AI infrastructure continues to reshape the main investment theme for Greater China semiconductors

Overview

This report is a Greater China semiconductor investor presentation focused on future AI infrastructure buildout and the beneficiary landscape for GPUs, XPUs, and domestic AI accelerators. Its core view is that AI demand remains the strongest structural driver in the semiconductor cycle, with cloud capex, advanced packaging, AI ASIC, Agentic AI CPU, memory shortages, rising test complexity, and China AI inference demand jointly creating multiple investable supply chains.

Core views

Morgan Stanley believes the overall semiconductor industry view is Attractive, but growth divergence is clear: AI and memory are significantly stronger than non-AI semiconductors, and non-AI semiconductors may decline in 2026 after excluding memory and NVIDIA AI GPU revenue. The report prefers AI, memory, China AI semiconductors/WFE, test equipment, and consumables, while flagging chip inflation and AI crowding out non-AI demand. In China, DeepSeek is driving low-cost inference demand, while improvements in domestic foundry and infrastructure capabilities are expected to boost the share and self-sufficiency of local AI accelerators.

Analysis framework

The report uses supply chain data, cloud capex tracking, TAM scenario analysis, advanced packaging capacity projections, memory supply-demand models, AI accelerator TCO/performance-cost comparisons, and cross-subsector valuation comparisons to identify beneficiary areas and key companies.

Methodology notes

• Demand and TAM modelingAI Semi TAM and Agentic CPU TAM

  Estimate future market size through demand for cloud AI, inference, training, CPU orchestration, and China AI accelerators.

  The report presents estimates such as global AI semiconductors reaching about US$753bn by 2030, a CPU orchestration bull-case scenario of US$238bn, and China AI chips reaching about US$91bn by 2030e, to assess long-term growth potential.

• Supply chain and capex trackingMorgan Stanley cloud capex tracker

  Use major CSP capex, capex-to-EBITDA, and relationships with TSMC capex to observe AI infrastructure demand.

  Top 4 CSP capex grew 95% YoY in 1Q26, while 2026 cloud capex for the top 11 listed global CSPs is about US$811bn, supporting an optimistic view on cloud AI semiconductor demand.

• Supply-demand modelAdvanced packaging, memory, and test equipment supply-demand analysis

  Compare capacity and demand for CoWoS, SoIC, NAND, NOR, DDR4, and test equipment.

  The report emphasizes that TSMC may expand CoWoS capacity to 200kwpm, AI memory is causing NAND shortages, NOR shortages will continue into 2026, DDR4 shortages will extend into 2H26, and rising test time and high pin-count demand will increase elasticity for equipment and consumables.

• Cost and performance comparisonTCO, cost per token, and TPS analysis

  Use total cost of ownership, cost per token, and tokens per second to compare domestic China AI accelerators with NVIDIA processors.

  The report argues that in China scenarios, domestic chips can achieve lower TCO and similar cost per token due to significantly lower prices, thereby improving price-performance.

• Rating and valuation comparisonFoundry, Back-end, Memory, IDM, Semi Cap, Fabless, and other valuation comparisons

  Compare valuation, growth, and ratings across semiconductor subsectors.

  The report ranks areas such as AI, memory, test equipment, China AI semiconductors, and mature nodes on a relative basis, and assigns Top Pick, OW, EW, and UW groupings.

Asset mapping & comparison

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

• MediaTek

  AI Top Pick

  Strengths

  Benefiting from AI, edge device replacement, and XPU-related opportunities.

  Weaknesses

  The main body of the report does not elaborate on detailed weaknesses in the provided excerpt.

  Comparison

  Part of the preferred AI basket, alongside TSMC, SMIC, Aspeed, Alchip, and other AI semiconductor plays.

  Risks

  Weaker-than-expected AI end demand and chip inflation compressing fabless margins.

• TSMC

  Core beneficiary of advanced AI nodes and advanced packaging

  Strengths

  Benefits from CoWoS, SoIC, 2nm/3nm/4-5nm advanced nodes, NVIDIA and cloud AI wafer demand; AI semiconductor revenue may exceed 30% of 2026e revenue.

  Weaknesses

  Capacity expansion requires sustained capex, and the pace of advanced packaging delivery and customer demand may affect revenue recognition.

