AI Compute Bottlenecks Expand from GPUs to ASICs, Advanced Packaging, MLCCs and Power Semiconductors
AI summary card
AI Compute Bottlenecks Expand from GPUs to ASICs, Advanced Packaging, MLCCs and Power Semiconductors
Morgan Stanley believes AI servers and demand for large models in China are driving custom ASICs, domestic AI accelerators, MLCCs and power semiconductors into a new supply-demand inflection point.
- Demand for higher-capacity, low-ESL and embedded MLCCs is rising in AI servers, and the report estimates AI server MLCC demand will approach $1 billion by 2027.
- Cloud providers still need custom chips even with strong NVIDIA GPU capacity; shipment forecasts for AWS Trainium and Google TPU suggest continued growth in ASIC programs.
- China AI demand is supported by large-model token growth at companies such as ByteDance, higher China prices for NVIDIA 5090, and cloud capex trends; domestic chips show strong performance per dollar at significantly lower prices.
- China's infrastructure capabilities, multi-chip packaging, larger rack-scale clusters and expansion path are viewed as key to narrowing the technology gap.
- Demand for power semiconductors is somewhat bifurcated between automotive and industrial end markets, but supply additions are limited; capital spending by global leaders has declined for two consecutive years, and new capacity may slow in 2026-2028.
Report interpretation
Overview
This report is Morgan Stanley's industry tracking of European semiconductors and the global AI hardware supply chain, covering AI GPU/ASIC, China LLMs, advanced packaging, MLCC and power semiconductors. The core view is that the bottleneck to AI growth is no longer concentrated only in high-end GPUs themselves, but has expanded to custom ASICs, packaging, power delivery architecture, passive components and localized compute supply chains.
Core views
The report argues that upgrades in AI server architecture will materially increase MLCC usage and specification requirements, and the high industry concentration means leading suppliers are more likely to benefit; cloud providers will continue to pursue in-house or custom ASICs, and the AWS Trainium and Google TPU roadmaps indicate continued expansion in future programs and shipment scale; China AI compute demand remains strong, and domestic AI accelerators are competitive thanks to lower prices, availability and local infrastructure support; power semiconductors may move higher on supply constraints, as capex has declined and new capacity growth is slowing.
Analysis framework
The report cross-checks multiple dimensions, including supply-chain decomposition, cloud provider ASIC shipment forecasts, China AI GPU demand tracking, token pricing and inference economics comparisons, domestic AI accelerator products and customer orders, MLCC industry concentration and volume recovery, and power semiconductor demand and supply cycles.
Methodology notes
Extending from GPUs to packaging, MLCCs, power devices and localized compute
The report breaks AI hardware bottlenecks into compute chips, packaging, power delivery, passive components and supply-chain availability to identify which links may become the next stage of incremental investment opportunities.
Trainium, TPU and major cloud providers' custom-chip forecasts
Generational product planning and annual shipment forecasts are used to judge the continuity and scale growth of cloud vendors' custom-chip programs.
Cost comparison between domestic AI chips and NVIDIA processors in the China market
The report focuses on total cost of ownership, per-token cost and performance per dollar, and concludes that domestic chips have economic advantages in China's inference use cases because their pricing is meaningfully lower.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- Cambricon Technology Corporation (688256.SS)A representative Chinese AI accelerator and ASIC company, marked OW in the report
- Strengths
- Leading inference performance and strong customer anchoring; the report expects 2025-2028e revenue CAGR of 90%.
- Weaknesses
- Still constrained by advanced process availability, supply chain, customer concentration and the maturity of the domestic ecosystem.
- Comparison
- The report compares it with MetaX and Iluvatar, positioning it more toward ASIC and inference-performance advantages.
- Risks
- Slower-than-expected technology iteration, weaker order conversion, process supply constraints and intensifying competition.
- Iluvatar CoreX Semiconductor Co. Ltd (9903.HK)A Chinese AI GPGPU vendor, marked OW in the report
- Strengths
- Strong supply-chain resilience and order visibility, with an active product roadmap; the report expects 2025-2028e revenue CAGR of 122%.
- Weaknesses
- Gross margin is described as modest, and commercialization and scale-up still need to be proven.
- Comparison
- In the Cambricon, MetaX and Iluvatar comparison, it leans toward the GPGPU route.
- Risks
- Listing progress, order execution, gross margin, process node and ecosystem compatibility risks.
- MetaX Integrated Circuits (688802.SS)A Chinese AI GPGPU vendor, marked EW in the report
- Strengths
- Covers AI training, inference and graphics rendering product lines, with C, N and G series.
- Weaknesses
- Compared with Cambricon and Iluvatar, the report assigns it a more neutral investment rating.
