Agentic AI expands AI computing from a GPU-centric opportunity to a full-stack opportunity spanning CPUs, memory, and MLCCs

Overview

This is a Morgan Stanley industry research report on Asia-Pacific and South Korean technology, covering Agentic AI, the memory cycle, and MLCCs. Its core view is that AI workloads are shifting from “generation” to “autonomous action,” driving demand for hardware related to CPU orchestration, memory and retrieval, tool calling, state management, and power delivery. The beneficiary chain is expanding from GPUs to CPUs, DRAM/HBM, NAND, MLCCs, PCBs/substrates, and semiconductor equipment.

Core views

First, Agentic AI increases inference loops, tool calling, code execution, multi-agent distribution, and memory access, potentially making CPUs a new bottleneck and creating substantial TAM. Second, memory stocks should not be evaluated solely on P/E; P/B, inventories, year-over-year price changes, and the market's advance pricing of a downcycle are more important. Third, LTAs provide more sustainable free cash flow and shareholder returns, potentially driving a memory valuation re-rating. Fourth, AI server upgrades are increasing MLCC content and specifications. The industry is highly concentrated, positioning leading suppliers such as Murata and SEMCO to benefit.

Analysis framework

The report uses thematic decomposition and supply-chain mapping. It first breaks down the Agentic AI architecture into CPU orchestration, GPU execution, and memory layers, then maps these to CPUs, DRAM, NAND, MLCCs, PCBs/substrates, BMCs, interface chips, SPEs, and other components. It then forms investment conclusions through TAM estimates, HBM/NAND supply-demand tables, LTA valuation sensitivity, historical relative performance, and case comparisons.

Methodology notes

• AI infrastructure frameworkThree pillars of Agentic AI

  Orchestration, memory, and execution

  The report divides Agentic AI into three capabilities—CPU orchestration, knowledge/memory, and GPU execution—to explain why AI hardware opportunities are spreading from GPUs to CPUs, DRAM, NAND, and MLCCs.

• Semiconductor cycle frameworkThree memory debates

  AI spending, LTA revaluation, and cycle inflection point

  The report uses hyperscaler capital expenditure, LTA pricing, and year-over-year changes in inventories and prices to assess the position of the memory cycle, concluding that price momentum may peak while the cycle is extended.

• Valuation frameworkP/B and LTA sensitivity

  Memory valuation re-rating

  The report argues that P/B is more informative than P/E for memory stocks near cyclical peaks or troughs, and discusses valuation support using LTA P/E sensitivity tables for Samsung and SK hynix.

Asset mapping & comparison

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

• CPU

  Core beneficiary asset of the Agentic AI orchestration layer

  Strengths

  Inference loops, tool calling, multi-agent distribution, and state management increase CPU workloads.

  Weaknesses

  The opportunity depends on actual Agentic AI deployment and changes in server architecture.

  Comparison

  Compared with the GPU execution layer, the CPU opportunity arises from scheduling and orchestration bottlenecks.

  Risks

  If workloads remain highly concentrated on GPUs or Agentic AI commercialization is slower than expected, incremental CPU TAM could fall below estimates.

• DRAM/HBM

  Direct beneficiary asset of AI training, inference, and memory-layer demand

  Strengths

  HBM demand is growing rapidly, and supply-demand tables indicate a shortage persisting through 2027e.

  Weaknesses

  Memory prices are highly cyclical, with a risk of peaking around 4Q26.

  Comparison

  P/B and the inventory cycle are more suitable than P/E for observing inflection points in memory stocks.

  Risks

  A slowdown in AI capital expenditure, inventory correction, or downward revisions to Street earnings expectations.

• NAND/AI SSD

  Beneficiary asset of AI server storage and data access demand

  Strengths

  AI SSD demand could maintain high growth in 2028. If YMTC maintains supply discipline, NAND could remain tight.

  Weaknesses

  Greenfield capacity expansion on the supply side could quickly alter the balance.

  Comparison

  Compared with HBM, NAND's key variables are more concentrated in AI SSD demand and the release of YMTC capacity.

  Risks

  The release of capacity at YMTC Fab4/Fab5 and additional facilities exceeding expectations could lead to oversupply.

• MLCC

  Beneficiary asset of AI server power delivery and signal integrity upgrades

  Strengths

  High-capacitance, low-ESL, and embedded MLCCs located near CPUs/GPUs are benefiting, with VR200 showing a significant increase in MLCC content per rack versus GB300.

  Weaknesses

  Demand still depends on AI server shipments and the pace of platform specification upgrades.

  Comparison

  The MLCC industry is relatively concentrated, making leading suppliers more likely to benefit from specification upgrades.

  Risks

  A slowdown in AI server platform transitions, lower-than-expected increases in content per system, or price competition.

Key data

• Agentic AI CPU opportunityup to US$238bn by 2030CPU opportunity estimate under the bull-case scenario.
• Agentic AI DRAM demand221EB by 2030DRAM demand generated by Agentic AI under the bull-case scenario.
• HBM market sizeUS$94bn in 2027eThe table shows the HBM market increasing from US$3bn in 2023 to US$94bn in 2027e under the base case.
• Total HBM DRAM sufficiency-15% in 2027eThe report table shows that total DRAM supply and demand still have a deficit in 2027e.
• NAND demand scenarioAI SSD demand +30-60% YoY in 2028The base case assumes non-AI NAND demand growth of +5% YoY and AI SSD demand growth of +30-60% YoY.
• MLCC content increase per rackUS$4,320 vs US$1,530VR200 has higher MLCC content per rack than GB300, with an increase of more than 180%.
• AI server MLCC demandapproach US$1bn by 2027Driven by content growth and specification upgrades.
• MLCC industry concentrationtop five suppliers about 87% in CY25The five largest global suppliers account for approximately 87%, led by Murata and SEMCO.

Impact & implications

The investment implication is that the beneficiary scope of the AI hardware chain is broadening. GPUs remain important, but demand elasticity for CPUs, memory, NAND storage, and passive components is increasing. Within the South Korean technology chain, Samsung Electronics, SK hynix, Samsung Electro-Mechanics, and other memory- and MLCC-related names have thematic relevance. At the same time, supply discipline, LTA execution, price inflection points, and the sustainability of AI capital expenditure will determine whether the trend continues.

Risks

• A slowdown in AI capital expenditure or weaker-than-expected monetization could reduce incremental demand for CPUs, memory, and MLCCs.
• Year-over-year momentum in DRAM contract prices may peak around 4Q26, and the market may begin pricing in a cyclical downturn early.
• Faster greenfield expansion by YMTC could result in NAND oversupply.
• If LTAs are renegotiated or result in forced inventory, the free-cash-flow and valuation-re-rating thesis for memory could be damaged.
• Morgan Stanley discloses investment banking or other service relationships with several covered companies, so investors should be aware of potential conflicts of interest.

What to watch

• Whether hyperscaler capital expenditure continues to grow in 2Q26 and beyond.
• Whether Agentic AI applications generate more inference loops, tool calling, and enterprise workflow deployments.
• The gap between DRAM contract and spot prices, year-over-year inventory changes, and the 4Q26 price inflection point.
• HBM capacity, yields, and supply discipline at Samsung, SK hynix, and Micron.
• The pace of capacity release at YMTC Fab4/Fab5 and its impact on global NAND share.
• MLCC content and shipment trends as AI server platforms transition from GB300 to VR200/Rubin.