Goldman Sachs: Memory Super-Cycle Begins, Raises Target Prices for Samsung/Hynix, Upgrades Kioxia

Overview

The core conclusion of this report is that the memory chip industry is undergoing a 'peak-to-stay-higher' super-cycle distinct from past cycles, with sustainability extending at least through 2028. Based on three pillars—structural changes in AI-driven demand, slowed supply growth due to HBM cannibalization, and earnings stability from long-term agreements—Goldman Sachs substantially raised earnings forecasts and target prices for Samsung Electronics and SK Hynix, and initiated Kioxia Holdings with a Buy rating. The report emphasizes that the market currently values memory stocks at single-digit P/Es, failing to fully reflect the sustainable high profitability of this cycle and the potential for re-rating.

Core views

Demand Side: AI servers have become the primary engine for memory demand, with significance far exceeding the cloud data center expansion period of 2017-2018. Currently, servers account for approximately 50% of DRAM demand and 40% of NAND demand, compared to only 16% in 2017. With the expansion of Agentic AI, the dependence of large models on KV Cache and high-bandwidth memory turns memory from an auxiliary component into a performance bottleneck. Even with widespread HBM adoption, bandwidth and capacity remain insufficient to meet accelerator needs, providing a solid foundation for long-term high visibility of memory demand. Supply Side: Expansion of traditional memory capacity is significantly slower than in the previous cycle. The CAGR for traditional DRAM capacity of the top three vendors was 12% during 2017-2018, but is projected to be only 7-8% from 2026-2030. This is due to: first, new cleanroom construction cycles take nearly three years, preventing rapid scaling in the short term; second, HBM consumes 3-4 times more wafers per unit than traditional DRAM, and yield ramping is difficult. Consequently, most new capacity is prioritized for HBM, further squeezing traditional memory supply. Traditional DRAM supply CAGR is expected to be only 15% during 2026-2028, down 4 percentage points from the 19% seen in 2017-2018. Contract Mechanisms: Long-term agreements are altering industry bargaining dynamics. Customers actively seek LTAs to secure mid-to-long-term supply, while suppliers use them to enhance business stability. Next-generation LTAs include substantial upfront payments, penalty clauses for breach, and guaranteed price floors, creating binding strength far stronger than historical versions. This mechanism not only enhances earnings predictability but also helps manufacturers plan capital expenditures more efficiently by locking in sales volumes, thereby reducing the risk of severe overcapacity in the future. Cases in the silicon wafer industry indicate that similar structures can maintain higher gross margin stability. Price and Earnings Forecasts: Based on these supply-demand assessments, Goldman Sachs substantially upgrades memory ASP forecasts. Samsung's 2026E traditional DRAM ASP is projected to rise 326% YoY, with NAND ASP rising 283%. In 2027E, they are expected to rise an additional 27% and 33%, respectively. HBM ASP is forecast to achieve catch-up growth in 2027, rising 44% YoY, propelling total HBM TAM from $56 billion in 2026 to $116 billion in 2027, and further to $168 billion in 2028. Correspondingly, Samsung's operating profit forecasts for 2026E-2028E are raised to KRW 3.74 trillion/5.30 trillion/6.10 trillion, and SK Hynix's to KRW 2.71 trillion/4.01 trillion/4.54 trillion. Valuation Methodology Shift: Given enhanced earnings sustainability, Goldman Sachs switches SK Hynix's valuation anchor from P/B to P/E, applying a target P/E of 9x (based on average EPS for 2026-2027), yielding a target price of 3.5 million KRW. This multiple aligns with current peer averages and approaches Micron's average valuation during its upcycle. For Samsung, a SOTP valuation method is used, implying approx. 9x average P/E for 2026-2027, resulting in a target price of 480,000 KRW. The report notes that if memory companies can demonstrate their high profitability is structural rather than purely cyclical, valuations could converge towards those of high-margin growth companies with moats (which typically trade around 20x P/E).

