AI compute expansion is pushing advanced packaging, MLCCs, and domestic AI accelerators toward a new supply-demand bottleneck
AI summary card
AI compute expansion is pushing advanced packaging, MLCCs, and domestic AI accelerators toward a new supply-demand bottleneck
Morgan Stanley tracks AI GPU/ASICs, compute demand from China's large models, advanced packaging, MLCCs, and power semiconductors, and believes higher AI server content, cloud providers' in-house chips, and the cost advantage of domestic Chinese chips will be the key investment cues.
- AI servers require more high-capacitance, low-ESL, embedded MLCCs, and AI server MLCC demand is expected to approach US$1bn by 2027.
- The global MLCC industry is highly concentrated, with the top five suppliers accounting for about 87% of the 2025 global market; Murata and SEMCO are marked OW.
- Even with NVIDIA providing powerful AI GPUs, cloud providers still need in-house ASICs, and shipment forecasts for AWS Trainium and Google TPU continue to be revised higher.
- China AI demand is supported by token growth from ByteDance, Volcano Engine, and Doubao, while domestic chips have a relative advantage in performance per dollar and TCO thanks to lower pricing.
- Power semiconductor demand is mixed, but lower capex and slower additions to capacity in 2026-2028 could create a supply-driven upcycle.
Report interpretation
Overview
This report is Morgan Stanley's data-tracking style industry research on European semiconductors and the global AI hardware supply chain, covering AI GPU/ASICs, demand from China's large language models, advanced packaging, MLCCs, and power semiconductors. The report argues that growth in AI compute will not only drive GPUs and ASICs, but also push bottlenecks into packaging, power delivery, passive components, power devices, and local supply chains.
Core views
The core views are as follows: first, AI server architecture upgrades increase MLCC count, specifications, and demand for embedded solutions, and the high concentration of the industry favors the leaders; second, global cloud providers will continue to push ahead with self-developed AI ASICs, with projects such as Trainium and TPU showing continued shipment growth; third, China's AI infrastructure, domestic chip pricing, and inference economics narrow the perceived gap versus overseas solutions; fourth, domestic AI accelerator vendors such as Cambricon, MetaX, and Iluvatar are key names to watch for orders, supply chain resilience, cost performance, and customer anchoring; fifth, while power semiconductors are seeing uneven demand recovery, their lower capex and limited new capacity give them the characteristics of a supply-driven upcycle.
Analysis framework
The report uses supply chain decomposition, company order and shipment forecasts, cloud provider in-house chip roadmaps, AI inference TCO and per-token cost comparisons, domestic AI accelerator performance/price comparisons, MLCC supply structure and monthly sales trends, and power semiconductor supply-demand cycle tracking to cross-validate AI hardware bottlenecks and investment implications.
Methodology notes
Use the pace of AWS Trainium, Google TPU, and major CSP projects to gauge AI ASIC demand.
The report presents Trainium and TPU shipment forecasts for 2023-2028e and emphasizes that even if cloud providers buy NVIDIA GPUs, they still need custom chips to optimize cost, performance, and supply security.
Compare domestic chips with NVIDIA solutions using the total cost of ownership and per-token cost of AI LLM inference.
The report notes that in the Chinese market, domestic chips may have a relative advantage in TCO and performance per dollar because of materially lower prices.
Determine cycle positioning by combining sales growth, earnings revisions, industry concentration, capex, and new capacity additions.
Monthly sales growth for Taiwanese MLCC suppliers has turned positive, earnings revisions have improved, and capex at leading power semiconductor names has declined for several consecutive periods, suggesting new capacity additions may slow in 2026-2028.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- MLCC suppliers: Murata Manufacturing, Samsung Electro-Mechanics, etc.Beneficiaries of higher AI server content and specification upgrades
- Strengths
- The industry is highly concentrated, with the top five suppliers accounting for about 87% of the global market; AI servers require higher capacitance, low-ESL, and embedded MLCCs placed close to CPUs/GPUs.
- Weaknesses
- Traditional consumer electronics and some downstream demand may still face cyclical volatility.
- Comparison
- Compared with ordinary passive components, high-end MLCCs benefit more directly from AI server power delivery and signal integrity requirements.
- Risks
- Demand recovery falls short of expectations, pricing pressure, and overly rapid capacity expansion.
- China AI accelerators: Cambricon, MetaX, IluvatarBeneficiaries of domestic AI compute substitution and inference economics
- Strengths
- Domestic chips are priced lower, and the report believes they have an advantage in TCO and performance per dollar in the Chinese market; Cambricon and Iluvatar are marked OW.
