Snowshagal ResearchNo.112026.10.26

Has China's
DRAM Really
Caught Up?

The distance to the leaders differs by productCount the bits that can be sold, not just the fabs

The Big ThreeCXMTLane = product; distance = how far CXMT trails

Special Research · Information as of Oct. 26, 19:00 KST

Cover illustration: a yellow runner pursues a navy runner in three lanes labeled DDR4/DDR5, LPDDR6, and HBM. They are nearly level in DDR4/DDR5, one stride apart in LPDDR6, and far apart in HBM.

RESEARCH NO. 11OCT 26, 2026 (MON)SPECIAL RESEARCH

Has China's DRAM Really Caught Up?

China has made up ground. The question now is when, and by how much, its expansion will weigh on DRAM prices and Korean chipmakers' earnings.

China now has the world's fourth-largest DRAM supplier. But installed capacity, technological capability and genuine pricing power are three different things.

Snowshagal ResearchAbout 20 minutes to read

M1−5.46%KOSPI on October 13Second trading day after CXMT's announcement

M29.5%CXMT's share of DRAM revenueQ2 2026 · TrendForce

M32.8xSK hynix 12-month forward P/EOct. 21 · SK Securities

HOW THIS RESEARCH UNFOLDS

01KOSPI falls 5.46% after CXMT announcementWhat frightened the market?
02Measuring the gap product by productWhere the world's No. 4 supplier stands
034F²: smaller cells, cheaper chips?A visual guide to the technology and its cost
04Where Washington draws the lineHow much do export controls constrain China?
052028: from fabs to shippable bitsWhat matters more than factory count
06How much can HBM and long-term contracts protect?Reading a 2.8x forward P/E
07The question hanging over half the KOSPIThe next evidence to watch
ENDHow to judge the evidenceWhat is confirmed, estimated and still unknown

Information as of Monday, October 26, 2026, 19:00 KST · Before SK hynix (Oct. 28) and Samsung Electronics (Oct. 29) Q3 earnings releases

China does not need to make the best-performing DRAM to compete on price. Nor does building more fabs automatically guarantee a glut. This report examines the space between those claims through products, yield, shipments and contracts.

AT A GLANCE

CXMT's announcement lit the fuse, but it was not the only factor. The U.S. semiconductor ETF SOXX also lost 11.8% that week.
CXMT has become the world's fourth-largest DRAM supplier. It held 9.5% of global DRAM revenue in Q2, while its monthly wafer input capacity is estimated at more than 300,000 wafers.
The pace of catch-up varies sharply by product. A 2025 teardown put G4 roughly three years behind. G5's announced mass production still needs independent yield and cost verification; LPDDR6 has appeared in a commercially available Xiaomi phone. HBM presents a separate set of barriers.
New fabs do not automatically mean oversupply. Prices respond to saleable bits after yield losses, and to how the incumbent Big Three allocate wafers between HBM and conventional DRAM.
A 2.8x forward P/E reflects doubts about how long earnings can last. The multiple alone cannot tell us when or how far profits will fall. The first tests are the scope of long-term contract protection and HBM profitability at the two Korean companies.

Conceptual illustration of a 300mm wafer under amber light in a lithography room — CH 01300mm wafer · Lithography-room lightingConceptual illustration · Not a photograph

01WHAT HAPPENED THAT WEEK

The KOSPI fell 5.46% after CXMT's announcement.
What was the market afraid of?

The October 12–13 selloff combined CXMT's announcement with fears that semiconductor profits were peaking. U.S. inflation data on the evening of October 14 and Musk's “2028” posts on October 15 added to the anxiety later that week. The contribution of each factor cannot be isolated. The share-price decline is the starting point for this investigation, not proof of Chinese technological prowess.

On Saturday, October 10, Cao Kangyu, president of Chinese DRAM maker CXMT (ChangXin Memory Technologies), said at an industry conference that the company planned to launch server DDR5 modules using a new 4F² cell architecture before year-end.^{1, 2} It was a plan, with no product specifications, named customers or yield figures. The market started reacting on Monday, two days later.

October 10–16:
Six days of announcements and market reactions
Date	Event	Market reaction
Oct. 10 (Sat)	At the fourth Integrated Circuits and Chiplets Conference, CXMT announces plans to release 4F² server DDR5 RDIMMs by year-end	Weekend
Oct. 12 (Mon)	First trading session after the announcement	KOSPI −3.65% · Micron −3.88% · SanDisk −6.0%³
Oct. 13 (Tue)	Japan's first session after the Sports Day holiday on Oct. 12	KOSPI −5.46%(6,035.35) · Kioxia −13.28%^{4, 5}
Oct. 14 (Wed), evening	U.S. September headline CPI: +3.7% YoY (consensus +3.6%)^{6, 7}	Renewed rate concerns
Oct. 15 (Thu)	Musk posts on X that China “will make a lot of RAM in 2028” (as reported by Kyunghyang Shinmun)	Samsung Electronics −9.48% · SK hynix −11.06%⁸
Oct. 16 (Fri)	Weekly tally	KOSPI, weekly −14.5% · SK hynix −24.3% · Micron −14.6%⁹

The KOSPI's weekly decline was its worst since October 2008.⁸ Musk posted several times between October 11 and 16.

Two sharp selloffs this year;
both coincided with Chinese memory-chip news

SK hynix daily closing price · July 1–October 23, 2026 · KRW

Earnings expectations, AI spending concerns and interest rates were also moving on those dates. Source: Yahoo Finance (000660.KS daily closes)¹⁰and cited reports

Was China really the reason?

That same week SOXX fell 11.8%, Lam Research 17.2%, and Nvidia 6.6%. Equipment makers and accelerator companies, not just memory producers, came under pressure.¹¹
CXMT itself fell 5.2% on October 12 and another 2.8% on October 13. Power-equipment, nuclear and robotics shares also plunged at much the same time. Yuanta Securities sees a possibility that basket selling mattered more than semiconductor-specific sentiment.¹²
Hanwha Investment & Securities, by contrast, regards CXMT-driven memory weakness amid high rates as the trigger for the October 13 selloff.⁵ Brokerages differ in how much weight they give each cause.

That other sectors fell does not mean Chinese news had no effect. A similar episode unfolded in July. CXMT listed on Shanghai's STAR Market on July 27 and closed its debut session 466% above its IPO price. The same day, reports of domestically made immersion DUV lithography equipment in China helped send ASML shares down more than 8%.^{13, 14} SK hynix fell 14.6% the following day. Daishin Securities argues that Chinese memory headlines repeatedly act less as proof of a changed industry structure than as confirmation of existing fears that semiconductor earnings have peaked.¹⁵

We therefore separate two questions. In which products, and to what extent, is China already a competitive supplier (Chapters 02–04)? And when, and by how much, might that supply pressure prices and Korean chipmakers' profits (Chapters 05–06)?

CH 02Wafer-bin map analogyVolume · Technology · HBMConceptual diagram · Not measured data

02CATCHING UP, PRODUCT BY PRODUCT

Now the world's fourth-largest supplier,
but the gap varies by product

CXMT has quickly gained market share and manufacturing capacity. Technology must be assessed product by product, from prototypes to shipping devices. HBM faces different barriers from conventional DRAM.

“Caught up” cannot be reduced to a single league table. Process generation, product specifications, commercialization dates and mass-production economics are different measures. Start with CXMT's scale, then examine its products separately.

How big is CXMT?
Revenue, manufacturing capacity and actual supply are different measures
Metric	CXMT	Comparison	How to interpret it
Market positionRevenue share	9.5% (Q2 2026; No. 4 worldwide, TrendForce)¹⁶ · Counterpoint: 4% in Q2 2025, 8% in Q1 2026 and 10% in Q2 2026¹⁷	Samsung Electronics 39.4% · SK hynix 24.9% · Micron 23.3%	Rapid gains
Manufacturing baseMonthly wafer input capacity	About 320,000 wafers → forecast 420,000 in 2027 (Commercial Times)¹⁶	Estimated 350,000 by year-end (Daishin Securities)¹⁵	Expanding
Actual supplyShare of bit shipments	8% estimated for 2027 (UBS), versus 14% estimated capacity share¹⁸	Reflects lower yields and limits in high-value products (UBS)	Below capacity share
CostPer bit	DDR5 cost per bit estimated to be over 30% above the Big Three (July report)¹³	More exposed when selling prices fall	Estimate · Needs verification

Revenue share includes differences in pricing and product mix; it is not volume share. The two research firms' Q2 estimates, 9.5% and 10%, are nearly identical.