Analysis framework

Goldman Sachs' analytical framework revolves around 'why this cycle is different,' employing historical comparison against the strongest upcycle of 2017-2018 across three dimensions: demand structure, supply elasticity, and contract enforceability. On the demand side, it analyzes server memory share trends, AI cluster Capex trajectories, and token consumption growth rates to argue that memory has evolved from a generic accessory to a compute bottleneck in the AI era. On the supply side, it quantifies physical capacity expansion limits using hard constraints like available cleanroom area, HBM wafer consumption coefficients, and factory lead times. At the contract level, it introduces silicon wafer industry LTAs as a reference to validate the effectiveness of long-term agreements in smoothing cyclical volatility. Ultimately, these three structural changes are translated into quantitative adjustments for ASP, TAM, and OPM, leading to a paradigm shift in valuation—when earnings visibility transitions from 'pulse-like' to 'platform-like,' the valuation anchor naturally shifts from asset-based P/B to earnings-based P/E.

Methodology notes

• Cycle and Business Health FrameworkTurning Point Analysis

  Analysis of Structural Changes in Memory Cycle Length and Intensity

  Traditional memory cycle analysis focuses on short-term fluctuations in inventory and prices. However, by comparing the 2017-2018 period with the current cycle, this report identifies how AI demand, HBM supply constraints, and LTA contracts jointly extend the duration of the peak business health period. This represents a deepened application of the turning point analysis paradigm, shifting from merely tracking price signals to assessing whether the underlying structures supporting business health have undergone qualitative changes.

• Industry/Industrial Analysis FrameworkSupply-demand framework

  Quantifying the Cannibalization Effect of HBM on Traditional Memory Capacity

  Beyond qualitatively pointing out that HBM occupies capacity, the report uses a conversion coefficient ('HBM consumes 3-4 times the wafers of traditional DRAM') combined with cleanroom allocation ratios to quantitatively calculate that the growth rate of actual available capacity for traditional memory is compressed by 4-5 percentage points. This is a typical application of the supply-demand framework where product mix changes impact effective supply.

• Valuation MethodologyPE/PEG valuation

  Conditions for Switching the Valuation Anchor of Cyclical Stocks from P/B to P/E

  Historically, memory stocks were valued using P/B due to frequent losses. However, this report argues that when mechanisms like LTAs make high earnings sustainable, P/E valuation becomes reasonable. This demonstrates that the choice of valuation method is not static but depends on structural improvements in earnings quality and visibility, representing a dynamic calibration of the applicability boundary of PE valuation.

• Competition and Strategy FrameworkValue chain analysis

  Enhanced Bargaining Power of the Memory Segment within the AI Value Chain

  By analyzing the evolution of AI data center architectures, the report highlights that memory has transformed from a CPU accessory into a key bottleneck constraining model expansion, elevating its strategic position within the AI value chain. This positional change directly translates into the commercial reality of customers willing to sign LTAs with upfront payments and price floors, representing a concrete application of value chain analysis in the semiconductor sub-sector.

• Event Gaming and Behavioral FinanceExpectation Gap/Expectation Management

  Valuation Repair Space Created by Market Lag in Recognizing 'Peak-to-Stay-Higher'

  The report explicitly states that the market still values memory stocks at single-digit P/Es, indicating that investors have not yet fully digested the structural changes of this cycle. This expectation gap forms the basis for upward valuation revision, a typical expectation gap analysis scenario: when fundamental changes are persistent but market reactions remain stuck in old paradigms, systematic revaluation opportunities arise.

Asset mapping & comparison

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

• Samsung Electronics (005930.KS)

  Primary Beneficiary: Largest exposure in traditional DRAM, with accelerating HBM progress

  Strengths

  Leading share in traditional DRAM market; Increasing share in Google TPU and Nvidia HBM4; 2026E ROE expected to reach 52%

  Weaknesses

  Mobile business dragged by rising memory costs; 2026E mobile OPM expected to fall to 2%, a historical low; Labor agreements may bring additional costs

  Comparison

  Greater elasticity in traditional DRAM shortage compared to SK Hynix; however, slightly behind in HBM technology progression

  Risks

  Worsening memory supply-demand; Sharp contraction in mobile margins; Loss of mobile OLED market share

