The selling in Micron below $1,000 is not a verdict on the chip trade. It is a verdict on one layer of it. On the day the Kospi fell 3.65% and Korea's semiconductor index dropped 5.2%, Taiwan finished broadly flat. That gap is the signal. Capital is leaving the commodity layer of the AI hardware stack while the layers with structural pricing power hold their ground. Memory has been reclassified. It was a growth sector priced on the slope of its earnings. It is now a cyclical sector priced on the duration of those earnings. The investors who owned it on the first basis are selling to investors who will only own it on the second.
Micron is the last of the memory leaders to register this, not the first. The order of the decline matters more than its size.
Leaders Falling in the Order They Rose
Measured from the June 22 close, SK Hynix is down about 45%, while the Kospi and Samsung are each down roughly 30%. Micron is down close to 20% from its June record and is near the threshold of a bear market. Sandisk fell 7% on the session, and Goldman's TMT desk notes that both it and Micron have closed lower in six of their last seven sessions.
None of this tracks a deterioration in reported numbers. Barely two weeks ago Micron guided fiscal first-quarter revenue to $60 billion to $63 billion against a $56.8 billion consensus. Goldman then estimated that Micron alone would deliver 19% of the S&P 500's third-quarter earnings growth. Samsung's preliminary figures showed operating profit up nearly ninefold, and the stock sold off.
When a sector posts its best results ever and the shares fall, the market is not disputing the results. It is disputing what comes after them. Rich Privorotsky, who heads one-delta trading for Goldman in EMEA, framed the issue as "peak memory earnings despite extraordinary near term fundamentals." The important word is "despite." Nobody is challenging the next two quarters.
Positioning data shows the same shift among US funds. Goldman's prime book shows US information technology moving from the most net-bought sector in four of the previous five weeks to the most net-sold, at the fastest pace in seven weeks. Tech hardware, storage and peripherals, and semiconductors led the selling. That is a rotation inside tech away from the hardware most exposed to the memory price cycle. It is not a general exit from the asset class.
Contract Resets Are a Level, Not a Trend
The clearest evidence that the market is pricing duration rather than magnitude came from Taipei. Nanya Technology reported third-quarter revenue of NT$133.6 billion, up 62% quarter on quarter and 612% year on year, with a gross margin of 85.7% and an operating margin of 80%. Goldman's local desk called it "another astonishing quarter."
The composition of that quarter carries the argument. Blended average selling prices rose more than 60% quarter on quarter. Over half of that increase came from contract resets as long-term agreements signed in 2025 expired and were repriced. A reset is a step change. A contract signed at old prices moves to current prices once. After the reset, further earnings growth requires either another leg of price gains or more bits shipped.
Nanya's own outlook narrows both routes. Management expects double-digit contract price gains and gross margins above 80% to stay "achievable" in 2027. It also expects no bit shipment growth in early 2027, because channel inventory is near zero, and does not see volumes inflecting until new fab output begins in the fourth quarter of 2027. Margins that are already above 80% leave little room for further expansion.
That is a sector where the level of earnings stays near a record while the rate of change flattens. Growth capital pays for rate of change. Pictet's Jon Withaar described memory makers to Bloomberg as having "become value names, cheap at current earnings," and noted that "retail and fast money hedge funds do not chase value." Goldman's Taipei desk said much the same: there is "not much excitement recently to drive further re-ratings."
A low multiple on peak earnings is not a margin of safety when the buyer base is changing. It is the price at which the incoming value buyer agrees to take the stock off the outgoing growth buyer, and value buyers in cyclicals often wait for a visible trough before paying up.
Seoul Lost Its Marginal Buyer First
Korea did not lead the decline because Korean investors understood the memory cycle better. It led because its flow structure gave way first.
Samsung and SK Hynix ran a combined repurchase program of 55 trillion won, about $40 billion and the largest in Korean history. It ran faster than planned. Samsung's portion effectively ended on October 6, and Hynix has only a few sessions left. The Kospi has fallen in every session since and is down 8% across the three sessions after Samsung's buying stopped. Monday's close of 6,384.2 left the index just above its 200-day moving average at 6,376.
The flow data shows what took the corporate bid's place. Foreigners extended a five-day net-selling streak with $1.3 billion of tech sales on the day, including $782 million of Hynix and $415 million of Samsung. Retail bought $1.2 billion of tech, with $710 million going into Samsung alone, but Goldman's Seoul desk notes that retail net buying is shrinking.
The depth of the retail pool is falling too. Margin loans have dropped from a June peak of 38.6 trillion won to roughly 33 trillion. Investor deposits have fallen to around 100 trillion won from close to 140 trillion. Market turnover is down 70% from its late-May peak, and foreigners have withdrawn $131 billion from Korean equities this year, the most of any major Asian market.
Micron never had a single corporate buyer of that scale holding up its shares. Its resistance through the summer rested on US earnings revisions and benchmark ownership. That explains the lag. Korea's price discovery was suppressed by the buyback and then released all at once. Micron's is arriving more gradually as US funds reach the same conclusion through their own channels.