A product-by-product competitive map

Company claims, independent validation and real customers
are at different stages for each product
Product	Company's claims	Independent validation · Customers	Assessment
DDR4·DDR5G4 process	Commercially available	2025 teardown of a shipping product: 16nm feature size, 0.0020µm² cell area, approximately three years behind the Big Three^{19, 20} · PCs: adopted by HP and Dell¹⁸ · Servers: ByteDance supply agreement worth over US$7bn over five years; Tencent agreement over RMB 20bn (Reuters, citing multiple sources; contracts not public; Tencent's term variously reported as up to three or five years)^{21, 22} · Apple qualification testing (July)²³	Competes on price
G5 process24Gb LPDDR5X	Mass production announced Sept. 20 · Self-aligned quadruple patterning (SAQP) delivers an 11.95nm active-area half-pitch in the cell array · Over 50% more gross dies per wafer versus G4 (normalized to 8Gb)²⁴	No public teardown, yield or cost data · Customers undisclosed	Company-announcement stage
LPDDR6	Entry into mass production reported in the Aug. 29 interim filing and Sept. 7 company announcement · 16GB package; up to 12,800Mbps (10,667Mbps when paired with an SoC, in company testing)^{25, 26}	Xiaomi Chairman Lei Jun confirms first adoption in the 18 Fold²⁷ · Commercial product review confirms LPDDR6 in the highest-memory configuration²⁸ · No chip teardown, shipment-volume or yield data	Shipping product integration
Server DDR54F² architecture	Targeting a year-end launch (Oct. 10)¹	Specifications, qualification and yields all undisclosed	Planned
HBM	8-high HBM3 samples · Targeting 12-high HBM3E mass production in 2027²⁹	Estimated yield of roughly 25% on 8-high HBM3 (a July report citing SemiAnalysis estimates; methodology not disclosed)¹³ · Korean suppliers are shipping HBM4E samples¹⁸	Separate barrier

The 11.95nm figure is the half-pitch of active regions in the cell array, not directly comparable to the numbers in competitors' process-node names. The claim of “over 50% more dies” for G5 refers to gross die count normalized to an 8Gb capacity, relative to G4. It is not a yield-adjusted increase in shippable output.²⁴

In short, CXMT is already competing on price in standard DDR4/DDR5 and mobile DRAM. It has announced G5 mass production, but cost and yield remain difficult to verify. LPDDR6 has reached commercial deployment in a Xiaomi device, while the 4F² server product remains a launch plan. HBM is a different market, requiring die stacking, advanced packaging and customer qualification.

With supply tight today, CXMT is not necessarily selling at a discount. On July 24, Reuters cited two sources saying CXMT had recently charged more for certain 64GB DDR5 server modules than Samsung's roughly US$1,240 price.²¹ In early August, industry sources also said Apple had asked CXMT for LPDDR5X pricing below Samsung's and SK hynix's offers, but CXMT had quoted similar or higher prices.³⁰ These examples cannot be extrapolated to the pricing power of CXMT's entire portfolio. They do, however, challenge the assumption that Chinese chips are invariably cheap. Pricing pressure is likely to intensify when supply loosens.

Even the No. 4 supplier makes money in a shortage

CXMT's earnings:
from losses to RMB 50bn in quarterly profit within two years
Period	Revenue (RMB 100m)	Net income attributable to parent shareholders (RMB 100m)	Net margin
2024	241.8	−71.4	Loss
2025	618.0	18.75	3%
Q1 2026	508	247.6	49%
Q2	Approx. 995	528.4	53%
Q3 (preliminary range)	Approx. 1,147–1,197	Approx. 534–549	Approx. 46%

Q2 and Q3 revenue were derived from the preliminary nine-month figures and disclosed sequential changes. Q3 is an unaudited preliminary range. Sources: IPO prospectus^{31, 32}and preliminary nine-month results (Oct. 16)^{33, 34}

Based on the preliminary range, Q3 revenue grew 15–20% sequentially, while net income attributable to parent shareholders rose only 1–4%. CXMT did not explain the divergence. Huxiu proposes several hypotheses: a slowdown in TrendForce's forecast quarterly contract-price increases, from 90–95% for DRAM in Q1 to 13–18% for server DRAM in Q3; depreciation; taxes; and minority interests.²⁹ The pace of price increases slowed; prices did not fall. CXMT shares sank nearly 9% intraday on October 19, the first session after the preliminary results, touching RMB 41—the lowest level since listing.³⁴

In a shortage, even the fourth-largest supplier can generate substantial profits. The real test for a higher-cost producer comes when prices decline.

03THE TECHNOLOGY, EXPLAINED VISUALLY

4F²,
Smaller cells do not necessarily mean cheaper chips

4F² is an architecture that fits more cells into the same area at a given feature width. Investors should focus not on the architecture's name but on actual die area, good-die yield and cost per bit.

A DRAM die contains tens of billions of memory cells arranged like a grid. Each cell has a switch (transistor) and a tiny charge-storage vessel (capacitor). A stored charge represents one binary state and its absence the other. More cells per unit area increase density and capacity. But the actual manufacturing cost of each sellable die also depends on yield and process cost.

Using the same feature width,
cell area shrinks by one-third

6F² versus 4F² cell layouts · Each gray square is F×F

Conceptual illustration, not measured data. At the same F, a 4F² cell occupies two-thirds the area of a 6F² cell; this does not imply an equivalent improvement in whole-die density or cost.

Two routes to packing in more cells

Print finer lines (reduce F). Extreme ultraviolet (EUV) lithography is the most direct tool, but not the only route. In 2022 Micron produced its 1β DRAM without EUV, using multiple patterning steps to define finer features.³⁵ CXMT said its September G5 mass-production process also uses quadruple patterning to shrink the cell array.²⁴ The trade-off is more process steps and more demanding alignment control.
Reduce the squares per cell (6F² → 4F²). Stand a transistor vertically instead of laying it out horizontally. At the same F, the cell footprint shrinks by one-third.

CXMT is pursuing both approaches. ASML says it has never shipped an EUV system to China.³⁶ Without EUV, ever-smaller printed features require longer process flows. That makes 4F²—fitting more cells at the same feature width—an attractive alternative. Chinese technology outlet Kuai Ke Ji described a possible 30–50% increase in density within the same area, and CXMT President Cao Kangyu said the company had more than 300 published patent families related to 4F².¹

A horizontal transistor takes more area;
a vertical one takes less

Schematic cell cross-sections · Conventional horizontal channel (left) vs 4F² vertical channel with a surrounding gate (right)

Schematic only · Actual geometry and dimensions differ

Between a smaller cell and a cheaper chip

Shrinking cells is not the same as reducing the cost of sellable dies. A chip also needs peripheral circuitry and capacitor structures, while a new architecture changes the process flow and materials. What ultimately matters is how many good dies a wafer produces, and at what cost.

ILLUSTRATIVE ASSUMPTIONS · NOT CXMT DATA

Old architecture100 dies per wafer × 90% yield = 90 good dies

New architecture150 dies per wafer × 60% yield = 90 good dies

A 50% increase in dies per wafer yields no additional sellable chips if yield fails to keep pace. If the new process also costs more, the cost per good die can actually rise.

Vertical transistors can be prone to charge leakage. CXMT says its vertical surrounding gate (VSG) encloses the channel on all sides to suppress leakage.¹
Some proposed implementations use hybrid bonding: forming the cell array and peripheral circuitry separately, then joining them.² Daishin Securities (Oct. 15) argues that process and material expenses, including silicon-on-insulator (SOI) wafers, could increase costs by around 10%, and that full implementation would be difficult without EUV. These are estimates conditional on the report's assumptions.³⁷
At an October 21 webinar, Micron President and CTO Scott DeBoer argued that DRAM bit density is determined by capacitor pitch, not whether the cell is 6F² or 4F². He called the assumed density advantage of 4F² largely a myth, saying Micron had evaluated the technology but declined to pursue it because of added complexity and cost.³⁸ That is Micron's assessment for its own current roadmap.

A competitor's judgment is not proof of physical impossibility. Samsung unveiled a 16Gb prototype combining 4F² cells and wafer bonding at ISSCC in February, with commercialization reportedly expected after 2028; SK hynix is developing a vertical-gate 4F² approach.² CXMT stresses the density gains; Micron emphasizes practical density and cost constraints. The verdict must await shipping silicon: die area, good-die yield and cost per bit.

The new 4F² server product:
five gates from announcement to volume production

1Technology disclosurePapers · PatentsCompany's accountThree papers since 2021 and 300+ patent families relating to 4F² (company statements)¹
2Product specificationsCapacity · Speed · Scope of 4F² implementationUndisclosedReports suggest the first device may use 4F² in only part of the design (MoneyToday)²
3Customer qualificationCPU and motherboard compatibility testingUndisclosedServer modules must pass around-the-clock operation and hundreds of compatibility tests¹
4Product launchYear-end targetTrendForce calls the target “particularly ambitious”²
5Yield · High-volume productionUndisclosedA commercial launch is not equivalent to high-volume manufacturing readiness²

Kiwoom Securities points out that volume production, yield, performance and customer qualification have yet to be verified for the proposed server product.³ Not disclosed does not mean nonexistent. Of the five gates, only the first is confirmed in public materials.

04HOW FAR CAN EXPORT CONTROLS REACH?

Where Washington draws the line,
CXMT is already inside it

The United States defines “advanced DRAM” using measurable cell-area and density thresholds. Controls have changed the pace and cost of China's progress, but have not stopped it.

U.S. export-control thresholds for DRAM are numerical. Shipping U.S.-origin equipment or technology to Chinese fabs producing DRAM that crosses any one of the specified cell-area, bit-density or through-silicon-via thresholds requires U.S. government authorization.^{39, 40}

The threshold changed in just six weeks;
CXMT's G4 moved inside the new definition

U.S. “advanced DRAM” bit-density thresholds (Gb/mm²) and CXMT 16Gb DDR5 (G4)

The cell-area test moved in the same direction: under 0.0019µm² in the old rule versus under 0.0026µm² in the revised rule. CXMT G4 measures 0.0020µm². Sources: U.S. Commerce Department rule^{40, 41}and TechInsights¹⁹

The December 2024 rule defined advanced DRAM as having a cell area below 0.0019µm² or a bit density above 0.288Gb/mm². The January 16, 2025 amendment broadened the thresholds to 0.0026µm² and 0.20Gb/mm² and added dies with more than 3,000 through-silicon vias (TSVs). The Commerce Department explained that the earlier definition did not capture all DRAM equivalent to an “18nm half-pitch” and adjusted it to reflect actual production.⁴⁰ CXMT G4, as measured in a TechInsights teardown, fell outside the old thresholds but inside the revised definition.

