Report Interpretation

Overview

This initiation report examines whether CXMT can turn China's localization drive into durable DRAM share gains despite lagging process technology and export restrictions. Bernstein concludes that strong current pricing, aggressive capacity additions, sufficient conventional-DRAM technology and policy-supported HBM demand justify an Outperform rating, while acknowledging a weaker long-term technology runway without EUV.

Core views

Bernstein's positive case begins with the combination of an elevated DRAM cycle, China's localization demand and CXMT's relative valuation. The report expects CXMT to grow faster than the industry through 2027 and to experience a milder downturn than global peers when conventional DRAM pricing normalizes in 2028. From 2026 to 2028, the model implies roughly 75% capacity growth, 55% revenue growth and 50% EPS growth. Bernstein initiates coverage at Outperform with a CNY70 one-year target. Capacity expansion is the first major driver. CXMT currently has about 300K wafer starts per month, while existing Hefei and Beijing fabs appear largely full. Bernstein assumes additional capacity from Shanghai, Beijing and Hefei, including a reported Shanghai fab that is not identified in company filings but that the firm's checks suggest is being equipped and may be linked to HBM. Capacity is modeled at 380K wafers per month in 4Q27 and 530K in 4Q28, about 75% above the current level. That would represent approximately 19%-20% of global DRAM wafer capacity, make CXMT the third-largest producer by wafer capacity and put it ahead of Micron on that measure. Because CXMT produces fewer bits per wafer than leading suppliers, Bernstein explicitly discounts wafer share when estimating its actual market presence. The expansion requires exceptionally high investment. Bernstein forecasts capex of approximately CNY81 billion, CNY142 billion and CNY236-238 billion in 2026, 2027 and 2028, equivalent to about US$12 billion, US$21 billion and US$35 billion. The 2026 figure is more than 70% above the prior year, with spending then expected to rise by about 70% annually. Unlike earlier expansion phases that relied heavily on central and local government funding, the report argues that elevated memory prices and stronger operating cash flow should allow CXMT to finance much of the new investment internally. Physical constraints such as obtaining land, constructing fabs, recruiting workers and qualifying equipment may be more limiting than cash. On technology, teardown-based bit-density comparisons place CXMT's G4 process roughly alongside competitors' 1z-1a nodes and about 3-4 years behind global leaders at 1c or 1-gamma. Bernstein expects migration to G5 from the second half of 2027, maintaining rather than closing that gap. The report nevertheless argues that the technology is good enough for most conventional DRAM. DDR5 and LPDDR5 supplied 40% of DRAM revenue in 1H25 and are said to represent more than half currently; servers contributed 27%, or roughly 25%-30%, of 2025 revenue, close to the global server mix of 25%-30%. Xiaomi's planned use of CXMT LPDDR6 is cited as evidence that the company can enter newer product generations. Customer checks indicate that CXMT's pricing is broadly similar to competitors rather than dependent on deep discounts. Profitability also supports the claim that CXMT is commercially competitive despite its process lag. Gross margin reached 88% in 2Q26, close to the 87%-89% range reported for Samsung, SK hynix and Micron. The report attributes this partly to the strong cycle and absence of lower-margin HBM in the current mix, while also considering possible construction, operating and policy-supported pricing advantages. EBITDA margins were already approaching incumbent levels in 2024 and were roughly aligned in 1H25, which Bernstein interprets as evidence of comparable cash costs and an ability to weather downturns. Gross margin is modeled at 88.1% in 2026, 90.5% in 2027 and 85.1% in 2028; operating margin is forecast at 82.5%, 85.8% and 80.5%, respectively. Export controls are the central constraint and a key assumption behind the forecasts. Restrictions announced in January 2025 limit access to equipment capable of producing DRAM below specified cell-area or above specified bit-density thresholds, equipment supporting more than 3,000 TSVs per die and advanced DUV or EUV systems. Yet CXMT continued to expand and improve for approximately 1.5 years after the restrictions, including before the current price cycle began. Bernstein therefore assumes that previously stockpiled equipment and domestic alternatives will permit expansion through 2028. The report believes DUV can sustain competitiveness for another 3-5 years, potentially supporting two or three more process generations, although with worsening yields and economics. Non-lithography equipment is considered less restrictive because Chinese alternatives are generally more capable in those areas. The longer-term technology outlook is less favorable. CXMT is researching 4F² cell architecture and 3D DRAM as ways to reduce dependence on EUV. A 4F² structure can shrink the cell without better lithography but is characterized as a