The Zhitong Finance App learned that global AI transactions are undergoing a stress test. The sharp decline in memory chip stocks brought the market's doubts about the sustainability of AI narratives to the front. From Seoul to Taipei to Philadelphia, the semiconductor sector is under collective pressure. The Kospi Index plummeted more than 5% on Tuesday, and the Philadelphia Semiconductor Index has fallen more than 7% from this month's high. The core of this round of sell-off is not on the demand side, but on the supply side, the competitive landscape has suddenly changed and the capital chain is tight.
Admittedly, similar concerns have impacted AI stocks in the past, 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 the demand for computing power, making memory chips more critical. But one thing is clear: the risks surrounding AI trading continue to pile up, raising questions about how long investors are willing to ignore these risks.
Korean stocks have become a “vulnerable part” of AI trading
The sharp decline in the South Korean stock market is the most glaring sign of this round of adjustments. In Tuesday's intraday session, SK Hynix fell more than 10%, Samsung Electronics fell nearly 7%, and the Kospi Index has already retracted by a third from its all-time high in June.
Foreign capital is being withdrawn at a much faster rate than in previous years. According to Korea Exchange data, the cumulative net sales of foreign investors in the Kospi market have exceeded 197 trillion won since this year, while net sales for the whole of last year were only about 4 trillion won. Among them, Samsung Electronics sold 86.06 trillion won and SK Hynix sold 87.14 trillion won. Together, the two stocks accounted for 87% of total net sales. The proportion of foreign investors holding Samsung Electronics fell from 52.4% at the beginning of the year to 46.38%, a new low since January 2008.
The trigger for the sharp decline came from China's Changxin Storage. The company announced the official mass production of the fifth-generation DRAM technology platform, and plans to launch a next-generation DDR5 RDIMM incorporating 4F² architecture and hybrid bonding technology before the end of the year, entering the server memory market controlled by Samsung, SK Hynix, and Micron for the first time. According to Counterpoint data, 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, ranking fourth in the world.
But that's not all of the reasons. Samsung's Q3 results announced last week were “epic” — operating profit of 107.4 trillion won, surging 782.5% year over year, breaking the 100 trillion won mark for the first time in a single quarter.
However, the stock price did not rise on the same day but fell. The market is no longer concerned with past profits, but rather that Samsung's capital expenditure in the first half of 2026 has surged compared to the same period last year, and whether this expense will suddenly eat up all profits in a certain quarter. Samsung and SK Hynix's repurchase plans totaling about 55 trillion won are nearing completion, and back-order purchases are disappearing.
Capital becomes the new bottleneck for AI
Broadcom CEO Chen Fuyang's statement at the DBS Asia Forward-Looking Summit in Singapore on Monday provided a key framework for understanding the current predicament. He made it clear that as the industry continues to expand infrastructure spending, capital — not chip supply — will become the biggest constraint on AI development.
Chen Fuyang looked back on his experiences over the past two years: lack of memory, insufficient wafers, and tight electricity. These bottlenecks appeared one by one, but were overcome one by one. And after the “wafer, storage, electricity, land” string, the biggest constraint is capital. He threw cold water on the crowds of investors, saying that many people's strategies “are more like gamblers, and they can hardly understand who they bet on.”
Chen Fuyang was not a bystander. Broadcom is in talks to arrange debt financing of more than 50 billion US dollars for OpenAI, and Apollo Global Management and Blackstone Group have reached out to become the lender. Broadcom also raised 60 billion US dollars in debt financing for Anthropic, and the bank has issued a syndicate invitation letter of 42 billion US dollars for Class A priority secured debt to investors. When chip suppliers began arranging loans for customers to buy their own chips, the financing model of this industry chain has fundamentally changed.
Intel CEO Chen Liwu made a unanimous judgment at the same forum: “If there are any current challenges, they are capital and capital costs. In the long run, people want to see a return on investment”. The signals from the two major chip giants point in the same direction: the restrictions on AI expansion have switched from production capacity bottlenecks to capital bottlenecks.