Falling within the definition does not automatically prohibit every equipment transaction. The rules work through multiple provisions governing items, end uses and support by U.S. persons. Nor did falling outside the old thresholds confer immunity from other controls. The clear limitation is that controls narrow access to new equipment; they cannot undo already installed tools or accumulated know-how. CXMT announced G5 mass production and expanded capacity even after the revision.^{16, 24}

Equipment, HBM and restricted-party lists

U.S. measures affecting Chinese memory manufacturers
Measure	What it does	CXMT · YMTC
Advanced-node DRAM fabs	Authorization required to export U.S. equipment or technology to Chinese fabs producing DRAM with cell area under 0.0026µm², density above 0.20Gb/mm², or more than 3,000 TSVs per die	CXMT G4 meets the thresholds
HBM	Licensing requirement for HBM exports to China above 2GB/s/mm² of bandwidth density (Dec. 2024)⁴¹	Chinese customers face barriers to purchasing the latest Korean HBM
EUV	No EUV exports to China under Wassenaar-related restrictions and subsequent export controls³⁶	DUV multipatterning enables further scaling but increases steps and cost
Entity List	U.S. Commerce Department restricted-party list	YMTC added in December 2022⁴² · CXMT not listed
1260H	U.S. Defense Department list of “Chinese military companies”	CXMT added in January 2025; sued in August 2026 to challenge its designation⁴³

Restrictions on buyers are the next variable

Apple sought reassurance from the U.S. administration that CXMT would not be placed on the Entity List (FT, June 27) and began qualification testing in July.^{23, 44} Industry sources said price negotiations had broken down in early August.³⁰
On July 16, House Select Committee on China Chairman Moolenaar and others asked the Commerce Secretary to consider placing CXMT on the Entity List and banning purchases of Chinese memory by U.S. firms in areas including AI, data centers and federal IT.⁴⁵ On July 30, a group of seven senators, including Schumer, wrote to Apple demanding a commitment not to use CXMT or YMTC chips in any product worldwide.⁴⁶ These were congressional requests, not regulations enacted by the administration.
The “50% Affiliates Rule,” which extends restrictions to subsidiaries more than 50% owned by listed entities, is suspended until November 9 under the Federal Register. After the U.S.–China truce was extended in late September to January 10, 2027, an extension of the suspension was anticipated, but no extension notice had appeared by October 22.^{47, 48, 49} CXMT is not currently on the Entity List, so this rule does not directly apply to it. But if CXMT were listed, the rule would affect the reach of restrictions to its subsidiaries. It therefore matters as an indicator of policy scope and uncertainty.

How much Chinese memory reaches global markets in 2028 will not be determined by technology alone. Watch how U.S. restrictions change for particular companies, products and end uses, and whether they affect overseas customers' actual adoption.

Conceptual side view of stacked wafers reflecting rainbow colors — CH 05Stack of wafersIPO proceeds → Capacity investment → SupplyConceptual illustration · Not a photograph

05THE BEAR CASE, AND ITS LIMITATIONS

More fabs are coming.
Only saleable bits move prices.

The manufacturing base in Korea, the U.S. and China is set to expand in 2027–29. But a fab schedule is the starting point for a supply forecast, not proof that oversupply has been calculated.

The strongest version of the bear case runs as follows. DRAM is scarce today because there are too few fabs and cleanrooms. That bottleneck eases in 2027–28. On top of the new capacity comes Chinese supply driven by domestic-substitution objectives rather than near-term returns.

Manufacturing capacity grows in 2027–29,
but shipments will ramp in stages

Major memory fab schedules · Hollow dots = first cleanroom; solid dots = operation or mass-production target

Target dates based on company statements and reporting. SK hynix says cleanroom expansion and equipment installation will proceed in stages as demand warrants.⁵⁰ NAND fabs are shown separately from DRAM supply. Sources: TrendForce^{16, 51}and Tom's Hardware¹³

Morgan Stanley estimates that CXMT's monthly wafer capacity will reach 500,000 by 2028, overtaking Micron, and warns that Chinese suppliers may show less pricing discipline in a downturn.⁵²
Nomura assumes CXMT's share of global DRAM production rises to 18% by the end of 2028, from roughly 10% today.¹³
On October 15, Musk wrote, “China will make a lot of RAM in 2028.” The next day he added that rapid Chinese expansion in LPDDR5 and LPDDR6 could make 2028 a different story, while HBM would be “a much steeper hill to climb” (as reported by Kyunghyang Shinmun).⁸ That is the outlook of a major buyer, not a supply-and-demand model.

CXMT capacity forecasts
diverge to as much as 950,000 wafers a month by 2030

CXMT monthly wafer input capacity · 10,000 wafers · Estimates and forecasts by institution

Forecasts are not connected into one series because institutions use different bases. CXMT has never disclosed an official capacity figure. Sources: Morgan Stanley⁵²Taiwan's Commercial Times (citing TrendForce)¹⁶and Citrini Research (cited by Kyunghyang Shinmun)⁸

From wafers to saleable bits

Saleable bits=Wafer startsCapacity × Utilization×Bits per waferProcess · Product mix×Good-die yield×Share available to ship

Under Morgan Stanley's assumptions, growing from 300,000 monthly wafers in 2026 to 500,000 in 2028 represents a 67% capacity increase.⁵² That alone does not tell us how much global DRAM supply will rise. Taking UBS's 2027 estimates at face value—a 14% share of capacity and an 8% share of bit shipments—implies that CXMT ships roughly 53% as many bits per unit of capacity as the average of all other suppliers.¹⁸ Put another way, for every 100 units of bits produced from 100 units of capacity by the other suppliers, CXMT produces about 53. This combines utilization, process density and product mix, not just yield. The rate at which this efficiency improves matters as much as the number of new fabs.

Efficiency matters more than capacity:
calculating CXMT's share of bit shipments
Capacity share	Efficiency ≈53.4%Back-solved from UBS	70% efficiency	90% efficiency
14%UBS 2027 estimate	8.0%	10.2%	12.8%
18%	10.5%	13.3%	16.5%
22%	13.1%	16.5%	20.2%

Illustrative calculation. Efficiency means bit shipments per unit of CXMT capacity relative to the average for all other suppliers (set at 100%). Against the worldwide average, UBS's estimate is about 57%. Bit share = capacity share × efficiency ÷ (capacity share × efficiency + other suppliers' capacity share). The 18% and 22% capacity shares and 70% and 90% efficiency levels are hypothetical; changes in competitors' efficiency and product mix are excluded.

Comparing two cells in the table makes the point. Even at just 14% of capacity, reaching 90% relative efficiency would lift CXMT's bit share to 12.8%. Conversely, a 22% capacity share at roughly 53.4% efficiency gives a bit share of only 13.1%. That is why yield and process migration deserve attention before fab-opening dates.

The table stops at bit-share estimates. How much those bits might cut Korean producers' selling prices depends on the mix of competitive DDR5 and LPDDR products, how much incumbent supply they displace, and demand growth. This report does not assign a percentage to that effect or treat a 2028 glut as inevitable.

How an IPO finances expansion

IPO proceeds at China's two memory makers:
CXMT raised roughly twice its original target
Item	CXMT (DRAM)	YMTC (NAND)
Listing	July 27, 2026 · STAR Market	Filed Aug. 21 · Local expectations pointed to December listing (March 2027 if CXMT's timeline were applied mechanically)⁵³
IPO size	RMB 57.9bn (RMB 66.6bn including the overallotment) · Original target RMB 29.5bn⁵⁴	Target RMB 33bn⁵⁵
Use of proceeds	DRAM technology upgrades: RMB 13bn · Next-generation DRAM R&D: RMB 9bn · Wafer-line improvements: RMB 7.5bn³²	Production-line upgrades: RMB 20.8bn · R&D: RMB 12.2bn⁵⁵
Recent earnings	Nine-month net income attributable to parent shareholders: RMB 131.0bn–132.5bn (preliminary)³³	Q1 net income attributable to parent shareholders: RMB 33.38bn⁵⁵

When memory prices rise, earnings rise. Better earnings draw money through listings and equity issuance. Two or three years later, that money returns as additional supply. It is an old memory-cycle pattern; China's capital market is now amplifying it.

We are also beginning to see where the money goes. At an extraordinary general meeting on October 19, CXMT approved using RMB 18bn in excess IPO proceeds for two new projects. RMB 13bn will fund a technology R&D project within its Hefei wafer-fab complex (RMB 24.1bn total investment, mostly equipment); RMB 5bn will fund Phase II of a back-end testing facility for DRAM testing and module assembly (RMB 10.8bn total investment).^{56, 57} This shows IPO proceeds moving into technology and manufacturing infrastructure. Because the investment is not solely in wafer-fabrication capacity, however, these amounts cannot be converted directly into wafer starts.

Even if the chips stay in China, the impact does not

TrendForce expects CXMT's incremental capacity to serve mainly domestic Chinese demand.¹⁶ That does not make it irrelevant to Korean producers. When Chinese buyers replace imported memory with CXMT chips, Korean and U.S. producers lose sales in China even if CXMT exports nothing. Displaced supply then seeks buyers in other markets, potentially putting pressure on prices. The magnitude depends on Chinese customers' import mix and the substitutability of each memory product.

