Global Memory: Memory shortage remains elevated through 2027, but China and CY2028 normalization reshape the longer-term outlook

Overview

This global-memory update assesses the durability of the current DRAM, NAND and HBM upcycle, revises company forecasts and valuations, and examines China’s growing competitive impact. Bernstein expects elevated pricing through CY2027, a more resilient but normalizing CY2028, and divergent company outcomes driven by HBM execution, capital returns and Chinese supplier progress.

Core views

Bernstein continues to view the memory cycle as elevated. It forecasts conventional DRAM and NAND ASPs to rise by high-teens to about 20% quarter-on-quarter in 3QCY26, following a weaker-than-expected 2QCY26 for DRAM. Conventional DRAM pricing has already quadrupled since 3QCY25. A material supply shortage is expected to persist through CY2027, but the report expects the rate of price increases to slow to mid- to high-single-digit percentages quarter-on-quarter in 4QCY26 and remain mild during CY2027. The constraint is long-term agreements: their price ceilings limit further escalation, while PC and mobile customers may reduce unit shipments and memory purchases as prices rise. For CY2028, Bernstein still expects the cycle to normalize, but now models a more gradual decline because current supply-demand visibility and LTAs should make pricing more resilient than in its previous revision. More capacity additions should eventually alleviate shortages. The report also argues that AI infrastructure spending may encounter financing, safety, environmental and job-displacement constraints, while very large industry revenue and profit pools could attract customer backlash or government intervention. Consensus expects memory suppliers to generate roughly US$1.5-1.6 trillion of annual revenue and US$1.2-1.3 trillion of operating profit over the next two years. Bernstein therefore expects supplier gross margins to ease from roughly 90% at the peak to high-70% levels for DRAM and mid-60% levels for NAND exiting CY2028, while remaining above prior-cycle peaks. The HBM outlook is more differentiated. Bernstein cuts the assumed size of the CY2027 HBM price increase for SK hynix after July and August Korean export trackers showed weak signals from its HBM production location, while Samsung’s signals were stronger. Reported HBM4 supply difficulties lead Bernstein to delay the HBM3E-to-HBM4 transition and retain meaningful HBM3E shipments in CY2027. It shifts some HBM share assumptions from SK hynix toward Samsung and expects HBM to be less profitable than conventional DRAM in CY2027. The model also assumes half of Rubin Ultra chips are de-specified from 12-high to 8-high HBM, reducing blended HBM content from 1,024GB to 640GB. This should free capacity for conventional server DRAM; Bernstein expects only a marginal effect on supplier financials because aggregate memory demand still exceeds supply. The report sees capital returns and enforceable LTAs, rather than large additional earnings upgrades, as the next potential support for memory equities. It believes consensus has largely caught up after about a year of the upcycle, making major positive earnings revisions less likely. However, suppliers’ free-cash-flow generation and balance sheets could support cash returns, while LTAs can improve confidence in earnings durability. Most suppliers aim eventually to place 50% or more of revenue or capacity under LTAs. Bernstein emphasizes financial guarantees as especially important because they determine whether contractual commitments are enforceable. It expects clearer capital-return policies and cash distributions to support gradual stock appreciation. Bernstein retains Outperform ratings on Samsung, SK hynix and Micron, prefers Samsung among incumbent suppliers, and retains Underperform on KIOXIA. Samsung is preferred because the report expects better HBM4 progress and capacity to support HBM share gains; it also expects Samsung Foundry to recover from a trough and begin generating operating profit. Samsung’s target remains KRW440,000, based on 6.5x forward 5Q-8Q EPS of KRW67,692. Micron’s target remains US$1,300, based on 7.6x forward 5Q-8Q EPS of US$171.1; its estimates are broadly raised because of relatively greater conventional-DRAM exposure. SK hynix’s target is cut to KRW2,700,000 from KRW3,300,000 and its valuation multiple reduced to 5.1x from 6.2x forward 5Q-8Q EPS, reflecting slower HBM4 progress and a more conservative HBM price outlook. The report nevertheless states that the revised target represented 52% upside as of September 28. KIOXIA remains the central negative stock-specific view. Bernstein raises its target to JPY49,000 from JPY40,000 on a better NAND-pricing outlook, but keeps Underperform because its SOTP analysis still indicates 10% downside. The report identifies YMTC as a greater-than-expected long-term threat: as of 1QCY26, YMTC had overtaken Micron, KIOXIA and SanDisk in NAND revenue and was only slightly behind SK hynix. YMTC’s gross margin had exceeded the peer average since CY2024 and was likely the highest among the group in 1QCY26, which Bernstein views as evidence of competitiveness in technology, pricing and cost. KIOXIA’s JPY800 billion buyback could nevertheless be a near-term positive catalyst, although the report notes that it would rely on expected FQ2 free cash flow given JPY791 billion of FQ1 cash and only a limited net-cash position. Bernstein is constructive on CXMT, which it initiated with an Outperform rating and a CNY70 target. It expects capacity to reach 530,000 wafers per month by the end of CY2028 and surpass Micron’s capacity, equating to nearly 20% of global DRAM wafer-start share. Despite export controls, the report sees continuing technology progress: CXMT’s G4 remains roughly three to four years behind global leaders in bit density, but G5 was announced for mass production on September 20, 2026 with an active-area half pitch similar to the latest generations elsewhere. HBM remains a challenge, but localization of China’s AI compute infrastructure is expected to provide demand and help absorb yield and cost difficulties. Bernstein forecasts CXMT’s global bit/revenue share to rise from 9%/10% in CY2026 to 12%/13% in CY2028, implying 42%-58% DRAM self-sufficiency for Chinese OEM needs. Its CY2026-28 EPS forecasts are CNY4.0, CNY6.8 and CNY6.0, respectively; CY2026-27 estimates are 20%-50% above consensus. The report values CXMT at 10x blended two-year-forward EPS of CNY7.16 and flags sizeable lock-up expirations in January and July 2027 as a risk. SanDisk remains Outperform with an unchanged US$3,000 target. Bernstein values it at 11x FY28 EPS or 14x through-cycle FY2026-30 average EPS. Across the industry, the report’s key downside risks are an earlier end to favorable pricing from weaker demand or greater supply, weaker investor sentiment and valuation, and faster Chinese memory progress, particularly in NAND.

