On Tuesday 13 October 2026, South Korea's benchmark KOSPI index closed at 6,035.35, down 5.46% on the day after falling more than 7% at one point during the session. It was the fifth consecutive declining day. Depending on which measurement window you use, the five-session loss was somewhere between roughly 13.8% and more than 15%, leaving the index around a third below the record high it set in June.

The damage was concentrated in two names. Samsung Electronics fell 7.39% to close at 232,000 won. SK Hynix fell 10.09% to 1.443 million won. Both had already dropped more than 4% the previous session. One estimate put the total Korean market value erased during the slide at about 229 trillion won. Foreign investors sold around 4.5 trillion won of KOSPI shares on 13 October alone.

The proximate cause was not a central bank, a war, or an earnings miss. It was a presentation. At an industry conference in Shanghai a few days earlier, Cao Kanyu, chief executive of ChangXin Memory Technologies, China's largest maker of DRAM memory chips, said the company would launch a new generation of server memory by the end of the year using a denser cell architecture known as 4F2. The claim alone was enough.

In this article we explore what CXMT actually announced, why a memory chip design can move an entire national stock index, how the Korea Exchange's circuit breakers worked during the fall, why Samsung's record profit guidance failed to help, and what the episode reveals about how much of the global AI rally rests on assumptions rather than facts.

What DRAM Is, and Why It Matters So Much

DRAM stands for dynamic random access memory. It is the chip that holds data a computer is actively working on, as distinct from storage, which holds data the computer is not currently using. Every server, phone and laptop needs it.

For most of its history DRAM behaved like a commodity. The chips were close to interchangeable, so prices swung violently with supply and demand, and manufacturers earned thin margins in bad years and fat ones in good years. What changed with artificial intelligence is that training and running large models requires enormous quantities of memory sitting very close to the processor, feeding it data fast enough to keep it busy. That created a premium product called high bandwidth memory, or HBM, which stacks DRAM chips vertically to move far more data per second. HBM is difficult to manufacture, commands much higher prices, and is dominated by a very small number of firms.

Three companies control almost the entire global DRAM market. According to Counterpoint Research, in the first quarter of 2026 Samsung held about 38% of global DRAM revenue, SK Hynix about 29% and the American firm Micron about 22%. That is roughly 89% between them. CXMT held about 8%.

What CXMT Actually Claimed

The technical heart of the announcement is the phrase 4F2. In chip design, F refers to the smallest feature size the manufacturing process can produce, and the number in front describes how many units of area each memory cell occupies. Most DRAM in production uses a 6F2 layout. Moving to 4F2 means each cell takes up roughly a third less area, so more memory fits on the same piece of silicon, which lowers cost per gigabyte.

The reason the industry has not simply done this already is physics. Pack cells closer together and electrical current leaks between them, corrupting the stored data. CXMT said its approach uses a Vertical Surrounding Gate transistor structure, meaning the control gate wraps around the vertical channel rather than sitting beside it, to contain that leakage. TrendForce estimated that a successfully implemented denser design of this kind could lift memory density by 30% to 50%.

Two things are worth holding onto. First, 4F2 is a threshold that Samsung and SK Hynix had not yet achieved in production. Second, every technical claim here traces back to remarks made at a conference. No detailed specification or independently verified benchmark accompanied them.

Why Two Stocks Can Sink a National Index

The KOSPI is capitalisation weighted, meaning each company's influence on the index is proportional to its total market value rather than its share price or its headcount. Samsung Electronics and SK Hynix together account for roughly half of the KOSPI's entire market value.

That is an extraordinary concentration. It means the Korean benchmark is, in practice, a leveraged bet on global memory demand wearing the costume of a diversified national index.

The mechanism of June's record also explains October's collapse. Samsung and SK Hynix pushed the KOSPI past 9,000 in June 2026 on AI-driven memory demand. The same two stocks dragged it to more than 30% below that peak four months later.

The Sidecar, and How Fast Falls Get Faster

At 9:48 a.m. on 13 October, with the index down more than 4%, the Korea Exchange activated what it calls a sell-side sidecar. This suspended program trading in KOSPI-listed shares for five minutes. At the moment it triggered, the KOSPI 200 futures index stood at 955.02, down 5.08% from the previous close. It was the first such curb since 19 August.

A sidecar is worth understanding because it reveals something about modern market structure. Program trading means large baskets of shares bought or sold automatically as a single instruction, often by computers responding to moves in the futures market. When futures fall sharply, automated systems sell the underlying shares to keep their hedges balanced, which pushes the shares down, which pushes futures down further. The sidecar interrupts that loop for a few minutes to let human judgment catch up. It is a speed bump, not a floor. The index closed down 5.46% regardless.