The Big Three's wafer allocation also shapes supply

China's additional wafer capacity cannot simply be added to global bit supply. Samsung, SK hynix and Micron allocate wafers between HBM and conventional DRAM. TrendForce projects HBM's share of the Big Three's DRAM wafer input to rise from about 18% at end-2025 to 22% at end-2026 and 30% at end-2027, increasingly crowding out conventional DRAM capacity.⁵⁸ HBM uses substantially more wafers to produce the same number of bits. TrendForce projects HBM's share of DRAM bit supply to rise from roughly 8% to 13% over that period. In other words, by end-2027 HBM could absorb 30% of the Big Three's wafer input yet account for only about 13% of their DRAM bit output.^{58, 59}

More wafers going to HBM leaves less capacity for conventional DRAM, with some of the gap potentially filled by Chinese expansion. But the Big Three's total wafer starts are also rising, so whether the absolute supply of conventional DRAM actually falls requires separate verification. Conversely, if HBM demand slows, the Big Three could redirect wafers to conventional DRAM. Process migration and qualification take time, so the change would not be instantaneous, but it could amplify price pressure if it coincides with Chinese expansion. A 2028 glut becomes plausible not from China alone, but if Chinese additions overlap with a ramp-up in conventional DRAM output by the Big Three and demand fails to absorb the combined supply.

Demand-side numbers matter too

On its July earnings call, Samsung said supply constraints could worsen in 2027 versus 2026 and that meaningful additional supply would be difficult to bring online before 2028.⁶⁰ On September 30 Micron likewise said supply-demand balances in fiscal 2027 and 2028 could be tighter than in 2026.⁵¹
On October 14 ASML said its planned equipment supply for 2027 and 2028 would still fall short of customer demand, and it was evaluating a 2028 capacity increase beyond its existing 30% expansion plan.⁶¹ Heavy equipment orders do not by themselves prove equally strong end-market demand for memory.
Morgan Stanley projects DRAM supply shortfalls of 17% in 2026 and 15% in 2027, putting oversupply risk in 2028 or later.⁵² TrendForce estimates a DRAM supply-fulfillment gap of approximately −1% to −2% in 2026, widening in 2027.⁵⁹ Both institutions expect shortages through 2027, but their estimates differ greatly in magnitude. We cannot verify that Morgan Stanley's shortage rate and TrendForce's supply-fulfillment rate use comparable formulas or coverage, so their figures should not be compared directly. Agreement extends only to the direction: continued shortages.
TrendForce raised its forecast for 2027 capex growth at nine major cloud companies to 60% (cited by DS Investment & Securities).⁶² The same firm expects 2027 server shipments to grow faster than in 2026 (+17%), while smartphone and laptop shipments are projected to fall again as rising component costs pass through to retail prices.⁵⁹ A 60% rise in capex does not imply 60% growth in DRAM bit demand.

Most of these forecasts come from suppliers and research firms and may contain optimism. Still, the 2028 fear is less “China caught up” than “everyone expands at once.” China's localization goals and state support could lead CXMT to make different production-cut decisions from the Big Three in a downturn. Whether it actually prioritizes market share over price will only become clear in the next downcycle.

The sequence matters too. Anticipation of capacity additions can influence contract prices before physical supply arrives. Whether lower prices persist, however, depends on realized shipments and demand.

False-color conceptual cross-section of stacked memory dies with through-silicon vias — CH 0612-high HBM cross-sectionConceptual illustration · Not a photograph

06HBM, LONG-TERM CONTRACTS AND 2.8x P/E

A 2.8x forward P/E:
how much does the market distrust future earnings?

One low multiple cannot tell us when earnings will decline or by how much. The question is under what conditions HBM and long-term contracts can protect profits.

On October 21, SK Securities calculated 12-month forward P/E ratios of 2.8x for SK hynix and 3.5x for Samsung Electronics. Both were below even the lowest multiples reached in past downcycles. SK Securities interpreted the valuations as pricing in a sharp earnings decline after 2028.⁶³ That is one possible interpretation. Low multiples can also reflect fears of peak earnings, distrust of the forecasts themselves, or a greater discount for risk.

Just two years of forecast FCF—2027 and 2028—
equal two-thirds of market capitalization

SK hynix market cap (Oct. 23) versus SK Securities' forecast free cash flow · KRW tn

FCF = operating cash flow − capital expenditure (SK Securities definition). This is an undiscounted sum, not cash necessarily returned to shareholders (SK Securities assumes 50% of FCF is returned). The 2026 estimate is shaded gray because most of that year has elapsed. Market cap = October 23 closing price of KRW 1,277,000 × 730.49m issued shares (including treasury shares). Source: SK Securities (Oct. 21)⁶³, Yahoo Finance¹⁰, DART⁶⁴

Under SK Securities' projections, just the 2027 and 2028 free cash flow would total KRW 625.5tn, or 67% of SK hynix's market capitalization. The gap between the market's valuation and the brokerage's earnings and cash-flow estimates is that large. Interpreting it requires examining both durability and risk. Even if profits hold up, a higher required return can keep the multiple depressed. For now, HBM economics and the scope of long-term contract protections are the best tests of earnings durability.

How much can HBM and long-term contracts protect profits?

What the two companies said about long-term contracts on July earnings calls:
what is public and what remains undisclosed
Item	Samsung ElectronicsJuly 30	SK hynixJuly 29
Customers	Agreements completed with five major data-center customers; five more in advanced discussions	Signed with approximately 10 customers, including key accounts; talking with other large customers
Duration	Five-year baseline, with one-year extensions negotiated annually	Typically around five years; varies by customer and product
Pricing	Minimum price floors for mainstream products, described as sufficient to reduce investment risk from market-price swings	Customer- and product-specific structures designed to manage price volatility; existence and level of floors not disclosed
Enforcement	Substantial prepayments, one-quarter of which has already been received	Financial mechanisms such as deposits, alongside long-term purchase commitments
Coverage	60–70% of planned DRAM and NAND capacity if contracts currently under negotiation are finalized	Share of revenue undisclosed; managed at an appropriate level according to market and customer demand
Not disclosed	Floor levels, covered products, and penalties for nonperformance	Pricing formulas, contracted share, and 2027 HBM prices and volumes (still being negotiated)

Samsung figures are based on remarks by EVP Kim Jae-june on July 30;⁶⁰SK hynix figures are from its July 29 results release and earnings call.⁶⁵ Both companies are expanding multi-year agreements, but undisclosed floor levels and covered volumes make the degree of protection hard to compare. Samsung's contracts do not protect SK hynix's earnings. A report suggesting that 75–80% of next year's supply is contracted⁸may use a different date, period and denominator and is therefore not directly comparable.

Long-term agreements cut both ways. Volume commitments and price floors can cushion earnings in a downturn, but can also limit additional price increases in an upcycle. TrendForce noted in July that multi-year contracts with U.S. cloud providers were constraining price increases for those customers; from Q3, server DRAM contract-price gains were expected to come mainly from customers without such agreements.⁶⁶ That is why the percentage under contract is not enough to gauge earnings protection. What matters is which products are covered, what price-adjustment terms apply, and how much supply remains uncontracted.

Nor is HBM always more profitable than conventional DRAM. TrendForce estimates that in Q1 this year, 64GB DDR5 server modules overtook HBM in revenue per wafer, and HBM profitability has been lower since that quarter. The same research argues that the more HBM crowds out conventional DRAM capacity, the stronger the case for large HBM contract-price increases in 2027.⁵⁸ How much HBM protects earnings thus depends on wafer allocation between the two product categories and the outcome of 2027 price negotiations.

Even with a cumulative 32% drop in DRAM prices,
profits remain at 82–88% of 2027 levels

DS Investment & Securities price sensitivity · SK hynix operating profit · KRW tn · 2027 forecast baseline KRW 406.8tn

Base case: DRAM prices −20%, −10%, −5% in successive years (cumulative ≈−32%). Bear case: −35%, −20%, −10% (≈−53%). Both assume SK hynix DRAM bit shipments grow 15%, 12%, 13% over those years. Numerical cost assumptions were not disclosed. Source: DS Investment & Securities (Oct. 19)⁶²

Profits decline less than prices because rising shipments partly offset weaker average selling prices. In the base case's first year, a 20% price drop paired with 15% bit-shipment growth implies DRAM revenue down about 8%. DS Investment & Securities also argues that customer roadmaps and contracts, rather than spot-market balances, primarily determine HBM prices.⁶² But the FCF estimates from SK Securities and the DS Investment & Securities scenarios rest on different assumptions. They must not be treated as parts of a single model.

Four arguments for resilient earnings—and their conditions

Daishin SecuritiesCost and equipment

4F² could raise costs by around 10%. The brokerage argues that operating margins could remain above 50% even if DRAM prices fell 50%, while shortages of ASML lithography tools would constrain aggressive capacity additions.³⁷

DS Investment & SecuritiesHBM and shipment growth

HBM prices are set by contracts and should decline less than conventional DRAM, while shipment growth partly offsets weaker prices. The brokerage also sees a smaller oversupply ratio than in past cycles.⁶²

SK SecuritiesLong-term contracts

Long-term agreements (LTAs) raise the price floor and dampen cycle volatility. Buyers who break commitments risk lower supply priority in the next cycle.⁶³

Kiwoom SecuritiesNew products remain unverified

CXMT's new server product has not yet demonstrated mass production, yield, performance or qualification. The brokerage sees it as premature to price a glut into chip stocks today.³

The claim of 50% operating margins after a 50% price decline assumes unchanged volume and costs. If revenue is 100 and cost is 25 (75% operating margin), halving the selling price still leaves a 50% margin; with cost of 30, the margin falls to 40%. Figures are illustrative.