one-time improvement, while 3D DRAM could move density gains into the vertical dimension. Bernstein does not assume either approach becomes commercially competitive within its forecast horizon. Because global rivals have investigated these alternatives but still chosen EUV, the report expects CXMT to become increasingly less competitive beyond the next 3-5 years unless these technologies succeed or equipment access improves. HBM presents an even greater technical hurdle. CXMT must deal with faster DRAM dies, limited experience in TSV and bonding, lower expected yields, higher costs and increasingly dense interconnect requirements. Existing restrictions cover equipment above 3,000 TSVs per die, while the report cites more than 5,600 TSVs in HBM3 and more than 20,000 in HBM4. Advanced logic base dies are another bottleneck. Bernstein nevertheless argues that controls on GPUs and HBM have divided the market: Chinese AI customers cannot obtain sufficient imported HBM and will therefore need domestic supply. Localization policy should allow CXMT to charge enough to share its higher costs with customers and retain satisfactory margins. The model moves CXMT from HBM3 or an equivalent product in 2026 toward HBM3E in 2027, still about 3-4 years behind leaders, with HBM reaching 9% of DRAM revenue in 2028. Bernstein also challenges the view that CXMT's addressable market is confined to domestic consumption. Chinese OEMs can install its memory in PCs and smartphones sold abroad, and checks suggest that some non-Chinese consumer OEMs and module makers will accept Chinese memory in products sold globally outside the United States. Geopolitically sensitive customers such as foreign AI laboratories, hyperscalers and government agencies are less accessible. Chinese OEMs are estimated to represent 20%-25% of global DRAM demand, and CXMT currently satisfies only about 40%-45% of that requirement. The report calculates China's DRAM self-sufficiency at 42% currently and 58% in 2028. Including non-Chinese consumer products sold in China and other non-U.S. markets expands CXMT's addressable design TAM to 42% of global DRAM demand in 2026 and 36% in 2030; CXMT's current penetration of that expanded TAM is estimated at only 21%. After adjusting for lower bit density and HBM limitations, Bernstein models CXMT's global bit share rising from about 9% in 2026 to 12% in 2028, with conventional-DRAM share increasing from around 10% to 14%. One section heading gives a wider 8%-13% bit-share progression. Because conventional DRAM should remain the majority of industry revenue and CXMT's blended pricing is expected to resemble peers, revenue share is forecast to rise from 10% to 13%. HBM is expected to remain around 20% of the overall DRAM market in the next several years, while inference and agentic workloads support continued demand for conventional memory. The financial model combines this share expansion with a cyclical price path. DRAM prices are expected to remain elevated or rise modestly through most of 2027 before declining steadily in 2028. CXMT revenue is forecast at CNY452.9 billion, CNY749.4 billion and CNY702.4 billion in 2026-2028, representing growth of 633%, 65.5% and negative 6.3%. EPS is modeled at CNY4.04, CNY6.82 and CNY6.00, with 2027 growth near 69%-70% and a 12% decline in 2028. Bernstein is 20%-30% above consensus on revenue and 40%-50% above on EPS for 2026 and 2027, but slightly below consensus on 2028 revenue and broadly in line on EPS because it assumes a more pronounced price normalization. Two financial-structure issues temper the earnings outlook. First, the model assumes no dividends or buybacks through 2028 because continued expansion qualifies as major capital expenditure and CXMT is expected to retain cash. Second, minority interests create substantial EPS uncertainty. CXMT owns only about 30% of certain consolidated fabs that account for roughly two-thirds of wafer capacity. Minority interests absorbed 63% of consolidated profit in 2H25 but only 25%-29% in 1H26, with no clear disclosure explaining the change. Bernstein assumes a 25% minority share going forward, while identifying the ownership structure of future fabs as a major sensitivity. For valuation, CXMT trades at roughly 13x one-year forward P/E and 7.3x one-year forward P/B in the report's discussion, making it more expensive than overseas memory producers. Bernstein expects that premium to persist because of A-share technology valuations and localization demand. Against domestic semiconductor groups spanning foundries, memory, equipment and AI processors, however, CXMT appears materially cheaper on both one- and two-year forward measures; the report says it would remain inexpensive even if two-year forward EPS were half of consensus. The CNY70 target applies 10x to CNY7.16 of blended Q5-Q8 EPS and corresponds to 4.4x two-year forward P/B and 6.2x forward EV/EBITDA. It represented 26% upside from the CNY55.54 close on 18 September and 31% using the 17 September reference cited in the valuation section. The report cautions that 2.2% of total shares unlock on 27 January 2027, increasing current free float by 34%, followed by a much larger 33% of outstanding shares on 27 July 2027.

