How Memory Supercycles End: What CXMT's 4F² Claim Really Tells Us
How Memory Supercycles End: What CXMT's 4F² Claim Really Tells Us
One press report did not end the memory supercycle, but it did reprice how much of it investors were willing to pay for. On October 13, KOSPI fell 5.46% to 6,035.35, SK hynix dropped roughly 10%, and U.S. chip shares extended a multi-day slide after reports that China's CXMT had made key progress on a denser DRAM design. A Broadcom CEO remark comparing AI investing to a casino, and a reset in OpenAI's reported run-rate revenue, landed on the same days.
This piece separates what has been reported from what has been proven, shows how earlier memory cycles actually ended, and lays out the most probable path from here, along with the risks that would break it.
Source: Korea Exchange
Source: FT, Reuters, Axios (Oct 8)
Source: market reports
What happened in three days: a rout driven by narrative, not earnings
The sell-off began with expectations, not with a weak quarter. Memory names were already priced for a long AI-driven upcycle, so any credible challenge to pricing power hit hard. The table below lays out the sequence.
| Date | Event | Market reaction |
|---|---|---|
| Oct 8 | FT and Reuters report OpenAI run-rate near $50B, below earlier figures | AI-linked equities soften; Korea closed for a holiday on Oct 9 |
| Oct 12 | Broadcom CEO likens AI investing to a casino; Korea reopens | KOSPI −3.65% to 6,384.24 |
| Oct 13 | Report of CXMT "key progress" on 4F² DRAM; U.S. chip stocks slide | KOSPI −5.46%; Micron about −4.5%; Philadelphia chip index about −1.5% (Reuters) |
| Oct 14 | ASML Q3 revenue slightly below expectations; TSMC results due Oct 15 | KOSPI −0.44% to 6,008.81; sixth straight decline |
Sources: Seoul Economic Daily, Reuters via KFGO, Forbes, TheStreet (Oct 13–14, 2026). Percentages for non-Korean names are as reported intraday.
Why one report moved a whole sector
Markets price second derivatives: the change in expected scarcity, not scarcity itself. Memory earnings are unusually sensitive to supply, because a modest oversupply can swing prices sharply. Investors who had capitalized several years of tight supply were therefore exposed to any hint that a new supplier could add bits. That is why a claim with no published specifications could still move SK hynix by double digits in a session.
Concentration amplified it. In Korea, two memory makers carry a large share of index weight, and foreign investors were net sellers for a seventh straight session by October 14 (Seoul Economic Daily; the one-day figure of roughly 4.5 trillion won on Oct 13 is single-source).
What 4F² actually means, and what remains unverified
4F² is a claim about cell density, and density only matters if yields and customers follow. In DRAM, cell area is expressed in multiples of the minimum feature size squared (F²). Industry coverage commonly describes today's mainstream cells as 6F² and a 4F² layout as a more compact design that typically depends on vertical-channel transistors. TrendForce has estimated a 30–50% density gain for the same area, and Reuters reported that Samsung and SK hynix are developing similar structures. Treat the technical framing here as context; the primary specifications have not been published.
| Evidence stage | What would count as proof | Status today |
|---|---|---|
| Design disclosure | Published specs, die size, bit density | Not published |
| Yield | Credible yield data at volume | Not published |
| Customer qualification | Server and PC makers certify the parts | Not reported |
| Volume timing | Capacity and shipment schedule | "By end-2026" server product plan, reported |
Sources: Reuters, Korea JoongAng Daily (citing SCMP), TrendForce estimate as cited in press, Oct 13, 2026.
How memory supercycles have ended before
Memory upcycles end when supply catches up with demand, rarely because demand disappears. The 2016–2018 run was followed by new fab capacity and an inventory correction. The 2020–2021 surge reversed when pandemic pull-forward demand faded. The current AI and HBM cycle has not ended, and the open question is what ends it. A new entrant that sells commodity DRAM at scale is one of the classic mechanisms, which is why this report was taken seriously even without data.
I examined the mechanics of HBM crowding out commodity DRAM in an earlier piece: The Memory Supercycle's Crowding-Out Problem. The relevant point here is that a rival entering commodity DRAM matters most when HBM absorbs incumbents' capacity.
The AI "casino" remark and the OpenAI revenue reset
The revenue reset was mostly accounting, but it still shifted the burden of proof onto AI spending. The Financial Times and others reported that OpenAI's annualized revenue was near $50 billion, about $20 billion below figures investors had seen, with the gap attributed largely to how partner-related revenue is counted. Separately, the Broadcom CEO's remark framed parts of AI investment as speculative. One outlet also cited a Broadcom long-range AI chip revenue target far above current levels (single-source), which shows why the same company can sound both bullish and cautious.
None of this shows demand falling. It does raise the standard of evidence investors will demand from AI capex, and memory is the most levered part of that chain.