The counterargument from the bear camp

Long-term agreements were struck in a shortage. Market participants quoted by Kyunghyang Shinmun argue that even if Chinese memory is not yet a direct substitute, its presence at the 2028 negotiating table will make today's prices harder to defend.⁸
Samsung's price floors apply to mainstream products, and the floor levels remain undisclosed. SK hynix has not even disclosed whether its contracts contain floors. Their levels, product scope and durations will determine how much protection they provide to earnings in 2028.⁶⁰
In DS Investment & Securities' bear case, SK hynix operating profit in 2030 falls to KRW 203.3tn—about half its projected 2027 level.⁶²

Four questions for the earnings calls

SK hynix: Wed, Oct. 28, 09:00 KST · Samsung Electronics: Thu, Oct. 29, 10:00 KST

Long-term agreements What has changed since July: customer count, finalized coverage, and which products and periods are subject to floors and adjustment provisions
2027 HBM Pricing negotiations; HBM4 and HBM4E volumes; allocation of wafers between HBM and conventional DRAM
Capital spending 2027 budget; timing differences between cleanroom and equipment investment; actual shipment ramp
Chinese competition Commentary on conventional DRAM pricing and sales to Chinese customers

Separate what was already disclosed about LTAs in July from genuinely new disclosures. Companies may not reveal customer-specific pricing or penalties.

CH 07The KOSPI as a silicon waferOct. 16 market cap: KRW 4,521tnOne die ≈ 0.3% of market cap

07READING THE KOREAN MARKET

The question hanging over half the KOSPI:
look for evidence, not announcements

The two companies account for 48% of the KOSPI's market capitalization. Evidence that could change the assessment will emerge over the next two months.

Of the KOSPI's KRW 4,521tn market capitalization on October 16, Samsung Electronics common shares accounted for KRW 1,260tn and SK hynix for KRW 929tn.⁹ That is why Chinese DRAM headlines can shake the entire index. But the companies' share of market cap is not the same as DRAM's share of their profits. Samsung also makes mobile devices and consumer electronics and operates a foundry business.

According to Hana Securities, overseas investors sold a net KRW 11.1611tn of KOSPI stocks during October 12–16, including KRW 5.3565tn of SK hynix and KRW 3.4763tn of Samsung Electronics.⁹
Yet semiconductor earnings estimates at brokerages rose that same week. Hana Securities reported a 0.57% upward revision in consensus SK hynix EPS estimates over the week.⁶⁷

THE NEXT EVIDENCE CALENDAR

10.28WedSK hynix Q3 results and earnings callLong-term contracts · HBM pricing · Capex

10.29ThuSamsung Electronics final Q3 results and earnings callStatus of contracts outlined in July · Any commentary on China

10.31DeadlineCXMT's formal Q3 filingConfirmation of nine-month preliminary results · Reason for margin decline

11.09U.S. 50% Affiliates Rule suspension set to expireWhether a notice extends the suspension to match the truce

Year-endCXMT target launch of 4F² server DDR5Actual shipments · Server-platform qualification

Signs that would increase concern

CXMT actually ships its 4F² server modules by year-end and announces server-platform qualifications
Korean memory sales to China fall while CXMT shipments rise
Renewals of long-term agreements weaken pricing terms for 2028 and beyond

Signs that would ease concern

The year-end launch proves limited to samples, small batches or partial use of 4F²
Earnings calls disclose quantified coverage and protection periods for finalized LTAs
CXMT's margin decline proves to reflect its own yield or cost problems rather than a fall in industry prices

The meaning of CXMT's margin decline depends on its cause. Company-specific yield or cost issues would benefit incumbents; an early decline in industrywide prices would hurt Korean producers too.

ENDTHE EVIDENCE STANDARD

What is confirmed, what is estimated,
and what remains unknown

KNOWNConfirmed facts

CXMT plans to release a 4F² server DDR5 product by year-end (Oct. 10)
CXMT held 9.5% of DRAM revenue in Q2, ranking fourth worldwide
Structures of Samsung's and SK hynix's multi-year agreements (July earnings calls)

ESTIMATEForecasts and estimates

CXMT monthly capacity of 500,000 wafers in 2028 (Morgan Stanley)
CXMT 2027 bit-shipment share of 8% (UBS)
DS Investment & Securities base case: even with a cumulative ≈32% DRAM price decline over three years, SK hynix 2028–30 operating profit remains at 82–88% of the 2027 forecast

UNKNOWNStill unknown

Die area, yield and cost per bit for CXMT's 4F² product
Finalized LTA coverage and price floors or adjustment provisions
Wafer allocation between HBM and conventional DRAM after 2028, and the pace of switching
Direction of U.S. listings and purchasing restrictions affecting CXMT

CONCLUSION

China has become home to the world's fourth-largest DRAM supplier, but its progress differs by product. Even with a technological gap, CXMT can already compete on price in conventional DRAM. And more fabs do not make a glut inevitable.

The effect on Korean chipmakers' earnings will not be determined by Chinese shipments alone. The Big Three's wafer allocation between HBM and conventional DRAM, product-level pricing and long-term contract terms all matter. The evidence to watch, in order, is contract and HBM disclosures on earnings calls, actual shipments of the year-end 4F² product, and changes in CXMT's yields and costs.

This report examines a question using public information; it is not a trading recommendation. Source notes specify the relevant dates and calculation methods.

Glossary

DRAM cell: One transistor and one capacitor forming a storage element. The capacitor's stored charge encodes a binary 1 or 0.
F, 6F², 4F²: F is the minimum printable feature width of a process. 6F² and 4F² express the area occupied by one cell in units of F×F squares.
Half-pitch: Half the period of a repeating line-and-space pattern. Because measurement locations differ across companies, this figure should not be equated directly with the numbers in process-node names.
Multiple patterning · SAQP: A method of creating fine patterns through several lithography and pattern-transfer steps. Self-aligned quadruple patterning (SAQP) divides pitch by four but adds process steps.
VSG: Vertical Surrounded Gate: a structure in which the gate wraps around a vertical transistor channel on every side. CXMT says it uses this design in its 4F² approach.
Hybrid bonding: A technology that fabricates the cell array and peripheral circuits on separate wafers, then joins them through direct copper-pad bonding.
EUV · DUV: Extreme-ultraviolet and deep-ultraviolet lithography equipment. Shorter-wavelength EUV can print finer features in fewer exposures; DUV may require multiple patterning steps, lengthening the process flow.
Yield: The percentage of fabricated chips that meet specifications. Wafer-level die yield differs from the yield after stacking and packaging.
RDIMM: Registered dual in-line memory module for servers. An onboard register buffers signals to support larger, more stable memory configurations.
Conventional DRAM: Standards-based DRAM for PCs, servers and smartphones (DDR4, DDR5, LPDDR, etc.). Interchangeability across suppliers enables direct price competition. Distinguished here from HBM, with its demanding customer-specific qualifications.
HBM · TSV: High-bandwidth memory: stacked DRAM dies connected by thousands of through-silicon vias (TSVs), typically placed alongside AI accelerators.
Monthly wafer input capacity: The number of 300mm wafers a fab can start in production each month. Actual saleable bits depend on utilization, yield and product mix.
Supply shortfall rate · Supply-fulfillment rate: Measures of the extent to which supply falls short of demand. A negative supply-fulfillment rate means supply is below demand. Research firms use different formulas, so the figures may not be directly comparable.
LTA: Long-Term Agreement: a multi-year supply contract setting volume commitments and pricing terms in advance.
12-month forward P/E: Share price divided by expected earnings per share over the next 12 months. A lower figure generally indicates a lower valuation relative to forecast earnings.
FCF: Free cash flow: cash generated by operations minus capital expenditures. Not equivalent to cash actually returned to shareholders.
STAR Market · Excess IPO proceeds: The STAR Market is the Shanghai Stock Exchange's technology-company board. Excess IPO proceeds are funds raised above the original offering target. CXMT's board and extraordinary general meeting approved how to allocate them.
Entity List · 1260H: The Entity List is a U.S. Commerce Department restricted-party list that generally triggers export licensing requirements. Section 1260H is a U.S. Defense Department list of “Chinese military companies,” but does not itself directly ban exports.

Sources · Cutoff dates and calculation notes

Information cutoff: Monday, October 26, 2026, 19:00 KST. Prepared before Q3 earnings releases from SK hynix (Oct. 28) and Samsung Electronics (Oct. 29).

