: rapid asset price gains, capital concentrating into the same investment theme, and continuously expanding debt financing — the combination of these three is a classic bubble characteristic. Speaking at the Forbes Global CEO Conference in Singapore on October 7, Dalio said: "We are now before the breaking point, but we are getting closer, and I think it is very close." He particularly emphasized the transmission mechanism of rising interest rates to the AI investment cycle — early AI spending was mainly supported by corporate cash flow, but as investment scale expands, the share of debt instruments rises, and AI investment becomes more sensitive to interest rates. Of course, there are also sharply different voices in the market. Galaxy Digital founder Mike Novogratz, also at the forum, acknowledged that AI is in the "biggest bubble of our lifetime," but advised investors to get in immediately. "Bubbles don't end like this. I know how bubbles end — they end in spectacular fashion, and this isn't spectacular enough yet."