Record Profits Did Not Help

Five days before the worst of the sell-off, on 8 October, Samsung released preliminary third-quarter guidance. It projected operating profit of about 107.4 trillion won, roughly 80.2 billion dollars, up around ninefold from a year earlier. Revenue was guided at about 195 trillion won, up 127% year on year. On that preliminary guidance, it would be the first time any technology company had posted quarterly operating profit above 100 trillion won, with full audited results not due until late October.

Operating profit is what a business earns from its core activities before interest and tax. A ninefold increase in a single year is the signature of a commodity cycle at its most violent upswing, where prices rise far faster than costs.

The shares fell anyway. This is the single most instructive detail of the episode. Markets do not price what a company earned last quarter. They price the stream of earnings investors expect it to produce in future, discounted back to today. Samsung's record quarter told investors how good the present was. CXMT's announcement told them something about the future, specifically that the duopoly pricing power generating those earnings might have a shorter life than assumed. The second piece of information mattered more, and it was not even a fact.

The Case for the Sell-Off, and Against It

The sceptical view

CXMT has not published detailed technical specifications for the 4F2 or Vertical Surrounding Gate design. A Korea Times report said the company still struggles to stabilise yields on high bandwidth memory, meaning the proportion of usable chips per wafer remains unreliable, while Samsung and SK Hynix are already sampling the newer HBM4E generation with customers. The announced product targets server DDR5 memory, which is lower margin conventional DRAM, not the HBM that AI accelerators actually require. And CXMT's share of global DRAM revenue was around 8% in the first quarter of 2026, against about 89% for the three incumbents combined.

The worried view

CXMT's DRAM market share reportedly rose from about 4% a year earlier to roughly 10% by the second quarter of 2026. A firm moving that fast, with state backing, does not need to beat Samsung on technology to hurt it. In a commodity market, additional supply at the low end compresses prices across the board, and the extraordinary profits Samsung just reported came substantially from price. LS Securities analyst Kim Yun-jeong attributed the weakness in AI and technology stocks directly to CXMT's DRAM advancements.

The selling was not confined to Seoul. Micron shares fell more than 4% and SanDisk about 6% in overnight US trading as the news spread.

The Condition of the Market Mattered as Much as the News

The CXMT announcement landed on a market already unsteady. The same week brought a broader wobble in global technology and AI shares, including public comments from Broadcom chief executive Hock Tan comparing some AI investors to casino gamblers. Institutional buying support was thinning as chipmaker share buyback programmes neared completion, removing a reliable source of demand for the shares.

That combination matters. The same headline delivered into a market with plenty of buyers, modest valuations and no narrative fatigue would have produced a bad day. Delivered into a market where positioning was crowded, support was fading and confidence was already fraying, it produced a 229 trillion won hole.

Key Takeaways for Investors

  • Index diversification can be an illusion. Half the KOSPI's market value sits in two memory chipmakers, so a Korea allocation is largely a memory semiconductor allocation.
  • Record earnings are backward looking. Samsung guided to what would be the first quarterly operating profit above 100 trillion won by a technology company and fell 7.39% days later, because share prices reflect expected future earnings, not reported past ones.
  • Duopoly pricing power is the asset being valued, not the chips. Anything that credibly threatens the ability to set prices threatens the valuation, even if volumes are unaffected for years.
  • Unverified claims can move markets as effectively as verified ones.
  • Market condition amplifies news. Crowded positioning and fading buyback support turned a technical conference remark into a historic sell-off.
  • Circuit breakers slow disorderly trading, they do not change direction. The sidecar paused program trading for five minutes and the index still closed down 5.46%.

Conclusion

The uncomfortable thing about this episode is not that investors overreacted. It is that nobody can currently prove they did. The entire case rests on remarks made at a conference by a company that has published no verifiable data, and the entire counter-case rests on yield difficulties reported secondhand and a market share number from the first quarter of 2026. Both sides are arguing from fragments.

That is the real lesson in the KOSPI's five days. The AI trade has been priced on an assumption of durable scarcity, that a very small number of firms would control the critical inputs for a very long time, and that the extraordinary margins this produces would persist. Scarcity assumptions are unusually fragile, because they can be undermined by a single credible claim long before any product ships. When the market finally gets hard numbers on what CXMT can actually manufacture, those numbers will matter far less than the discovery of how little it took to put the question on the table in the first place.