- Weaknesses
- Advanced process nodes, software ecosystems, supply chains, and customer validation remain constraints.
- Comparison
- Compared with NVIDIA solutions, absolute performance may still lag, but China's infrastructure and cluster-based solutions can narrow the perceived gap.
- Risks
- Export controls, foundry constraints, weaker-than-expected order conversion, and insufficient ecosystem adaptation.
- Cloud provider in-house AI ASICs: AWS Trainium, Google TPU, etc.AI training and inference cost optimization tools
- Strengths
- CSPs can lower costs, improve supply controllability, and build differentiated cloud service capabilities through in-house chips.
- Weaknesses
- Development cycles are long, and the software stack and customer migration costs are high.
- Comparison
- NVIDIA GPUs still offer a strong general-purpose ecosystem, while in-house ASICs are used more for specific cloud workloads and cost optimization.
- Risks
- Changes in model demand, production delays, and underperformance versus expectations.
- Power semiconductorsA potential beneficiary of a supply-constrained cycle
- Strengths
- Capex at leading players is declining, and new capacity additions may slow in 2026-2028; industrial automation revenue is relatively strong.
- Weaknesses
- Automotive and industrial demand are mixed, and year-to-date wholesale EV growth in China has been weak.
- Comparison
- Compared with AI accelerators, the logic for power semiconductors is more about the supply-demand cycle and capacity discipline.
- Risks
- Slow end-demand recovery, prolonged inventory adjustments, and intensifying price competition.
Key data
- Report date2026-06-23The report states that the information is current as of June 23, 2026.
- Industry viewIn-LineThe European semiconductor industry view shown on the front page.
- AI server MLCC demandClose to US$1bn by 2027Driven by higher content and specification upgrades.
- MLCC market concentrationTop five suppliers account for about 87%The 2025 global MLCC market, with leaders including Murata and SEMCO.
- AWS Trainium total shipment forecast2025e 1,570k; 2026e 1,600k; 2027e 1,800k; 2028e 2,000kMS Asia Research forecasts from the report tables.
- Google TPU total shipment forecast2025e 1,750k; 2026e 3,700k; 2027e 6,150k; 2028e >6,000kFrom the TPU forecast table.
- Cambricon revenue forecast2025-2028e CAGR 90%The report says Cambricon leads in inference performance and customer anchoring, with an OW rating.
- Iluvatar revenue forecast2025-2028e CAGR 122%The report highlights its supply chain resilience and order visibility, with an OW rating.
- Power semiconductor end demandAutomotive and industrial account for about 70% of 2025 discrete device end demandEnd demand is mixed, but industrial automation revenue grew 21% year over year in 1Q26.
- Selected closing pricesNVDA.O 208.65 USD; 688256.SS 1413 CNY; 688802.SS 768.98 CNY; 9903.HK 735.50 HKDRecent closing prices shown in the report tables.
Impact & implications
The investment implication is that the bottleneck in AI compute is expanding from a single GPU supply constraint to in-house ASICs, advanced packaging, HBM, MLCCs, power delivery architectures, and China's domestic AI accelerator supply chain. Beneficiaries may include leading high-end MLCC makers, domestic AI accelerator vendors with cloud customer orders and supply chain resilience, companies tied to advanced packaging, and power semiconductor companies with constrained new capacity; however, the industry as a whole is still rated In-Line, meaning opportunities are more likely to come from structural differentiation than from a broad-based upside move across the sector.
Risks
- AI capex or cloud budgets may come in below expectations.
- Energy constraints in the United States and other regions may limit AI data center expansion.
- China may face insufficient chip capacity, advanced process, HBM, and advanced packaging supply.
- Regulation, export controls, and geopolitical tensions may increase supply chain uncertainty.
- Domestic AI accelerators may underperform expectations in terms of performance, software ecosystem, or customer validation.
- If new supply in MLCCs and power semiconductors grows too quickly, pricing and earnings leverage could weaken.
What to watch
- Production and shipment pace for self-developed chips such as AWS Trainium3/4 and Google TPU v8/v9/v10.
- Token volume, token pricing, and inference demand changes for China's major large models.
- Large-customer orders, gross margin, and supply chain progress at Cambricon, MetaX, and Iluvatar.
- Specification upgrades in AI server MLCCs, supplier monthly sales, and earnings revision trends.
- Supply bottlenecks in semiconductor solutions such as CoWoS/SoIC, HBM, CPO, and GaN HVDC 800V.
- Capex, new capacity additions, and recovery in automotive/industrial demand at power semiconductor companies.