  Comparison

  Compared with Intel EMIB, TSMC CoWoS has developed a more mature supply chain; the report also discusses the trend toward larger package sizes.

  Risks

  Slower-than-expected CoWoS/SoIC expansion, customer capex slowdown, advanced node yield issues, or geopolitical risks.

• Macronix

  Memory Top Pick

  Strengths

  Benefits from tight supply-demand conditions in NAND, NOR, and traditional memory.

  Weaknesses

  DDR4 spot prices are constrained, and memory price-cycle volatility may affect profits.

  Comparison

  Part of the beneficiary memory basket alongside AP Memory, Nanya Tech, Winbond, and GigaDevice.

  Risks

  Supply recovery faster than expected and weak downstream non-AI demand.

• Hon Precision

  Beneficiary of test equipment and handlers, OW

  Strengths

  Longer AI test times are expanding the handler market; the report expects its market share to rise from 28% in 2024 to 42% in 2027e, with 2025-28 revenue CAGR of 67%.

  Weaknesses

  Growth depends on expansion in AI/HPC test equipment and the pace of customer capacity additions.

  Comparison

  Belongs to the highly elastic test equipment and consumables segment alongside MPI and WinWay.

  Risks

  Delayed test equipment orders, shrinking customer capex, and execution and capacity expansion risks.

• MPI

  Probe card technology leader with CPO optionality, OW

  Strengths

  The report expects its probe card market share to rise from 8% in 2024 to 20% in 2028e, with EPS CAGR of 94% in 2025-28e.

  Weaknesses

  Sensitive to demand for advanced testing and the commercialization pace of CPO equipment.

  Comparison

  Compared with traditional test consumables, MPI is more focused on probe cards and CPO equipment drivers.

  Risks

  Slower-than-expected CPO adoption, intensifying probe card competition, and weaker-than-expected AI testing demand.

• WinWay

  Socket leader under rising AI packaging complexity, OW

  Strengths

  Vertically integrated probe card and socket solutions; the report expects share to rise from 8.6% in 2024 to 20% in 2028, while monthly pin capacity expands from 3.5mn in 2025 to 9mn in 2026.

  Weaknesses

  Capacity expansion and realization of demand for high pin-count sockets are key.

  Comparison

  Along with Hon Precision and MPI, it benefits from longer test times and the AI/HPC high pin-count trend.

  Risks

  Capacity ramp-up, customer qualification, and changes in AI packaging roadmaps.

• Cambricon

  Key China AI accelerator name, OW

  Strengths

  The report says it leads in inference performance and customer lock-in, and expects 2025-28e revenue CAGR of 90%.

  Weaknesses

  Still constrained by advanced process technology, supply chain, and customer order conversion.

  Comparison

  The report focuses on Cambricon, MetaX, and Iluvatar, with Cambricon standing out more in inference performance.

  Risks

  Weaker-than-expected domestic AI demand, insufficient advanced-node capacity, export controls, and regulatory restrictions.

• Iluvatar

  Key China AI accelerator name, OW

  Strengths

  The report emphasizes its supply chain resilience and order visibility, and expects 2025-28e revenue CAGR of 122%.

  Weaknesses

  Gross margin is described as moderate, and earnings quality may be affected by pricing and scale.

  Comparison

  Alongside Cambricon, it is one of the report's favored China AI GPGPU vendors, while MetaX is EW.

  Risks

  Order execution, supply chain execution, price competition, and regulatory risks.

• SICC

  Preferred SiC name, OW

  Strengths

  The report prefers SiC over GaN and expects SiC penetration to exceed 50% by 2030.

  Weaknesses

  Large-scale capacity expansion may bring depreciation pressure.

  Comparison

  Compared with InnoScience's GaN exposure, the report has a stronger preference for SiC.

  Risks

  Weaker-than-expected NEV and industrial demand, overcapacity, and earnings dragged by depreciation.

• InnoScience

  GaN name, EW

  Strengths

  The report says it has a dominant position in the GaN market, with 2025 TAM of about US$584mn.

  Weaknesses

  The report has a lower preference for it relative to SiC.

  Comparison

  Compared with SICC, the report explicitly states a preference for SiC over GaN.