- Comparison
- Listed alongside Cambricon and Iluvatar as a key Chinese AI GPGPU supplier.
- Risks
- Product competitiveness, customer orders, domestic substitution pace and supply-chain constraints.
- Murata ManufacturingA global MLCC leader, marked OW in the report
- Strengths
- One of the top five MLCC suppliers globally, benefiting from higher MLCC content and specification upgrades in AI servers.
- Weaknesses
- Demand recovery still depends on downstream cycles in consumer electronics, servers and automotive.
- Comparison
- Listed together with Samsung Electro-Mechanics as a beneficiary among MLCC leaders.
- Risks
- Price pressure, inventory cycles and AI server demand coming in below expectations.
- Samsung Electro-Mechanics (009150.KS)An MLCC supplier, marked OW in the report and disclosed with historical share price and target price data
- Strengths
- Operating in a highly concentrated MLCC industry and may benefit from rising AI server demand.
- Weaknesses
- Industry cyclicality and changes in end demand affect profitability.
- Comparison
- Along with Murata, it is one of the MLCC leaders mentioned in the report.
- Risks
- Weaker-than-expected demand recovery, price competition, FX and customer-mix changes.
- NVIDIA Corp. (NVDA.O)The core AI GPU supplier and the benchmark for China AI demand comparisons
- Strengths
- Strong supply capability for high-performance AI GPUs and still the global benchmark for AI compute.
- Weaknesses
- In China, price, availability and regulatory restrictions create room for domestic substitution and ASIC projects.
- Comparison
- The report compares domestic chips with NVIDIA processors on TCO and per-token cost.
- Risks
- Regulatory restrictions, supply-chain constraints and customer migration to in-house ASICs.
Key data
- Report date2026-06-23The report states that all information is as of June 23, 2026.
- AI server MLCC demandClose to $1 billionThe report estimates that AI server MLCC demand will approach US$1bn by 2027, driven by higher content and spec upgrades.
- Global MLCC market concentrationTop five suppliers account for about 87%CY25 estimate; Murata and SEMCO are marked OW.
- AWS Trainium total shipment forecast2025e 1570k units; 2026e 1600k units; 2027e 1800k units; 2028e 2000k unitsFrom the Trainium Forecasts table in the report.
- Google TPU total shipment forecast2025e 1750k units; 2026e 3700k units; 2027e 6150k units; 2028e >6000k unitsFrom the TPU Forecasts table in the report.
- Cambricon revenue forecast2025-2028e CAGR 90%The report expects Cambricon revenue to expand at a 90% CAGR during 2025-2028e.
- Iluvatar revenue forecast2025-2028e CAGR 122%The report expects Iluvatar revenue to expand at a 122% CAGR during 2025-2028e.
- Discrete semiconductor end-demand mixAutomotive and industrial account for about 70%The report says automotive and industrial account for about 70% of discrete semiconductor end demand in 2025.
- Industrial automation revenue growthUp 21% YoY in 1Q26The report says industrial automation company revenue grew 21% YoY in 1Q26.
Impact & implications
The investment implication is that opportunities across the AI compute chain may broaden from a single GPU leader to custom ASICs, advanced packaging, MLCCs, power semiconductors and China-based AI accelerators. For European semiconductors, the overall view is In-Line, but supply constraints and AI-related demand provide structural opportunities; for China AI hardware, the key catalysts are better inference economics, greater order visibility and improved supply-chain resilience for domestic chips.
Risks
- AI application demand or token growth may fall short of expectations, causing GPU, ASIC, MLCC and power semiconductor demand to undershoot forecasts.
- Cloud providers' in-house ASIC programs may miss expectations on schedule, yield or cost.
- China's domestic AI chips may be constrained by process nodes, packaging, HBM, software ecosystems and customer qualification cycles.
- MLCCs and power semiconductors face inventory cycles, price competition and volatile end-demand.
- Geopolitics, export controls and regulatory changes could alter the AI chip supply chain and capex path.
- The report discloses that Morgan Stanley has investment-banking or other service relationships with several covered companies, and investors should note the potential conflicts of interest.
What to watch
- Execution of AWS Trainium3/4 and Google TPU v8/v9 shipment plans.
- Trends in China large-model token volumes, token pricing and inference costs.
- Order conversion from major CSPs and state-owned customers for domestic AI accelerator vendors.
- AI server MLCC content per server, spec upgrades and monthly sales trends among Taiwanese suppliers.
- Expansion and adoption progress for semiconductor solutions such as CoWoS, SoIC, HBM, CPO and GaN HVDC 800V.
- Capital spending by global power semiconductor leaders and the pace of new capacity additions in 2026-2028.