AI infrastructure continues to reshape the semiconductor cycle, but the memory sector needs to guard against price peaks and supply expansion

Overview

This report is Morgan Stanley's global technology key debates material, focusing on four industry chain themes: AI, memory, substrates, and MLCCs. It views AI infrastructure investment as the main medium-term demand driver, while treating memory prices, long-term agreements, P/B cycle positioning, YMTC NAND capacity expansion, CPU and memory demand driven by AI agents, and passive component and substrate upgrades in AI servers as the main investment debate points.

Core views

The core views of the report are: first, cloud vendor capital expenditures remain strong, supporting demand for AI semiconductors, HBM, ASICs, GPUs, and related supply chains; second, although the memory sector is supported by HBM and AI demand, YoY DRAM contract prices may already be near a cyclical peak, and the risk of a peak in 4Q26 deserves attention; third, long-term agreements may enhance memory suppliers' revenue visibility, cash flow stability, and valuation re-rating potential; fourth, AI agents are evolving from generative AI toward autonomous action, potentially making CPU orchestration and memory the new bottlenecks; fifth, AI server power architecture and advanced packaging upgrades are boosting demand for ABF substrates and MLCCs.

Analysis framework

The report uses a combination of top-down and bottom-up analysis: it first determines demand direction from changes in cloud capital expenditures and AI computing architecture, then breaks down the supply-demand, pricing, and valuation implications for HBM, NAND, DRAM, CPU, ABF substrates, and MLCCs; it also uses P/B, P/E, long-term agreement sensitivity, HBM sufficiency, YMTC capacity scenarios, and supply chain checks to build its investment framework.

Methodology notes

• Valuation and cycleP/B cycle turning-point framework

  P/B is more suitable than P/E for identifying peaks and troughs in memory stocks

  The report believes memory stocks often look cheap when the market is already discounting the downcycle in advance, while they may look expensive near the peak of the cycle. Therefore, P/B is more informative for judging peak or trough multiples.

• Industry supply and demandHBM sufficiency analysis

  Use capacity, yield, UTR, and demand to estimate HBM/DRAM supply-demand sufficiency

  The report presents the HBM TSV capacity, yields, and implied output of Samsung, SK Hynix, and Micron, and compares them with HBM and commodity DRAM demand to assess supply-demand tightness in 2026e and 2027e.

• Contracts and valuationLTA long-term agreement framework

  Long-term agreements improve revenue visibility and valuation stability

  The report compares traditional LTAs with the current LTA framework, noting that the new agreements are usually longer-term, include supply allocation or volume commitments, pricing formulas or floor-price mechanisms, and may bring cash flow and valuation re-rating.

• Scenario analysisYMTC NAND supply scenario

  Assess NAND oversupply risk based on the pace of YMTC capacity expansion

  Using YMTC Fab4, Fab5, and all announced fabs as the basis, the report compares non-AI NAND demand, AI SSD demand growth, and YMTC capacity scenarios to judge whether NAND can remain tight.

Asset mapping & comparison

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

• Samsung Electronics

  One of the core names in the memory, HBM, and DRAM cycle

  Strengths

  It has large-scale DRAM and HBM capacity and may benefit from LTAs, HBM demand, and AI infrastructure expansion.

  Weaknesses

  It has high exposure to the traditional memory cycle; if DRAM prices peak or inventories rise, valuation and earnings expectations may come under pressure.

  Comparison

  The report compares it alongside SK Hynix and Micron in HBM sufficiency and P/E sensitivity.

  Risks

  HBM execution, yields, price-cycle declines, and changes in the competitive landscape.

• SK Hynix

  Core beneficiary of HBM and high-end DRAM

  Strengths

  HBM capacity expansion is significant, with strong relevance to the AI GPU supply chain.

  Weaknesses

  Sensitive to HBM customer demand and the memory price cycle.

  Comparison

  The report jointly estimates HBM capacity and valuation sensitivity for it with Samsung and Micron.

  Risks

  Price pressure after HBM supply becomes sufficient, customer concentration, and cyclical reversal.

• Micron

  Global DRAM and HBM supplier

  Strengths

  Participates in HBM capacity expansion and benefits from rising AI memory demand.

  Weaknesses

  Its HBM capacity base shown in the report is lower than Samsung's and SK Hynix's.

  Comparison

  Included as one of the three major memory makers in the average P/B and HBM sufficiency analysis.

  Risks

  Capacity ramp-up, yields, price cycle, and capital expenditure returns.

• YMTC

  NAND supply variable and potential source of oversupply risk

  Strengths

  If expansion is kept under control, the NAND market may still remain tight.

