On Tuesday, October 13, the global semiconductor trade hit turbulence. A single weekend announcement from a Chinese memory maker, ChangXin Memory Technologies, collided with a wave of doubt about runaway AI spending, and chip stocks around the world sold off in unison. South Korea's KOSPI benchmark fell 5.5 percent to close at 6,035.34, its worst session in months. Samsung Electronics dropped 7.4 percent. SK Hynix cratered 10.1 percent. The selling reached Wall Street overnight and kept going into the next morning's premarket: Micron down 4 percent, SanDisk down 5.2 percent, Western Digital and Seagate off 3.5 to 4.5 percent, and the Philadelphia Semiconductor Index down 2.7 percent.
The proximate trigger was a claim. CXMT's president, Cao Kanyu, told an industry conference in Shanghai over the weekend that the company will ship a new generation of data-center server memory chips, built on a next-generation cell architecture called 4F², before the end of the year. If true, CXMT would beat Samsung and SK Hynix to a threshold the two Korean giants have chased for years. It would have done so while cut off, by US sanctions, from the top-end lithography tools made by ASML that such advances were thought to require.
Markets took the claim seriously first and will verify it later. But the selloff was not really about one chip design. It was about two fears arriving at the same moment: that China's memory makers are closing the technology gap faster than expected, and that the money pouring into AI infrastructure may not come back.
What CXMT actually claimed
The technical claim is narrow and, on its face, significant. The 4F² design packs the same number of memory cells into a smaller area by stacking transistors with what CXMT describes as a vertical surrounding gate, a structure that reduces current leakage between cells. TrendForce estimates the approach could raise memory density by 30 to 50 percent. CXMT's stated product is a DDR5 RDIMM module for data-center servers, due before the end of 2026.
Server DRAM is the high-value, high-margin end of the memory market, the tier where Samsung and SK Hynix earn much of their profit. Bloomberg characterized CXMT's announcement as an unsubstantiated claim, and no independent verification has emerged. But CXMT's trajectory the market could not dismiss: the company's share of global DRAM revenue has climbed from roughly 4 percent to about 10 percent in the span of a year, through the second quarter of 2026.
The selloff, in numbers
Enjoying this story?
Get the five most important stories in tech, every morning. Free.
The Korean session on October 13 was historic by the standards of recent years. The KOSPI fell as much as 7.4 percent intraday before recovering part of the loss, closing down 5.47 percent at 6,035.34. The KOSPI 200 futures index fell more than 5 percent within a minute about an hour after the open, triggering the market's sell-side circuit breaker for the first time since August 20. The Korea Exchange temporarily restricted program selling as the decline deepened.
Samsung closed down 7.38 percent at 232,000 won. SK Hynix closed down 10.09 percent at 1.443 million won. Together the two chipmakers account for more than half of the KOSPI by market capitalization, so a bad day for both is a bad day for the whole index: the close extended the KOSPI's losing streak to five sessions, with cumulative losses approaching 14 percent.
The selling was global. Taiwan's TAIEX closed 2 percent lower at 48,341, with TSMC down 2 percent and foreign investors selling a net TWD 165.34 billion of Taiwan-listed shares. On Wall Street, the Invesco QQQ fund tracking the Nasdaq 100 fell 1.6 percent to 739.27.
Two fears arrived at once
The October 13 Chip Selloff
One-day price moves, October 13, 2026. Bars scaled to the largest drop.
The first fear is the memory moat. For a decade, the investment case for Samsung, SK Hynix, and Micron rested on an assumption: memory is a brutal commodity business, but the technical frontier moves only at the pace the three giants can afford. CXMT's 4F² claim attacks the premise. Even if the product ships late or at low yields, the announcement compresses the perceived timeline of Chinese competition.
The second fear is the AI capex question, and it gave Tuesday's selloff its breadth. On Monday, Broadcom chief executive Hock Tan told an audience that some of the money flowing into AI resembles casino gambling, according to Bloomberg's account. The same day, the International Monetary Fund warned that disappointing AI productivity could trigger a sharp fall in AI-linked asset prices. The New York Times framed the day's selling as the convergence of two anxieties: fears of runaway AI spending meeting the threat from Chinese rivals to the memory leaders.
That convergence is what made the damage so broad. Memory is the bottleneck commodity of the AI buildout: every accelerator rack ships with racks of DRAM beside it. A market that doubts the return on AI infrastructure spending sells the memory makers first, because their margins are the most cyclically exposed.
Memory was supposed to be the corner of the chip industry where the incumbents' lead was safe. On Tuesday, the market decided it no longer believed that.
The gap between a claim and a chip
A dose of skepticism is warranted. Bloomberg called CXMT's announcement an unsubstantiated claim, and the history of memory technology is full of announcements that arrived years ahead of viable production. Jing Jie Yu, an analyst at Morningstar, noted that a successful and earlier implementation of the 4F² architecture would vastly improve CXMT's competitive position versus its peers, but added that a meaningful time gap should still be expected between the unveiling of the product and commercialization at viable yields.
That yield question is the entire ballgame in DRAM. Designing a denser cell is one discipline; producing billions of them with defect rates low enough to make money is another, and it is where the incumbents' decades of process knowledge live. SK Hynix has published a development roadmap for 4F²-class technology, so the architecture is not unknown to the leaders, only unshipped.
But the market's reaction carries its own information. CXMT's share gains are real: 4 to 10 percent of DRAM revenue in a year is not an announcement, it is output. Sanctions have not stopped the climb; they have arguably concentrated it, forcing Chinese fabs to innovate around equipment constraints rather than through them.
What happens next
The first checkpoint is Thursday, October 15, when TSMC holds its earnings conference in Taipei. TSMC's read on AI demand will set the tone for the entire sector and test whether Tuesday's selling was a localized memory panic or the start of a broader repricing of AI infrastructure. After that, the calendar turns to fundamentals: whether CXMT actually ships its DDR5 RDIMM product before year-end, what yields it achieves, and whether pricing in server DRAM begins to reflect a new supplier's entry.
For the industry's structure, the longer question is whether 4F² is a one-time leap or the first of several. If CXMT can ship competitive server DRAM on schedule, the next contested ground is high-bandwidth memory, the stacked chips that feed AI accelerators and carry the industry's fattest margins. The market has now learned to stop assuming the incumbents' lead is safe, and started pricing the gap between assumption and evidence.
0 Comments