  Risks

  Slower-than-expected GaN penetration, competition, and capacity expansion pressure.

Key data

• Global AI semiconductor TAMabout US$753bn by 2030The report expects the AI semiconductor market to expand significantly by 2030.
• Cloud AI semiconductor TAM bull-case scenarioUS$485bn in 2026eA bull-case scenario driven by supply chain data.
• Top 4 CSP capex+95% YoY in 1Q26Including Amazon, Google, Microsoft, and Meta.
• Top 11 listed global CSP cloud capexUS$811bn in 2026Excluding sovereign AI.
• TSMC AI semiconductor revenue share>30% of 2026e revenueDriven by demand for advanced AI nodes and advanced packaging.
• AI compute wafer consumptionUS$27bn in 2026Mostly attributable to NVIDIA.
• CPU orchestration TAMbase case US$79bn；bull case US$238bnAgentic AI creates CPU opportunities.
• Agentic CPU TAM CAGR251% over 2026-30From Morgan Stanley's top-down model.
• Test equipment market CAGR35% during 2024-27Higher AI/HPC test time and complexity.
• China AI chip TAMUS$91bn by 2030eDriven by domestic AI inference and expansion of the local supply chain.
• China AI chip self-sufficiency rate70% in 2030eThe report uses an estimated basis of US$60bn / US$85bn.
• Cambricon revenue CAGR90% during 2025-28eBenefiting from inference performance and customer lock-in.

Impact & implications

The investment implication is that gains in the AI supply chain are not limited to GPUs, but also extend to advanced process foundries, CoWoS/SoIC, ASICs, CPUs, memory, test equipment, power semiconductors, and domestic China AI accelerators. Compared with the broad semiconductor cycle, the report places greater emphasis on structural stock selection: prioritize segments driven by AI capex, packaging complexity, tight memory supply-demand, and China localization demand, while avoiding names more exposed to chip inflation, weak non-AI demand, or competitive pressure.

Risks

• Chip inflation: rising wafer, OSAT, and memory costs may pressure margins for chip design companies.
• AI crowding-out effect: AI may replace some labor demand, while the supply chain prioritizes AI semiconductors, squeezing resources for non-AI semiconductors.
• Weak non-AI semiconductor growth: after excluding memory and NVIDIA AI GPU revenue, non-AI semiconductor growth may decline in 2026.
• Cloud capex slowdown: CSP budgets, capex-to-EBITDA constraints, or ROI uncertainty may suppress AI infrastructure demand.
• Energy and infrastructure constraints: the report mentions U.S. energy, China chip capacity, and regulation as constraints on AI growth.
• Advanced packaging and process capacity risks: if CoWoS, SoIC, advanced nodes, and domestic China capacity expand less than expected, AI chip deliveries will be constrained.
• Export controls and legal compliance risks: the report involves U.S. executive orders, export controls, and restricted entities, and related transactions may face legal restrictions.
• Memory and critical material shortages: shortages of NAND, NOR, DDR4, and T-Glass may raise costs or limit shipments.
• Order conversion risk for domestic China AI accelerators: local chip performance, supply chain, customer adoption, and price competition will affect revenue growth.

What to watch

• Whether cloud capex growth and capex-to-EBITDA for the Top 4 and Top 11 CSPs remain stable.
• Whether TSMC CoWoS can expand to 200kwpm as expected, as well as demand from SoIC and advanced-node customers.
• Supply-demand dynamics for NVIDIA GB200/300 racks and changes in AI wafer consumption mix.
• Progress of AWS Trainium, Google TPU, and other CSP in-house ASIC projects.
• Whether Agentic AI continues to drive CPU orchestration demand and adoption of new CPU platforms such as Vera.
• Whether NAND, NOR, DDR4 prices and supply-demand gaps continue into 2H26.
• China AI inference demand, the DeepSeek ecosystem, ByteDance/Volcano Engine/Doubao token volumes, and domestic cloud capex.
• Orders, TPS, TCO, and customer adoption for domestic AI accelerators such as Cambricon, Iluvatar, and MetaX.
• Order trends in the test equipment market, test time, pin count, and the pace of CPO equipment adoption.
• Changes in U.S. executive orders, export controls, and regulation of China's semiconductor supply chain.