Analysis framework

Bernstein starts with a bottom-up fab and capex build, translates wafer capacity into bit output after adjusting for process-density and yield disadvantages, and then estimates market share, addressable demand and China's self-sufficiency. It combines shipment growth with an industry DRAM price cycle, product mix, cost and minority-interest assumptions to forecast revenue, margins and EPS. The valuation then compares CXMT with overseas memory companies and domestic semiconductor peer groups before applying a forward P/E multiple to blended Q5-Q8 earnings.

Methodology notes

• Industry AnalysisSupply-demand framework

  DRAM supply-demand and pricing cycle

  The report links industry shortages and capacity additions to elevated pricing through 2027 and a supply-demand normalization with lower prices in 2028.

• Industry AnalysisVolume-price decomposition

  Bit shipments, blended ASP and cost per bit

  Revenue and margins are built from shipment growth, conventional DRAM and HBM pricing, product mix and cost-per-bit assumptions rather than from a single top-down growth rate.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  Localization-policy transmission through equipment, memory production and OEM demand

  The analysis follows how equipment access and domestic alternatives affect CXMT's production, and how localization requirements and OEM adoption translate that production into demand and pricing.

• Corporate Fundamentals and FinanceOperating and Financial Leverage Analysis

  Operating leverage across the DRAM cycle

  Bernstein expects earnings to rise faster than revenue in the upcycle and decline more sharply in 2028 as pricing, fixed costs and utilization move through the cycle.

• Valuation methodsP/E and PEG Valuation

  Forward Q5-Q8 P/E valuation

  The CNY70 target applies 10x to CNY7.16 of blended earnings from forward quarters five through eight, a period intended to look beyond the immediate peak.

• Valuation methodsPB valuation

  Forward price-to-book cross-check

  The target is cross-checked at 4.4x two-year forward P/B and compared with global memory and domestic semiconductor peer valuations.

• Valuation methodsEV/EBITDA valuation

  Forward EV/EBITDA cross-check

  Bernstein notes that the target corresponds to 6.2x forward EV/EBITDA, above global memory incumbents but below A-share semiconductor peers.

• Other

  Bit-density normalization and design-TAM self-sufficiency analysis

  The report adjusts wafer capacity for lower bits per wafer, then compares CXMT's bit shipments with Chinese OEM design demand and a broader non-U.S. consumer TAM to estimate market share and self-sufficiency.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• CXMT Corp (688825.CH)

  Primary covered company and expected beneficiary of China's DRAM localization, rising capacity and policy-supported HBM demand.

  Strengths

  China's largest DRAM producer; rapid capacity growth; conventional DRAM suitable for server and newer mobile applications; gross and cash margins approaching incumbent levels; substantial domestic and non-U.S. consumer TAM.