  Weaknesses

  If all announced fabs are directed toward NAND, it could significantly increase global share and intensify supply pressure.

  Comparison

  The report compares YMTC's 2028e capacity scenarios against non-AI NAND and AI SSD demand.

  Risks

  Faster greenfield expansion is the key risk for NAND oversupply.

• NVIDIA

  Core demand driver for HBM and the AI server supply chain

  Strengths

  The report believes it will still consume most HBM supply in 2027e, driving demand for HBM, MLCCs, and server components.

  Weaknesses

  Sensitive to the sustainability of AI capital spending and system architecture iterations.

  Comparison

  The report discusses related supply chain demand through GB200/300, Rubin systems, and AI server architecture.

  Risks

  Slower AI capital spending, competition from alternative ASICs, and supply chain bottlenecks.

• MLCC suppliers

  Beneficiary segment of AI server specification upgrades

  Strengths

  Demand for high-capacitance, low-ESL, and embedded MLCCs is rising; the industry is highly concentrated, and fundamentals are improving.

  Weaknesses

  Traditional consumer electronics and industrial cycles may still affect overall utilization.

  Comparison

  The report emphasizes that the top five suppliers account for about 87% of the global MLCC market in CY25.

  Risks

  AI server penetration below expectations, price competition, and uneven inventory replenishment timing.

• ABF substrate supply chain

  Beneficiary segment of packaging materials in the AI diffusion era

  Strengths

  The report believes demand from AI GPUs, ASICs, and servers may push ABF substrates into undersupply from 2027 onward.

  Weaknesses

  Expansion cycles, customer qualification, and product specification changes will affect the pace of realization.

  Comparison

  Compared with traditional PC demand, AI servers and high-performance computing provide stronger pull for high-end substrates.

  Risks

  AI hardware shipments below expectations, overly fast capacity release, or changes in alternative packaging technologies.

Key data

• 2027e cloud capital expendituresApproaching US$1.3tnBased on the Morgan Stanley cloud capex tracker, estimating the top 14 listed global CSPs, excluding sovereign AI.
• DRAM price cycle viewMay peak around 4Q26The report says YoY DRAM contract prices have come off cyclical highs, while valuations have not yet rerated, consistent with pricing peaking in 4Q26.
• 2027e HBM market sizeAbout US$94bn; about 56,085mn GbThe report's table shows 2023-2027e HBM market size CAGR of about 128%.
• 2027e HBM consumptionUp to about 50bn GbThe report says NVIDIA will still consume most HBM supply in 2027e.
• Agentic AI CPU opportunityUp to US$238bn in the bull caseThe report also estimates this could bring 221EB of DRAM demand by 2030.
• AI server MLCC demandClose to US$1bn in 2027Driven by rising content per AI server and specification upgrades.
• MLCC industry concentrationTop five suppliers account for about 87% of the global market in CY25The report says Murata and SEMCO are in leading positions.
• Current memory stock P/BAbout 2.5xThe report says the average for Samsung Electronics, Micron, and SK Hynix is above the historical average.

Impact & implications

For investors, the report suggests that AI infrastructure remains the strongest source of structural demand in technology hardware, but different segments are at different points in the cycle: HBM, ABF substrates, and high-end MLCCs benefit more directly; traditional DRAM and NAND require simultaneous tracking of prices, inventories, and new Chinese supply; in trading memory stocks, investors should not focus only on low P/E, but also pay attention to P/B, inventory revisions, and market expectation inflection points.

Risks

• YoY momentum in DRAM contract prices is declining, and prices may peak in 4Q26.
• An upward turn in memory inventories may pressure earnings expectations and valuations.
• If YMTC accelerates greenfield expansion, NAND may face oversupply risk.
• If AI capital spending slows, it will affect demand for HBM, GPUs, ASICs, CPUs, ABF substrates, and MLCCs.
• HBM supply expansion, yield improvement, or changes in customer demand may weaken the scarcity premium.
• The report is industry thematic material, and the ratings and target prices of multiple companies in the appendix should not be interpreted as investment advice for any single company.

What to watch

• DRAM contract prices, inventories, and P/B changes in 2H26.
• Whether Top CSPs continue to raise cloud capital expenditure guidance.
• HBM supply sufficiency, and capacity and yield ramp-up at Samsung/SK Hynix/Micron.
• Actual pull-through from NVIDIA, ASIC customers, and AI server platforms for HBM, MLCCs, and ABF substrates.
• The pace of NAND capacity additions at YMTC Fab4, Fab5, and subsequent projects.
• Monthly sales growth and the breadth of earnings revisions among Taiwan MLCC suppliers.
• The speed of AI agent adoption and its impact on CPU orchestration and DRAM demand.