Analysis framework

Bernstein combines memory supply-demand and ASP forecasts with HBM product-transition assumptions, export-tracker observations, company financial forecasts, consensus comparisons and valuation analysis. It uses forward P/E for Samsung, SK hynix, Micron and CXMT, SOTP for KIOXIA, and evaluates LTAs, financial guarantees, free cash flow, capital returns, capacity expansion and supplier market-share changes.

Methodology notes

• Industry AnalysisSupply-demand framework

  Memory supply-demand and price-cycle analysis

  The report forecasts DRAM and NAND prices from shortages, capacity additions, customer demand and LTA price ceilings, then uses those assumptions to model the CY2028 normalization.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  HBM capacity allocation and AI-server demand transmission

  It assesses how HBM4 delays and lower HBM content in Rubin Ultra could release capacity for conventional server DRAM and affect supplier mix, revenue, profit and market share.

• Valuation methodsP/E and PEG Valuation

  Forward P/E valuation

  Bernstein applies forward earnings multiples to Samsung, SK hynix, Micron and CXMT to derive price targets.

• Valuation methodsSOTP (Sum-of-the-Parts) Valuation

  KIOXIA sum-of-the-parts valuation

  KIOXIA’s value is divided into non-LTA business, LTA-covered business, LTA guarantee payments and excess cash to estimate its implied downside.

Asset mapping & comparison

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

• Samsung Electronics (005930.KS)

  Preferred incumbent memory supplier; expected to gain HBM share.

  Strengths

  Better expected HBM4 progress, more capacity, memory-market leadership and a prospective Foundry recovery.

  Weaknesses

  Employee-bonus-related cost increases caused a slight EPS reduction despite stronger conventional-memory revenue expectations.

  Comparison

  Bernstein considers Samsung’s market capitalization being similar to SK hynix’s unreasonable given Samsung’s memory leadership and non-memory businesses.

  Risks

  Earlier memory-price normalization, weaker demand, higher supply and China’s memory progress.

• SK hynix (000660.KS)

  Covered HBM supplier with an Outperform rating.

  Strengths

  Remains positioned in a tight memory market.

  Weaknesses

  Slower HBM4 progress, weaker export-tracker signals and reduced CY2026-27 HBM pricing assumptions.

  Comparison

  Bernstein shifts some projected HBM share from SK hynix to Samsung.

  Risks

  HBM4 delays and a smaller-than-expected HBM price increase.