“AI machines are still running at full speed, but their fuel — capital — is getting more expensive,” said Hebe Chen, an analyst at Vantage Global Prime. As Wednesday's US inflation data became the next key risk, “the tactical shift to defensive positions looks increasingly sensible,” she added.
The timing of the latest sell-off is also significant. It coincides with the company preparing to release financial reports, which may provide new clues about whether the huge investment in AI is paying off. Companies such as Asmack and TSMC will release results this week.
The debt market sounded the alarm before the stock market
Cautious signs are everywhere. The boom in AI financing is making investors in the US corporate bond market worth over 10 trillion US dollars uneasy, triggering a repricing of risks surrounding some large technology companies. Meanwhile, J.P. Morgan Chase and Goldman Sachs rarely limited their exposure to a loan from AI startup Volta Infrastructure Holdings Ltd.
In Asia, the IPO of Nvidia-supported data center operator Firmus Grid Ltd., has come to an end, making investors much more picky about AI infrastructure deals that rely on future profits.
According to Bloomberg data, the total amount of debt financing for the technology industry in the first nine months of this year reached about 500 billion US dollars. Goldman Sachs expects this scale to rise further to $1.2 trillion by 2027. Chris Della Fave, senior vice president of Post Oak Group, estimates that AI now accounts for 25% of total corporate bond issuance, compared to only 4% two years ago. Adjusted for inflation, the AI industry's expected loan size in 2026 will exceed the total amount of financing for cable operators during the entire Internet construction period or during the major expansion of American railways in the 19th century.
The credit market has taken the lead in sounding the alarm. Paramount's CDS spread soared to the highest level since the 2008 financial crisis; SpaceX bonds fell to a record low. The price of its 2056 bond maturing was about 85 cents, and the spread was close to junk level; the yield on Oracle's entire credit curve bond set a new historical record. The 10-year US Treasury yield once hit a decades-high of 5.36%, and is currently still hovering in the 5.27% to 5.30% range.
The decline in the primary market is also worth watching. When arranging a $5 billion loan for AI cloud startup Volta Infrastructure, J.P. Morgan Chase and Goldman Sachs rarely underwrote only $1.3 billion of these term loans without underwriting another $3.7 billion in chip loans. Volta's debt financing yield is around 11%, which is already one of the highest borrowing costs in the leveraged loan market in the AI finance sector. Even Wall Street's top investment banks, which are best at risk management, are selectively participating, indicating that the credit market's pricing of AI-related debt is shifting from “belief” to “settlement of accounts.”
Fragile structures under centralized holdings
Hedge fund position data further confirms the fragility of the market. According to Goldman Sachs data, hedge funds' net exposure to the top seven US tech companies has climbed to 22% from 14% in June, a record high, exceeding the previous peak of 21% in June 2024.
On the face of it, hedge funds are reducing their exposure to the tech industry; in reality, risk is not diversifying, but rather concentrating — from holding a basket of tech stocks to concentrating on a few giants. Dalio, founder of Qiaoshui Fund, hit the nail on the head: the rapid rise in asset prices, the concentration of capital flows to the same investment theme, and the continued expansion of debt financing. The combination of these three is a typical bubble characteristic.
Dalio said at the Forbes Global CEO Conference in Singapore on October 7: “We are currently at the point of collapse, but we are getting close; I think we are very close.” In particular, he emphasized the transmission mechanism of rising interest rates to the AI investment cycle — early AI expenses were mainly supported by corporate cash flow. As the scale of investment expanded and the share of debt instruments increased, the sensitivity of AI investment to interest rates increased.
Of course, there are also very different voices on the market. Galaxy Digital founder Mike Novogratz, also on the forum, acknowledged that AI is in the “biggest bubble of a lifetime,” but advised investors to enter the market immediately. “The bubble won't end like it is now. I know how bubbles ended — they ended in a spectacular way, and now it's not spectacular enough”.