Cutoff dates · Methodology · Differences among sources

Item	Details
Share prices and indices	Korean regular-session closing prices (Yahoo Finance 000660.KS); KOSPI closes as reported by brokerages. U.S. and Japanese prices use each local market's close.
Market capitalization	Oct. 23 closing price KRW 1,277,000 × 730,492,365 issued shares (share count before cancellation, including treasury shares, per Aug. 19 DART cancellation filing) = approximately KRW 932.8tn. KOSPI weights calculated from Hana Securities' Oct. 16 market caps (Samsung common shares 27.9%, SK hynix 20.6%).
FCF	SK hynix FCF in SK Securities' (Oct. 21) shareholder-return scenario (operating cash flow − capex). 2027E KRW 268.57tn + 2028E KRW 356.96tn = KRW 625.53tn (rounded to KRW 625.5tn); divided by KRW 932.8tn market cap = 67%. Including 2026E (KRW 170.7tn) would bring the figure to 85%, but most of 2026 has already elapsed and it cannot all be regarded as future cash flow. Undiscounted addition; not equivalent to shareholder distributions.
CXMT quarterly results	Q1 from the updated prospectus. Q2 revenue (approximately RMB 99.5bn) back-calculated from preliminary nine-month revenue of RMB 265bn–270bn and a Q3 sequential revenue increment of RMB 15.2bn–20.2bn. Q3 net income attributable to parent shareholders = preliminary nine-month total minus Q1 and Q2. Margin = net income attributable to parent shareholders ÷ revenue. Q3 figures are unaudited preliminary estimates.
Market share	By revenue. TrendForce (Q2 2026: 9.5%) and Counterpoint (Q2 2026: 10%) use different methodologies. Omdia estimated 7.67% in Q4 2025 (cited in the prospectus). Revenue share reflects price and product mix, not just shipment volume.
Capacity	Monthly wafer input capacity. Estimates are not joined into one chart series because institutions use different definitions and reference dates. CXMT has not officially disclosed a number.
Bit-share calculation	Backing out relative efficiency r from UBS's 2027 estimates (capacity share 14%, bit-shipment share 8%): 0.08 = 0.14r ÷ (0.14r + 0.86), giving r ≈ 0.534. Here r is efficiency relative to the average of all other producers excluding CXMT (versus the worldwide average: 0.08 ÷ 0.14 ≈ 0.571). Table formula: bit share = s×r ÷ (s×r + 1 − s), holding all other suppliers' efficiency constant at 1.
DS scenarios	Three-year cumulative ASP change calculated by multiplying annual factors: base case 0.8×0.9×0.95 = 0.684 (about −32%); bear case 0.65×0.8×0.9 = 0.468 (about −53%). Assumed SK hynix DRAM bit growth in 2028, 2029 and 2030: 15%, 12%, 13% (DS Table 1). First-year revenue effect: 0.8×1.15 = 0.92 (about −8%).
Regulations	The current definition in 15 CFR 772.1 reflects the January 16, 2025 amendment (any one of the three tests qualifies as advanced DRAM). Old values under the December 2, 2024 rule are from law firm Covington's interpretation. G4 figures are from 2025 reports of TechInsights' teardown; no public teardown of G5 is available.
Company statements	CXMT's G5, LPDDR6 and 4F² figures, Samsung's explanations of LTAs and ASML's supply forecast are company statements, distinguished here from independent verification. Musk's remarks are based on Kyunghyang Shinmun reporting.
Timetable	Target dates from company announcements and news reports. Earnings-call times from each company's IR notices (SK hynix Oct. 28, 09:00 KST; Samsung Oct. 29, 10:00 KST).
Generated illustrations	The wafer and HBM artwork opening Chapters 1, 5 and 6 consists of code-generated conceptual illustrations, not actual photographs.