The 2028 Supply Curve Runs Through a Smaller Cell
The duration question needs a reason why 2028 would look different from 2027. China supplied a candidate over the weekend.
CXMT president Cao Kanyu told an industry event in Shanghai that the company has made "key progress" on a 4F² DRAM cell and plans to launch a server-memory product based on it by the end of 2026, according to the South China Morning Post. Mainstream commercial DRAM uses a 6F² cell. CXMT's design depends on a proprietary vertical surrounding gate transistor that limits current leakage while shrinking the cell footprint.
Cutting the cell from 6F² to 4F² reduces its area by a third. That allows up to roughly 50% more cells on a die at the same line width, before real-world overhead takes its share. The density comes from architecture rather than lithography. For a company effectively shut out of EUV tools, that is the gain that matters. SK Hynix only disclosed its 4F² roadmap in 2025. Samsung targets mass production of sub-10nm DRAM on the design by 2028, according to The Elec.
The announcement should be discounted heavily as a product claim. "Key progress" says nothing about yields. First-generation cell architectures often reach the market at poor yields and qualify slowly in server applications, where reliability requirements are strict. A December launch could easily turn out to be a sampling milestone rather than volume production.
As a supply-curve claim, though, it adds to evidence that was already strong:
- Capacity. Goldman expects CXMT's DRAM supply to reach 41% of Samsung's and 50% of SK Hynix's by 2028, up from 28% and 35% in 2025.
- Tools. The Financial Times reported in September that CXMT and YMTC have stockpiled enough ASML DUV equipment to cover three years of expansion. That weakens the export-control backstop many bulls relied on.
- Customers. Acer, HP and Asus already ship CXMT memory in PCs. Acer CEO Jason Chen said the shortage "has already completely disappeared" and expects memory prices to peak around mid-2027.
China already holds about 10% of DRAM revenue. The cell announcement changes the timing more than the direction. It suggests Chinese density gains could coincide with new Korean fab output in the same window.
A further detail deserves more attention than it has received. ASML is reportedly raising spare-part prices for Korean customers by 10%, and Samsung and Hynix have accepted the increase starting in January 2027. When the memory makers absorb a price rise from their toolmaker in the middle of their own record cycle, pricing power has moved upstream. Combined with Taiwan's steady session, this points to a rotation toward the parts of the stack whose economics do not depend on DRAM contract prices. That last link is an inference. No flow data yet confirms money moving into equipment or foundry names specifically.
The Bull Case Has Real Numbers Behind It
The countercase rests on actual figures, not hope, and it deserves a fair statement.
Nanya's 2027 outlook is not a peak-cycle guide in the usual sense. Double-digit contract price gains, margins above 80%, and near-zero channel inventory describe a market that is still short of supply. Micron's guidance beat consensus by more than $3 billion at the midpoint. Privorotsky also pointed to a duller explanation for the selling: some "travel and arrive" positioning after Samsung's numbers, which beat expectations without overwhelming them. If the decline is mostly a crowded-trade unwind, it ends once forced selling is exhausted, whatever CXMT does in 2028.
There is also a structural point on the bulls' side. CXMT's progress so far is concentrated in conventional DRAM. High-bandwidth memory, where qualification with accelerator customers is the binding constraint, is a separate market.
That argument has a weakness, and Goldman's Korea team identified it. One of the items it will listen for on Samsung's October 29 call is the effect of HBM "despeccing." Pressure on HBM specifications would reduce the premium that separates HBM from commodity DRAM. That would bring the differentiated part of the business closer to the commodity cycle the market is already discounting.
The thesis fails on one specific condition: credible evidence that earnings visibility extends past 2027. That would mean Samsung matching Micron's call for tight supply in 2028 and announcing new long-term agreements that run that far. Multiyear contracts at current prices would turn a reset-driven step into a durable stream, and durable streams are what growth capital buys back.
What Settles the Duration Argument
The next three weeks offer clean tests. ASML reports on Wednesday and TSMC on Thursday, with SEMICON West in the same week. If equipment and foundry results hold while memory keeps weakening, the rotation from commodity layers to toll collectors is confirmed rather than inferred. Samsung's October 29 call comes next, and the first item on Goldman's list is whether management describes 2028 supply as tight. SK Hynix reports in the last week of October, close to the point where its buyback runs out.
The price level to watch is the Kospi's 200-day average at 6,376. Monday's close was eight points above it. A sustained break, with foreigners still selling and retail buying still shrinking, would show that Korea's flow problem is unresolved. Korea has led this cycle in both directions, and Micron's drawdown is about 25 points shallower than Hynix's. A continued Korean break would point to further catch-down in Micron.
The read changes if foreign investors become sustained net buyers of Hynix after its buyback ends. That would be a buyer stepping in without corporate support. Until it happens, record margins in memory describe 2026 and say little about 2028, which is the year the market is now pricing.