• Micron (MU)

  Covered supplier rated Outperform.

  Strengths

  Relatively greater conventional-DRAM exposure supports broadly raised estimates.

  Comparison

  YMTC had overtaken Micron in NAND revenue as of 1QCY26, according to Bernstein.

  Risks

  Earlier end to favorable memory pricing, weaker demand, supply expansion and Chinese competition.

• KIOXIA Holdings (285A.JP)

  Covered NAND supplier rated Underperform.

  Strengths

  A JPY800 billion buyback may provide a near-term catalyst and downside protection.

  Weaknesses

  Bernstein’s SOTP still indicates 10% downside and sees a worsening long-term competitive position.

  Comparison

  YMTC overtook KIOXIA in NAND revenue by 1QCY26 and had likely the highest gross margin among peers.

  Risks

  YMTC capacity and competitiveness could add NAND supply and pressure pricing; the buyback depends on expected FQ2 free cash flow.

• CXMT Corp (688825.CH)

  Covered Chinese DRAM supplier rated Outperform.

  Strengths

  Rapid capacity expansion, localization demand, improving technology, margins near global peers and CY2026-27 EPS forecasts above consensus.

  Weaknesses

  HBM technology remains materially behind global leaders, while yields and costs remain uncertain.

  Comparison

  Bernstein forecasts capacity to surpass Micron’s by CY2028 and values CXMT as cheaper than other A-share semiconductor stocks.

  Risks

  Export-control uncertainty and sizeable lock-up expirations in January and July 2027.

• SanDisk (SNDK)

  Covered NAND supplier rated Outperform.

  Strengths

  Bernstein maintains its US$3,000 target and values the shares on FY28 and through-cycle earnings.

  Weaknesses

  Near-term numbers appear high, and company disclosures and investor communication are described as confusing.

  Comparison

  YMTC had surpassed SanDisk in NAND revenue as of 1QCY26.

  Risks

  Cyclical NAND downside and the possibility that structural NAND weakness lowers DCF value and asset values below replacement cost.

Key data

• 3QCY26 conventional DRAM and NAND ASP growthHigh-teens to 20% QoQBernstein forecast based on recent contract-price reports.
• CY2028 supplier marginsDRAM high 70s%; NAND mid-60s%Expected exiting CY2028 after normalization from roughly 90% peak gross margins.
• SK hynix target priceKRW2,700,000Cut from KRW3,300,000 on slower HBM4 progress and more conservative HBM pricing.
• KIOXIA SOTP implied downside10%Bernstein’s SOTP value is JPY49,191 per share versus a JPY54,470 current price as of September 25.
• CXMT global bit/revenue share9%/10% in CY2026 to 12%/13% in CY2028Supported by rapid capacity expansion and China localization demand.
• CXMT CY2026-28 EPS forecastCNY4.0 / CNY6.8 / CNY6.0CY2026-27 forecasts are 20%-50% above consensus.

Impact & implications

Bernstein’s view supports a still-favorable industry backdrop through CY2027 but shifts the emphasis from earnings surprises to capital returns, LTA enforcement and relative execution. Samsung is the preferred incumbent because of expected HBM4 progress and capacity, while the report sees China’s YMTC and CXMT as increasingly consequential long-term forces, particularly for NAND and KIOXIA.

Risks

• A weaker demand environment or faster supply growth could end favorable memory pricing earlier than Bernstein expects.
• Investor sentiment and the valuation multiple assigned to memory suppliers may weaken.
• China’s progress in memory, especially NAND, could pressure incumbent suppliers and prices.
• CXMT faces uncertainty around access to semiconductor equipment for capacity expansion, potential global DRAM oversupply and a slowdown in AI-related DRAM demand.
• CXMT lock-up expirations in January and July 2027 could affect the market.
• SanDisk faces cyclical NAND downside, confusing investor communication and the risk of structural NAND weakness.

What to watch

• HBM4 supply progress, export-tracker signals and whether Samsung gains HBM share from SK hynix.
• The pace of conventional DRAM and NAND price increases, LTA price ceilings and the timing of CY2028 normalization.
• Supplier capital-return announcements and whether LTAs are backed by enforceable financial guarantees.
• YMTC revenue, margin and technology progress in NAND.
• CXMT capacity ramp, G5 execution, export-control effects and China AI-infrastructure localization demand.
• KIOXIA’s capital-return actions and funding of its buyback.