Source references

Kuai Ke Ji (republished by Sina Tech) · “Domestic DRAM breakthrough: CXMT to unveil 4F architecture server memory by year-end” · Oct. 10, 2026 · CXMT President Cao Kangyu at the 4th Integrated Circuits and Chiplets Conference (Oct. 9–10): target launch of 4F² server DDR5 RDIMMs by year-end · Article suggests 30–50% higher density in the same area · VSG transistor · Over 300 published patent families related to 4F² · Three papers since 2021 · Server-module requirements: continuous operation, hot-swap capability and hundreds of compatibility tests · finance.sina.com.cn
TrendForce News · “CXMT Accelerates 4F² DRAM Push, Eyes Year-End Server DDR5 RDIMM Launch” · Oct. 12, 2026 · Q2 2026 DRAM revenue share: CXMT 9.5% (Q1: 7.6%), Samsung 39.4%, SK hynix 24.9%, Micron 23.3% · Announcements of fifth-generation (G5) and LPDDR6 mass production in September · Hybrid bonding · Samsung ISSCC 2026 16Gb prototype combining 4F² and wafer-to-wafer (W2W) bonding, with commercialization reportedly after 2028 · SK hynix vertical-gate 4F² · Year-end target “particularly ambitious”; commercial launch ≠ high-volume production readiness · Cites MoneyToday report suggesting partial 4F² implementation · trendforce.com
Kiwoom Securities, Han Ji-young · “U.S. Stocks Mixed Despite Earnings Optimism as China-Linked Chip Weakness Weighs” (market commentary) · Oct. 13, 2026 · Oct. 12 Micron −3.9%, SanDisk −6.0%, Philadelphia Semiconductor Index −2.7% · KOSPI Oct. 12 at 6,384.24 (−3.65%) · CXMT has not demonstrated mass-production success, yields, performance or qualification · Brokerage PDF (no public URL) ·
IBK Investment & Securities, Kim Ye-seul · “Commentary on the Japanese Equity Selloff” · Oct. 14, 2026 · Oct. 13 Nikkei 225 −1.82%, Kioxia −13.28%; news of CXMT technological progress added pressure to semiconductor shares · Brokerage PDF (no public URL) ·
Hanwha Investment & Securities, Ha Jang-kwon · “Market Prices In Not Just Growth Disappointment But Falling ROE” (selloff commentary) · Oct. 14, 2026 · Oct. 13 KOSPI −5.5% (6,035.35); CXMT-driven memory weakness amid high rates viewed as selloff trigger; too early to regard a plan as a threat to the oligopoly · Brokerage PDF (no public URL) ·
U.S. Bureau of Labor Statistics (BLS) · “Consumer Price Index - September 2026” · Oct. 14, 2026, 08:30 ET · Headline CPI +3.7% over 12 months; core CPI +2.6% · bls.gov
IBK Investment & Securities, Investment Analysis Department · Market commentary · Oct. 15, 2026 · U.S. September CPI +3.7% YoY (consensus +3.6%); core +2.6% · Brokerage PDF (no public URL) ·
Kyunghyang Shinmun (English edition) · “Musk floats '2028 China DRAM'… A big buyer's ploy to shake prices?” · Oct. 19, 2026 · Musk's X posts Oct. 11–16 (Oct. 15: “China will make a lot of RAM in 2028”; Oct. 16: LPDDR5/6 and HBM references) · Oct. 15 Samsung −9.48%, SK hynix −11.06% · Previous week's KOSPI −14.46% (worst weekly fall since October 2008) · Report that 75–80% of Samsung's next-year supply is under LTAs · Citrini Research forecast CXMT capacity of 350,000 wafers a month this year, rising to 950,000 in 2030 · Industry view that “prices now may be hard to defend at the negotiating table in 2028” · khan.co.kr
Hana Securities, Kim Rok-ho · “DRAM Price Forecasts Rise as Fear Peaks” (Semiconductor and Equipment/Materials Weekly) · Oct. 19, 2026 · Week of Oct. 12–16: KOSPI −14.5%, Samsung −17.8%, SK hynix −24.3% · Table 1, Oct. 16 market caps: KOSPI KRW 4,521.029tn, Samsung KRW 1,259.873tn, SK hynix KRW 929.186tn; weekly foreign net buying: KOSPI −KRW 11.1611tn, Samsung −KRW 3.4763tn, SK hynix −KRW 5.3565tn · Table 2: Micron one-week −14.6% · Brokerage PDF (no public URL) ·
Yahoo Finance 000660.KS (SK hynix, Korean regular session), daily closes · July 1–Oct. 23, 2026 · July 27 KRW 1,816,000 → July 28 KRW 1,550,000 · Oct. 8 KRW 1,681,000 → Oct. 12 KRW 1,605,000 → Oct. 13 KRW 1,443,000 → Oct. 14 KRW 1,419,000 → Oct. 15 KRW 1,262,000 → Oct. 16 KRW 1,272,000 · Oct. 23 KRW 1,277,000 · finance.yahoo.com
Kyobo Securities, Choi Bo-young · “Look Beyond Capital Constraints to Medium-Term Semiconductor Earnings” · Oct. 19, 2026 · Oct. 12–16: SPY −2.0%, QQQ −5.5%, SOXX −11.8%, Nvidia −6.6%, Broadcom −9.1%, Micron −14.6%, Lam Research −17.2% · Brokerage PDF (no public URL) ·
Yuanta Securities, Koh Kyung-beom and Kim Hye-won · “Situational Awareness: High-Frequency Trading in Asian TRS, Domestic CFDs, DMA, and Open API Scalping” · Oct. 14, 2026 · p. 7: CXMT shares Oct. 12 −5.2%, Oct. 13 −2.8%; simultaneous declines in power-equipment, nuclear and robotics shares; possible basket-level selling · Brokerage PDF (no public URL) ·
Tom's Hardware · “Chinese memory maker CXMT posts blistering 466% leap in Shanghai IPO, bulk of spending to be focused on DRAM production, no HBM in sight” · July 27, 2026 · IPO price RMB 8.66, first-day close RMB 49 (+466%) · Planned use of proceeds RMB 29.5bn · Estimated 25% yield on 8-high HBM3 (SemiAnalysis model) · DDR5 cost per bit reportedly over 30% higher than the Big Three · Server revenue share 8.4% in 2024 → 26.5% in 2025 · Nomura's assumed DRAM production share: around 10% now to 18% by end-2028 · Shanghai fab expected to begin mass production in 2027 · tomshardware.com
The Standard (Hong Kong) · “China starts production of home grown immersion DUV chipmaking tools” · July 28, 2026 · Coverage of July 27 reports in The Information and Reuters; ASML shares fell more than 8% · thestandard.com.hk
Daishin Securities, Lee Kyung-min · “Why the KOSPI Plunged, Key Support Levels and the Next Inflection Point” (Daishin Strategy) · Oct. 14, 2026 · Oct. 13 KOSPI 6,035.35 (−5.46%), intraday sell-side circuit breaker for program trading · 12-month forward P/E 4.60x (2008 trough: 6.27x) · Table 4 discusses CXMT 4F² concerns and counterarguments (year-end monthly capacity 350,000 wafers; news repeatedly reinforces peak-earnings fears more than proving structural change) · Table 5 covers late-July news of CXMT's Beijing fab · Brokerage PDF (no public URL) ·
TrendForce News · “CXMT, YMTC Ramp Memory Capacity, but China's AI, Cloud Boom Could Soak Up New Supply Through 2027” · Sept. 24, 2026 · Citing Commercial Times: CXMT monthly wafer input capacity around 320,000, rising to 420,000 in 2027; YMTC Wuhan Fab 3 to start operations in late 2026 and ramp in 2027 · Most new capacity expected to serve Chinese domestic demand; global supply-demand impact likely after 2027 · Q2 2026 market share 9.5% · trendforce.com
Counterpoint Research · “Global DRAM and HBM Market Share: Quarterly” · Updated Sept. 1, 2026 · CXMT DRAM revenue share 4% in Q2 2025, 8% in Q1 2026, 10% in Q2 2026 (No. 4) · Q2: Samsung 38%, SK hynix 25%, Micron 24% (quarterly table cross-checked against Xenospectrum, Sept. 4) · counterpointresearch.com · xenospectrum.com
MoneyToday (English edition) · CXMT IPO article · July 28, 2026 · Jefferies: 1.5–2 generation technology gap · Korean producers' HBM4E samples versus CXMT's HBM3 samples · UBS: capacity share 12% → 14%, bit-shipment share 8% · Adopted by HP and Dell · mt.co.kr
SemiWiki (citing TechInsights teardown) · “China's CXMT's DDR5 reveals 3-year technology gap” · July 31, 2025 · CXMT G4 16Gb DDR5: F 16.0nm, cell area 0.0020µm², die area 66.99mm², bit density 0.239Gb/mm², wordline pitch 41.7nm and bitline pitch 47.9nm · semiwiki.com
Electronics360 · “Chinese memory vendors catching up quickly” · Feb. 19, 2025 · TechInsights: “three-year technology gap”; the Big Three had already brought 16Gb DDR5 into mass production in 2021 · electronics360.globalspec.com
Reuters · “China’s memory chip makers ride AI boom to new power, and U.S. scrutiny” (reproduced by WMBD) · July 24, 2026 · Three sources: CXMT–ByteDance five-year deal worth over US$7bn, Tencent June agreement worth more than US$3bn · Two sources: CXMT recently sold certain 64GB DDR5 server modules for more than Samsung's roughly US$1,240 price · Companies did not respond · wmbdradio.com
Reuters · “China’s CXMT wins 20 billion yuan memory supply deal with Tencent, sources say” (republished by The Standard) · June 29, 2026 · Three sources: server DRAM supply agreement worth more than RMB 20bn; two sources described term of up to three years, one said up to five years · thestandard.com.hk
MacRumors · “Apple Begins Testing Controversial Chinese Memory Chips” · July 8, 2026 · Cites Financial Times reporting that CXMT DRAM is in Apple's qualification-testing phase · macrumors.com
CXMT Newsroom · “CXMT Announces Mass Production of 5th-Generation DRAM Technology Platform” · Sept. 20, 2026 · G5 mass production; cell-array active-area half-pitch of 11.95nm with self-aligned quadruple patterning (SAQP) · Over 50% more gross dies per wafer versus G4 when normalized to 8Gb · Two 24Gb LPDDR5X dies brought into mass production · Yield, cost figures and customer names undisclosed · cxmt.com
EqualOcean · “CXMT Claims First Mass Production of LPDDR6, With Xiaomi 18 Fold as Launch Customer” · Aug. 31, 2026 · Aug. 29 interim filing claims LPDDR6 commercialization, up to 12,800Mbps and 16GB; Xiaomi 18 Fold integration (company announcement) · equalocean.com
CXMT Newsroom · “CXMT Pioneers Mass-Production of LPDDR6” · Sept. 7, 2026 · Entry into mass production, initial adoption in Xiaomi 18 Fold, 16GB package, up to 12,800Mbps (10,667Mbps when paired with SoC); performance figures based on CXMT's own tests · cxmt.com
The Korea Herald · CXMT LPDDR6 mass-production story · Aug. 31, 2026 · Xiaomi Chairman Lei Jun separately confirmed Xring O3 support for CXMT LPDDR6 and first adoption in the 18 Fold · koreaherald.com
Notebookcheck · “Xiaomi 18 Fold review: Big display, big performance, big compromises?” · Oct. 20, 2026 · Review of a commercially available device; confirms LPDDR6 in the highest-memory (16GB) configuration but does not identify memory-chip manufacturer · notebookcheck.net
Huxiu (虎嗅) · “CXMT forecasts over RMB 130bn profit for first nine months; Q3 attributable net margin falls to around 46%” · Oct. 2026 · Q3 sequential revenue increase RMB 15.2bn–20.2bn; net income attributable to parent shareholders increase RMB 0.55bn–2.05bn · Parent-attributable net margin Q2 53.1% → Q3 about 46% · TrendForce estimates quarterly DRAM contract-price increases: Q1 +90–95%, Q2 +58–63%, Q3 +13–18% · Target 12-high HBM3E mass production in 2027 · huxiu.com
News Tomato · “CXMT Turns Down Apple: A Boost to Korean Chipmakers' Negotiating Power?” · Aug. 7, 2026 · Industry sources say Apple sought LPDDR5X prices below those offered by Samsung and SK hynix, but CXMT refused and quoted similar or higher prices, supported by firm domestic demand · v.daum.net
Yicai (第一财经) · “CXMT Updates Prospectus: Q1 Net Income Exceeds RMB 20bn” · May 2026 · Net loss attributable to shareholders RMB 16.34bn in 2023 and RMB 7.14bn in 2024 · 2025 gross margin 41.02% · Q1 2026 revenue RMB 50.8bn; net income attributable to parent shareholders RMB 24.762bn · yicai.com
The Beijing News (新京报) · “After RMB 36.6bn in Accumulated Losses, CXMT Seeks STAR Market Listing” · May 2026 · Prospectus: 2025 revenue RMB 61.799bn, net income attributable to parent shareholders RMB 1.875bn, fab utilization 95.73%, three 12-inch fabs in Hefei and Beijing · Planned IPO proceeds: DRAM upgrades RMB 13bn, wafer-line improvements RMB 7.5bn, next-generation DRAM R&D RMB 9bn · LPDDR customers include Xiaomi, OPPO, vivo, Transsion and Lenovo · bjnews.com.cn
IT Home (IT之家) · “CXMT Expects RMB 131bn–132.5bn in Net Profit Attributable to Shareholders for the First Nine Months of 2026” · Oct. 16, 2026 · Nine-month preliminary revenue RMB 265bn–270bn and net income attributable to parent shareholders RMB 131bn–132.5bn (unaudited) · ithome.com