Why rates matter as much as the headline
At a 5.2% ten-year yield, every long-duration growth story has to clear a higher hurdle. The U.S. 10-year closed near 5.21% on October 13, and Fed officials including John Williams signaled that one more hike this year could be appropriate. Futures priced roughly a 20–25% chance of a hike at the October 27–28 meeting, with figures varying by source. The September CPI release on October 14 was the last major data point before that meeting.
Higher funding costs also pressure the financing structures built around AI hardware. I covered that mechanism in GPU-Backed Loans Explained, and the same logic applies to memory makers' capex plans.
What to watch over the next two weeks
The next two weeks supply more evidence than the last two did. Several scheduled events will either support or undercut each leg of the current debate, so it helps to know which result matters for which question.
| When | Event | Why it matters |
|---|---|---|
| Oct 15 | TSMC Q3 results; U.S. PPI | Read-through on AI demand and capex plans; another inflation data point after CPI |
| Oct 27–28 | FOMC meeting | Tests the one-more-hike narrative and the 5%+ yield backdrop |
| Late October | Samsung and SK hynix full Q3 results (dates per company announcements) | Management commentary on HBM and conventional DRAM pricing and on competition |
| Ongoing | Any CXMT specification, yield, or customer-qualification disclosure | The only evidence that converts a press report into a supply-side fact |
Dates for the Korean memory makers' full results were not confirmed at the time of writing; check each company's investor-relations calendar.
Market Outlook: the most probable path and what would break it
The most probable path is a test, not a verdict: volatility persists until evidence arrives. The data point to a market that is demanding proof on three fronts at once: CXMT's specifications and yields, AI revenue quality, and the rate path. My judgment is that this base case carries roughly 55% weight, because none of the three can be resolved within a few weeks and earnings, DRAM contract prices, and the FOMC decision will each re-test the narrative.
The stress case (about 20%) is a credible ramp, such as customer qualification news, which would challenge memory pricing power. The relief case (about 25%) is a slow burn in which yield or equipment hurdles delay a rival and easing yields steady chip valuations. These weights are my judgment from current information, not forecasts.
What it means for Korea
Korea is the market where this debate is most concentrated and most visible. KOSPI fell about 14% in seven sessions from its October 2 peak of 7,003.74, and retail leverage added to the speed of the decline: margin loan balances stayed above 34 trillion won, near a three-month high, according to the Korea Financial Investment Association as reported by Hankyung and SBS Biz. I examine that retail-leverage channel in my Korean-language post on the sister blog, written for domestic readers rather than translated.
My view: separate the claim from the capacity
The most useful habit in a supercycle is to ask what has to be true for the cycle to end. Across the cycles I have watched since the 1990s, the turn was rarely announced by a headline. It showed up as a rival's yield curve, a customer's second-source decision, or an inventory line. A press report is a reason to start tracking those indicators, not a reason to conclude the cycle is over.
Basic investment principles for volatile, concentrated sectors
- Diversify across drivers, not just tickers. Several holdings that depend on the same AI-capex narrative are one position in practice.
- Size positions for the drawdown you cannot tolerate. A 10% single-day move in a large cap is possible, so plan for it in advance.
- Avoid funding volatile assets with borrowed money. Leverage turns a drawdown into a forced sale.
- Update on evidence, not headlines. Specifications, yields, qualification, and earnings are evidence; reports of progress are not.
These are general risk-management principles, not recommendations on specific securities, weights, or timing.
Related reading
- The Memory Supercycle's Crowding-Out Problem: How HBM Is Squeezing Commodity DRAM
- GPU-Backed Loans Explained: Nvidia's $500B Plan at 5% Yields
- Sister blog in Korean: Naver Blog (blog.naver.com/woduf00), by the same author and written independently for retail investors in Korea.
This article is for informational and educational purposes only and is not investment advice or a recommendation to buy or sell any security. Figures were cross-checked against public sources at the time of writing and may be revised. Investing involves risk, including loss of principal. Consult a qualified professional before making decisions.
Sources
- Reuters via KFGO, "US chip stocks extend slide as China memory report, AI spending jitters hit sector," Oct 13, 2026 — link
- Forbes, "Why Are AI Chip Stocks Falling? China's CXMT Advances And Broadcom CEO's Warning," Oct 13, 2026 — link
- Seoul Economic Daily, "KOSPI Plunges 6%, Falls Below 6,000," Oct 13, 2026 — link
- Babypips, "Chip Stocks Slide on Broadcom CEO AI Comments and CXMT Design Improvements," Oct 13, 2026 — link
- TheStreet, "Stock Market Today (Oct. 13, 2026)" — link
- Axios, "OpenAI annualized revenue $20 billion less than previously reported," Oct 8, 2026 — link
- Hankyung, margin-loan balance report (Korea Financial Investment Association data), Oct 13, 2026 — link
Comments
Post a Comment