Agentic AI expands the value chain from GPUs to CPUs, memory, and MLCC
AI summary card
Agentic AI expands the value chain from GPUs to CPUs, memory, and MLCC
Morgan Stanley believes AI is moving from content generation to autonomous execution, with CPU orchestration, memory expansion, and high-spec MLCC becoming new growth drivers for the Asia-Pacific technology supply chain.
- Agentic AI requires more inference loops, tool calls, code execution, and multi-agent collaboration, and the report argues that CPUs are becoming the new system bottleneck.
- In the bull-case scenario, Agentic AI could create up to US$238bn of CPU opportunity and 221EB of DRAM demand by 2030.
- The core debates in storage include AI capex, LTA re-rating, and the position of the storage cycle; the report judges that the cycle is more likely to extend rather than collapse rapidly.
- NAND may sustain a tight balance through 2028 amid rising AI SSD demand and continued capacity discipline at YMTC.
- AI server architecture upgrades are increasing MLCC content per server; MLCC value per rack for VR200 is more than 180% higher than for GB300, and AI server MLCC demand could approach US$1bn by 2027.
Report interpretation
Overview
This report is Morgan Stanley's industry research on the Asia Pacific Summer School, Korean technology, and the AI supply chain, covering Agentic AI, CPU and memory opportunities, the storage cycle, HBM and NAND supply-demand dynamics, and rising MLCC value content in AI servers. The report describes AI evolution as a shift from "generation" to "autonomous action": LLMs are responsible for understanding, reasoning, and generation; CPUs handle orchestration, routing, tool calls, state management, and guardrails; GPUs continue to perform intensive execution; and memory plus external knowledge bases provide context and long-term memory.
Core views
The core views are: first, Agentic AI will increase inference, tool calling, code execution, and multi-agent fan-out, leading to a re-rating of CPU and memory demand; second, AI spending remains a key support for storage demand, but the market needs to watch for the possibility of DRAM prices peaking around 4Q26; third, LTAs can support free cash flow and shareholder returns, but the report remains neutral on valuation re-rating driven by LTAs; fourth, HBM and AI NAND remain structural growth areas, with NAND tightness depending on AI capex and YMTC's capacity expansion discipline; fifth, rising power and signal integrity requirements in AI servers are driving demand for high-capacitance, high-specification, low-ESL, and embedded MLCCs.
Analysis framework
The report uses thematic decomposition and supply-chain mapping, first identifying new bottlenecks from the Agentic AI architecture, then mapping demand to CPUs, DRAM, HBM, NAND, MLCC, PCB/substrates/materials, ODMs, and semiconductor equipment; it then uses scenario analysis, supply-demand sufficiency estimates, valuation sensitivity, and historical analogies to assess the investment implications of each segment.
Methodology notes
Breaks agentic AI into CPU orchestration, GPU execution, memory/knowledge base, and tool-calling layers.
This framework is used to determine whether the AI value chain is diffusing beyond GPUs into CPUs, memory, storage, MLCC, and server components.
Estimates supply-demand gaps based on GPU/ASIC shipments, HBM content, capacity, yields, and utilization.
The report uses HBM estimates for 2023 to 2027 and AI SSD/NAND scenarios for 2028 to assess whether storage tightness is sustainable.
Assesses cycle positioning by combining DRAM price YoY, inventory, LTAs, and P/B and P/E sensitivity.
The report argues that storage stocks trade more on cycle inflection points and P/B positioning, while P/E is less correlated with predicting future performance.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- Samsung ElectronicsCore covered name for DRAM, HBM, NAND, and Korean technology
- Strengths
- Benefits from AI memory demand, HBM capacity expansion, storage pricing, and improved LTA-related cash flow.
- Weaknesses
- If DRAM prices peak around 4Q26, earnings expectations could face downward revision pressure.
- Comparison
- Like SK hynix, it is a core Korean storage asset, but differences in HBM share, product mix, and execution could affect relative performance.
- Risks
- Downturn in the storage cycle, slower AI capex, HBM competition, and failure of LTAs to drive valuation re-rating.
- SK hynixHigh-exposure name for HBM and DRAM themes
- Strengths
- The report highlights it as a key name in DRAM/HBM supply and Korean technology coverage, benefiting from AI server memory demand.
- Weaknesses
- Valuation and earnings are sensitive to HBM supply-demand, DRAM pricing, and customer procurement cadence.
- Comparison
- Relative to Samsung Electronics, it is more purely exposed to the high-end memory cycle.
- Risks
- Rapid HBM supply increases, price peaks, customer concentration, and cyclical inventory corrections.
- Samsung Electro-MechanicsBeneficiary of MLCC and AI server power architecture upgrades
- Strengths
- Global MLCC industry concentration is high, and AI servers require more high-capacitance, low-ESL, and embedded MLCCs.
- Weaknesses
- AI server MLCC is still in a growth phase, and demand realization depends on the pace of server platform upgrades.
- Comparison
- Benefits from specification upgrades alongside global MLCC suppliers such as Murata, TDK, and Yageo.
- Risks
- AI server shipments below expectations, price competition, and delays in customer platform transitions.