• SK Hynix (000660.KS)

  HBM Leader, Core Subject of Re-rating

  Strengths

  Absolute leader in HBM market; 2027E HBM ASP expected to rise +50% YoY; DRAM/NAND OP margins maintained at approx. 80%/mid-60%

  Weaknesses

  2026E HBM shipment volume slightly below expectations; High dependence on a single product line

  Comparison

  Superior HBM technology compared to Samsung but smaller traditional DRAM share; Upside clearer after valuation switch

  Risks

  Worsening memory supply-demand and technical migration delays; Weak terminal demand; Samsung's HBM progress exceeding expectations; Reduction in AI Capex

• Kioxia Holdings (285A.T)

  Direct Beneficiary of NAND Prices Staying High for Longer

  Strengths

  Pure-play NAND manufacturer sensitive to eSSD demand growth; FY3/28E implied 7.8x P/E is attractive

  Weaknesses

  No DRAM/HBM business hedge; Smaller scale compared to Samsung/Hynix

  Comparison

  Performance elasticity concentrated on NAND cycle compared to integrated firms; Lower starting valuation

  Risks

  NAND recovery falling short of expectations; YMTC technology iteration; Obstacles in eSSD market expansion

Key data

• 2026E DRAM Supply/Demand Deficit-5.0%Expanded from prior -4.9%, primarily due to upward revision in server DRAM demand
• 2027E DRAM Supply/Demand Deficit-5.9%Significantly expanded from prior -2.5%, representing the tightest levels historically
• 2027E HBM TAM$116 billionUp 54% from prior $75 billion, mainly due to ASP catching up to traditional DRAM
• Samsung 2026E Traditional DRAM ASP Increase+326% yoyUp from prior +312%, reflecting tighter supply-demand conditions
• SK Hynix 2026E Traditional DRAM ASP Increase+316% yoyUp from prior +309%
• 2026-2030 Traditional DRAM Capacity CAGR7-8%4-5 percentage points lower than the 12% seen in 2017-2018
• Samsung 2026E Operating Profit ForecastKRW 3.74 trillionYoY growth exceeding 8x, ROE reaching 52%
• SK Hynix Target P/E Multiple9xBased on average 2026-2027E EPS, matching Micron's average during its upcycle

Impact & implications

The report posits that structural changes in this memory cycle will permanently raise the earnings baseline for major players rather than causing a brief pulse. For Samsung, its largest exposure in traditional DRAM makes it the primary beneficiary of this shortage, while accelerating progress in HBM (especially increasing shares in Google TPU and Nvidia HBM4) will further bolster profitability, with 2026E dividend yield potentially exceeding consensus by over threefold. For SK Hynix, the valuation methodology shift marks a transition in market perception from 'cyclical stock' to 'core supplier of AI infrastructure.' If it continues to deliver high earnings, there remains room for valuation to approach 20x P/E. For Kioxia, expectations of NAND prices staying high for longer directly support earnings recovery, and the Buy upgrade reflects confidence in confirming the cyclical bottom. Overall, the memory sector is transitioning from a pure Beta trade to a配置 target with both Alpha characteristics.

Risks

• Major deterioration in memory supply-demand relationship
• Weak demand in smartphones/PCs/servers dragging traditional memory demand
• Technology migration delays affecting capacity release timing
• Samsung's HBM progress exceeding expectations squeezing SK Hynix's share
• Reduction in AI-related capital expenditure leading to HBM demand falling short of expectations
• Sharp contraction in Samsung's mobile margins
• Loss of mobile OLED market share
• Uncertainty regarding additional costs from labor agreements

What to watch

• Whether HBM ASP achieves the projected 44% YoY increase in 2027 and completes the price catch-up with traditional DRAM
• Actual signing scale of long-term agreements, proportion of upfront payments, and execution of floor prices
• Progress of new cleanrooms commissioned by the top three vendors and allocation ratios for HBM capacity
• Whether AI server shipment volumes and token consumption growth rates continue to exceed expectations to support memory demand
• Whether SK Hynix can maintain a 9x P/E valuation or evolve toward higher multiples