  Weaknesses

  Process technology is approximately 3-4 years behind global leaders; HBM yields and costs are expected to be challenging; long-term scaling is constrained by lack of EUV.

  Comparison

  Pricier than overseas memory producers but materially cheaper than the report's A-share semiconductor peer groups on one- and two-year forward measures.

  Risks

  Equipment restrictions, tighter export controls, DRAM oversupply, weaker AI-related demand, minority-interest uncertainty and large 2027 share unlocks.

Key data

• Rating and targetOutperform; CNY70Coverage initiation; target based on 10x forward Q5-Q8 EPS
• Reference priceCNY55.54Closing price on 18 Sep 2026; target implies 26% upside
• Wafer capacity300K currently; 380K in 4Q27; 530K wpm in 4Q28Approximately 75% expansion through 4Q28
• Capex forecastCNY81.3B / CNY141.5B / approximately CNY236-238B2026E / 2027E / 2028E
• Technology gapApproximately 3-4 yearsG4 is compared with rivals' 1z-1a nodes; G5 migration expected from 2H27
• 2Q26 gross margin88%Comparable with the 87%-89% range for Samsung, SK hynix and Micron
• HBM revenue mix9% in 2028No meaningful revenue currently; HBM3E transition modeled for 2027
• Global revenue share10% in 2026 to 13% in 2028Modeled broadly in line with bit share
• China DRAM self-sufficiency42% currently to 58% in 2028Based on Chinese OEM design demand and CXMT bit shipments
• Revenue forecastCNY452.9B / CNY749.4B / CNY702.4B2026E / 2027E / 2028E; growth of 633%, 65.5% and -6.3%
• EPS forecastCNY4.04 / CNY6.82 / CNY6.002026E / 2027E / 2028E
• Consensus differenceRevenue 20%-30% above and EPS 40%-50% aboveBernstein versus consensus for 2026-2027
• Share unlocks2.2% on 27 Jan 2027; 33% on 27 Jul 2027The January amount raises the current free float by 34%; July is substantially larger

Impact & implications

Bernstein expects CXMT's expansion to lift it into the top three global DRAM producers by wafer capacity and materially increase China's memory self-sufficiency. This would intensify competitive pressure on incumbent suppliers while giving CXMT structural domestic demand and some protection from a future downcycle. The same strategy requires very high capex, depends on continued equipment access or domestic substitutes, and does not resolve the longer-term EUV technology constraint.

Risks

• Greater difficulty obtaining semiconductor equipment could delay capacity expansion or the G4-to-G5 transition.
• Further tightening of U.S. export restrictions could reduce access to advanced DUV, EUV, TSV and other production equipment.
• Faster-than-expected global DRAM supply growth could turn the current shortage into oversupply and drive prices lower.
• A significant slowdown in AI-related DRAM demand could weaken HBM and conventional-memory demand.
• HBM's low expected yields, high cost and advanced packaging requirements could delay progress or depress profitability.
• The unexplained change in minority interests and the ownership structure of future fabs create substantial uncertainty for parent-attributable EPS.
• Share unlocks in January and especially July 2027 could materially increase tradable supply and affect the stock.

What to watch

• Whether CXMT reaches 380K wafer starts per month in 4Q27 and 530K in 4Q28, including the assumed Shanghai, Beijing and Hefei ramps.
• Progress toward G5 production from 2H27 and whether domestic equipment alternatives can sustain yields.
• The planned move from HBM3 or an equivalent product to HBM3E in 2027 and HBM's path toward 9% of 2028 revenue.
• Whether DRAM prices remain elevated through 2027 and then normalize at the pace assumed for 2028.
• Any tightening or enforcement change in export controls affecting advanced lithography, TSV or other equipment.
• Minority interests' share of consolidated profit and the ownership structure chosen for new fabs.
• Market impact from the 27 January and 27 July 2027 share unlocks.