- Murata、TDK、YageoExposure names in the global MLCC supply chain
- Strengths
- The industry is highly concentrated, with the top five suppliers accounting for about 87% of the global market, and specification upgrades help increase high-end product value content.
- Weaknesses
- Traditional consumer electronics and industrial demand still affect overall utilization rates.
- Comparison
- Together with SEMCO, they form the main supplier group in the MLCC industry.
- Risks
- AI demand below expectations, recovery in commodity MLCC supply, and price declines.
- NVIDIA、AMD、Intel、ArmRelevant names for the Agentic AI CPU/GPU and computing architecture theme
- Strengths
- Agentic AI brings more inference, orchestration, and tool calls, expanding computing demand.
- Weaknesses
- Value distribution may shift among CPUs, GPUs, ASICs, and the software layer.
- Comparison
- NVIDIA remains the core GPU representative, while AMD/Intel/Arm are tied to CPU and platform expansion.
- Risks
- Slower AI capex, ASIC substitution, intensifying competition, and customer in-house chip development.
- TSMC and semiconductor equipment suppliersAssets tied to AI chip manufacturing and the capex chain
- Strengths
- Expansion in AI compute supports demand for advanced process nodes, packaging, and equipment.
- Weaknesses
- Affected by customer capex cadence and geopolitical technology fragmentation.
- Comparison
- Compared with storage and MLCC, these names are more exposed to upstream manufacturing and equipment cycles.
- Risks
- US-China decoupling in AI compute, capex volatility, and capacity investment mismatches.
Key data
- Agentic AI CPU opportunityup to US$238bnIn the report's bull-case scenario, the incremental CPU TAM opportunity that Agentic AI could create by 2030.
- Agentic AI DRAM demand221EBIn the report's bull-case scenario, the DRAM demand that Agentic AI could generate by 2030.
- DRAM pricing timingpeak around 4Q26The report notes that DRAM contract price YoY has rolled over from cyclical highs, consistent with a view that prices peak around 4Q26.
- HBM market sizeUS$94bn in 2027eIn the report's base-case table, the HBM market grows from US$3bn in 2023 to US$94bn in 2027e.
- HBM demand56,085mn Gb in 2027e2027e HBM demand in the report's HBM sufficiency estimates.
- Total DRAM sufficiency-15% in 2027eThe report estimates that total HBM+DRAM supply remains insufficient relative to demand in 2027e.
- YMTC potential global NAND shareabout 24%If YMTC uses all five announced fabs for NAND, it could reach about a 24% share of the global NAND market.
- AI server MLCC demandapproaching US$1bn by 2027The report estimates that AI server MLCC demand could approach US$1bn by 2027.
- VR200 MLCC value vs. GB300more than 180% higherThe report's table shows VR200 MLCC value per rack at US$4,320, above GB300's US$1,530.
- Global MLCC market concentrationtop five suppliers about 87%The report states that the top five suppliers account for about 87% of the global MLCC market in CY25, led by Murata and SEMCO.
Impact & implications
In investment terms, the report views Agentic AI as a catalyst for further diffusion across the hardware value chain: CPUs, memory, HBM, AI NAND, and high-spec MLCCs could all benefit from incremental demand; however, storage stock trading should avoid relying only on low P/E, because price cycles and inventory corrections often lead consensus expectations. For the Korean technology sector, companies such as Samsung Electronics, SK hynix, and Samsung Electro-Mechanics have high relevance to the AI memory and MLCC themes, although the report also discloses potential conflicts of interest related to investment banking business and shareholdings.
Risks
- If AI capex slows, demand estimates for CPUs, HBM, NAND, and MLCC could be revised down.
- DRAM contract prices may peak around 4Q26, putting pressure on storage stock earnings and valuations.
- LTAs may not deliver the valuation re-rating the market expects, and the report remains neutral on this point.
- If YMTC accelerates greenfield expansion, it could increase the risk of NAND oversupply.
- A low P/E for storage stocks does not necessarily mean they are cheap; it may reflect the market discounting a downcycle in advance.
- If AI server platform upgrades are delayed, the MLCC value uplift from new architectures such as VR200 could be postponed.
- The report discloses that Morgan Stanley has investment banking, shareholding, or service relationships with several covered companies, and investors should pay attention to potential conflicts of interest.
What to watch
- Changes in hyperscale capex from 2Q26 onward, to verify whether AI infrastructure monetization continues to support hardware demand.
- Whether the divergence between DRAM contract and spot prices converges, and whether prices peak around 4Q26.
- HBM capacity, yields, utilization, and customer procurement cadence, especially whether the 2027e supply-demand gap persists.
- The YoY growth rate of AI SSD demand in 2028, as well as capacity expansion discipline at YMTC Fab4, Fab5, and its potential five-fab footprint.
- Whether free cash flow, dividends, and buyback policies at Samsung Electronics and SK hynix can drive valuation re-rating.
- The ramp-up pace of new AI server platforms such as VR200, and whether the increase in high-capacitance MLCC value per rack materializes.
- Whether monthly sales growth and the breadth of earnings revisions among key Taiwanese MLCC suppliers continue to improve.