Sina Finance (新浪财经) · “CXMT Shares Drop Over 8% Intraday Despite Nine-Month Profit Above RMB 130bn as Q3 Growth Slows Sharply” · Oct. 19, 2026 · Intraday drop approached 9%, low RMB 41 (lowest since IPO debut) · Q2 net income attributable to parent shareholders RMB 52.843bn; Q3 sequential rise of about 1.0–3.9% · finance.sina.com.cn
Micron press release · “Micron Ships World’s Most Advanced DRAM Technology With 1-Beta Node” · Nov. 1, 2022 · 1β process used proprietary multiple patterning and immersion lithography without EUV, improving bit density by more than 35% over 1α · investors.micron.com
Tom's Hardware · “ASML denies US government report that its EUV chipmaking tool was shipped to China” · June 19, 2026 · ASML: “has never shipped an EUV machine to China”; shipments constrained under the Wassenaar Arrangement and later export controls · tomshardware.com
Daishin Securities, Ryu Hyung-keun · “An Excessive Selloff” (Semiconductor Issue & News) · Oct. 15, 2026 · Argues 4F² requires SOI wafers, could raise costs by approximately 10%, is hard to implement fully without EUV and may be relevant for only one or two generations before 3D DRAM · Big Three working on wafer-to-wafer bonding for DRAM 0a · Operating margin above 50% even after 50% DRAM price fall · Shortage of ASML lithography equipment · Brokerage PDF (no public URL) ·
Transcript of Micron technology webinar (via StockAnalysis) · Oct. 21, 2026 · President and CTO Scott DeBoer: bit density determined by capacitor pitch, not by 6F² versus 4F²; 4F² density benefit overstated, added complexity and cost led Micron to decline this route after evaluation · 3D DRAM as a parallel path between 1ζ and 1η · stockanalysis.com
U.S. Code of Federal Regulations, 15 CFR 772.1 · Definition of “Advanced-node integrated circuits” (DRAM: cell area below 0.0026µm², density above 0.20Gb/mm², or over 3,000 TSVs per die) · Hosted by Cornell LII · law.cornell.edu
U.S. Federal Register, 90 FR 5298 · “Implementation of Additional Due Diligence Measures for Advanced Computing Integrated Circuits; Amendments and Clarifications; and Extension of Comment Period” · Published and effective Jan. 16, 2025 · Revised DRAM thresholds to 0.0026µm², 0.20Gb/mm² and 3,000 TSVs; explains previous definitions did not capture all DRAM equivalent to “18 nm half-pitch” · federalregister.gov
Covington · “U.S. Department of Commerce Strengthens Export Controls on Advanced Computing and Semiconductor Manufacturing Items” · Dec. 2024 · Interim final rule dated Dec. 2, 2024 (FR 2024-28270): replaced DRAM thresholds with cell area under 0.0019µm² and density above 0.288Gb/mm²; HBM category 3A090.c for bandwidth density above 2GB/s/mm² · cov.com
The Register · U.S. Commerce Department adds 36 Chinese entities including YMTC to the Entity List · Dec. 16, 2022 · theregister.com
Evertiq · “CXMT sues Pentagon over Chinese military company designation” · Aug. 31, 2026 · First placed on the Section 1260H list in January 2025, retained after review in June · evertiq.com
Fortune (Bloomberg) · “Apple seeks U.S. approval to buy chips from blacklisted CXMT: FT” · June 27, 2026 · CXMT is on the Section 1260H list; Apple sought assurances it would not be placed on the Entity List · fortune.com
U.S. House Select Committee on the Chinese Communist Party · “Moolenaar, Whitesides to Secretary Lutnick: Hold Firm on Chinese Memory Chips Ban” · Letter dated July 16, 2026 · Urged review of CXMT's addition to Entity List and ban on purchasing memory from YMTC, CXMT and other listed entities in AI, data centers, federal IT and critical infrastructure · chinaselectcommittee.house.gov
MacRumors · “Senators Demand Apple Commit to Not Buying Chinese Memory Chips” · July 30, 2026 · Seven senators—Schumer, Banks, Crapo, Shaheen, Andy Kim, Risch and Ricketts—requested commitment not to use CXMT or YMTC chips in any Apple product worldwide and a reply by Aug. 21 · Apple had explained to lawmakers it would use such chips in devices sold in China · macrumors.com
U.S. Federal Register · “One Year Suspension of Expansion of End-User Controls for Affiliates of Certain Listed Entities” · Nov. 12, 2025 · Affiliates Rule suspended until Nov. 9, 2026 · federalregister.gov
Miller Canfield · “BIS Affiliates Rule Suspension Extended but Uncertainty Remains” · Oct. 1, 2026 · Treasury Secretary Bessent announced extension of the truce to Jan. 10, 2027 on Sept. 23; extension of the affiliates-rule suspension expected but not yet published in the Federal Register; without action the rule would resume Nov. 10 · millercanfield.com
Pillsbury · “What If the BIS Affiliates Rule Comes Into Force?” · Oct. 22, 2026 · No Federal Register notice had yet changed the Nov. 10 resumption date · pillsburylaw.com
W.Media · “SK hynix invests US$ 38.1 billion in two new AI memory fabs” · Aug. 7, 2026 · Yongin Y2 (HBM and other DRAM) first cleanroom June 2029; Cheongju M17 first cleanroom December 2028; “cleanroom expansions and equipment installations will happen as demand dictates” · w.media
TrendForce News · “Micron Sees Tighter Memory Supply Through FY28: When Will the Big Three's New Capacity Arrive?” · Oct. 6, 2026 · Micron's Sept. 30 earnings call suggests tighter supply through FY27–28 · First cleanroom at Yongin Y1 Feb. 2027, Cheongju M17 Dec. 2028, Yongin Y2 June 2029 · Samsung Pyeongtaek P5 operating target 2028, Yongin Fab 1 target 2029 · Micron Idaho ID1 and Taiwan Tongluo mid-2027; ID2, Japan and Singapore NAND in H2 2028 · trendforce.com
BigGo Finance (reporting on Morgan Stanley research) · “Morgan Stanley: China's Memory Duo Could Reshape Global Supply-Demand by 2028; ChangXin Capacity to Surpass Micron” · Aug. 29, 2026 · CXMT monthly capacity 180,000 wafers (2025), 300,000 (2026), 500,000 (2028) · DRAM shortage rate 17% in 2026 and 15% in 2027 · Potentially weaker pricing discipline in a downturn · finance.biggo.com
Hana Securities, Kim Sung-eun and Kim Kyung-hwan · “China Semiconductor Supply Competition: YMTC IPO vs CXMT Index Inclusion” (China Thematic Strategy) · Oct. 19, 2026 · YMTC application filed Aug. 21, review questions Sept. 4; local expectations for December 2026 IPO (March 2027 if CXMT's timeline applied mechanically) · CXMT's first lockup expiry at end-January 2027: approximately RMB 68.5bn at Oct. 16 closing price RMB 45.02 · STAR 50 announcement end-November, effective Dec. 14; estimated Q4 inflows RMB 17.1bn–20.8bn · Brokerage PDF (no public URL) ·
TrendForce DataTrack · “CXMT Lists on the STAR Market as China's DRAM Localization Enters a Capital-Intensive Expansion Phase” · July 2026 · IPO about RMB 57.919bn (RMB 66.607bn including overallotment), offer price RMB 8.66, float 6.73% · datatrack.trendforce.com
Sina Finance (新浪财经) · “YMTC Aims to Raise RMB 33bn in STAR Market IPO” · Aug. 21, 2026 · IPO target RMB 33bn (production-line upgrades RMB 20.8bn, R&D RMB 12.2bn) · Q1 2026 revenue RMB 47.042bn, net income attributable to parent shareholders RMB 33.379bn · finance.sina.com.cn
Sina Finance · CXMT plan to deploy RMB 18bn of excess IPO proceeds (board action Sept. 28) · Sept. 29, 2026 · Technology R&D project: total RMB 24.1bn (RMB 13bn excess proceeds; equipment 92.94%; 30-month schedule) · Hefei back-end testing facility Phase II: total RMB 10.8bn (RMB 5bn excess proceeds; 37-month schedule) · Not purely investment in wafer-production capacity; approximately RMB 18.8bn of excess proceeds remains · finance.sina.com.cn
CXMT filing No. 2026-015 (reproduced by Securities Star) · “CXMT: Resolution of the Second Extraordinary General Meeting of Shareholders in 2026” · Oct. 19, 2026 · All motions at the October 19 extraordinary meeting passed, including allocation of part of excess IPO proceeds to new projects · stockstar.com
TrendForce press release · HBM pricing power and 2027 contract-price outlook · June 2, 2026 · HBM share of Big Three DRAM wafer input: ~18% end-2025 → 22% end-2026 → 30% end-2027; more displacement of conventional DRAM capacity · HBM share of DRAM bit output rises from ~8% to 13% · Q1 64GB DDR5 RDIMMs overtook HBM in revenue per wafer; HBM profitability also lower since Q1 (estimates based on die size, yield and price per Gb) · Expects large HBM contract-price gains in 2027 · trendforce.com
TrendForce press release · “Diverging Memory Market Outlook in 2027 as DRAM Supply Remains Tight While NAND Flash Supply Conditions Ease” · July 30, 2026 · 2026 DRAM supply-fulfillment rate approximately −1% to −2%, deficit widening in 2027 · HBM consumes more wafers, limiting bit-output gains from input increases · 2027 server shipment growth exceeds 2026 (+17%) · Smartphone and laptop shipments decline again in 2027 as rising component costs pass through to buyers · trendforce.com
Samsung Electronics Q2 2026 earnings-call transcript (via StockAnalysis) · July 30, 2026 · EVP Kim Jae-june: multi-year supply agreements start at five years, with one-year extension negotiated annually; if finalized, contracts cover 60–70% of planned DRAM and NAND capacity · Signed with five leading data-center customers, five more in advanced talks · Substantial prepayments, one-quarter already received (amount undisclosed under NDAs) · Minimum price floors for mainstream products at levels meant to reduce investment risk from price volatility · 2027 supply constraints expected worse than 2026, meaningful added supply difficult through 2028 · stockanalysis.com
ASML Q3 results release (U.S. SEC Form 6-K) · Oct. 14, 2026 · CEO Fouquet: “our supply as currently planned for both 2027 and 2028 will not fully meet the demand”; evaluating 2028 expansion beyond existing 30% plan · sec.gov
DS Investment & Securities, Lee Su-rim · “Mind Over Matter” (Semiconductor Sector Comment) · Oct. 19, 2026 · Table 2 operating profit by ASP scenario: SK hynix 2027E KRW 406.8tn; base (DRAM −20/−10/−5%) KRW 347.6tn, 335.6tn, 357.8tn; bear (−35/−20/−10%) KRW 259.5tn, 210tn, 203.3tn; bull KRW 406.4tn, 424tn, 477tn · HBM affected by customer roadmaps and contracts · Brokerage PDF (no public URL) ·
SK Securities, Han Dong-hee · “This Time Is Different” (Semiconductor) · Oct. 21, 2026 · Samsung and SK hynix 12-month forward P/E: 3.5x and 2.8x, below historic downcycle troughs, implying deep post-2028 earnings cuts · Costs of breaching LTAs · SK hynix FCF: 2026E KRW 170.655tn, 2027E KRW 268.57tn, 2028E KRW 356.96tn · Brokerage PDF (no public URL) ·
SK hynix · “Decision on Cancellation of Shares” · DART · Aug. 19, 2026 · Total issued shares before cancellation: 730,492,365 · dart.fss.or.kr
SK hynix Q2 2026 results release (reproduced by Placera) and earnings-call coverage (Digital Today) · July 29, 2026 · LTAs signed with about 10 customers, including key accounts; typically around five years but varying by customer and product · Different pricing formulas to manage volatility · Deposits and other financial safeguards plus long-term purchase commitments · LTA share of revenue undisclosed · Discussions with major customers about 2027 HBM volumes and pricing ongoing (President Song Hyun-jong) · placera.se · digitaltoday.co.kr
TrendForce press release · “Long-Term Agreements Cap Price Increases; Server DRAM Contract Prices Expected to Rise 13-18% QoQ in 3Q26” · July 9, 2026 · Multi-year LTAs at U.S. cloud companies limit price increases to those customers; server DRAM price gains from Q3 led by customers without LTAs · trendforce.com
Hana Securities, Lee Kyung-soo · “Weekly Domestic and Global Earnings Trends” (Quant Comment) · Oct. 19, 2026 · Weekly SK hynix EPS revisions +0.57%, Samsung +0.78% · ASML formally plans 30% expansion in 2027 and considers an additional 30% in 2028 · Brokerage PDF (no public URL) ·