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

Overview

This is a Morgan Stanley Greater China semiconductor sector presentation for investors. Its core topic is the beneficiary paths across GPU, XPU, ASIC, CPU orchestration, advanced packaging, memory, testing, and domestic Chinese AI accelerators in future AI infrastructure buildout. The overall stance is positive, arguing that AI-related demand remains the strongest structural growth driver for the semiconductor industry, although non-AI semiconductors may come under pressure in 2026.

Core views

The report's core views include: first, global AI semiconductor demand continues to be driven by cloud capex, training and inference workloads, Agentic AI, and sovereign AI; second, TSMC's position in advanced processes, CoWoS, and SoIC makes it a key beneficiary; third, NVIDIA still dominates AI compute wafer consumption, but CSPs still need to develop their own custom chips; fourth, China AI accelerators are driven by domestic demand, supply-chain resilience, and cost advantages, with long-term TAM and self-sufficiency expected to rise; fifth, AI is creating tight supply-demand conditions and price support for NAND, NOR, DDR4, and test equipment demand.

Analysis framework

The report uses supply-chain data, cloud capex tracking, TAM estimation, advanced packaging capacity breakdowns, customer demand structure, cost/performance comparisons, TCO and cost-per-token analysis, market share forecasts, and valuation comparisons to assess the relative attractiveness of different assets across the industry chain.

Methodology notes

• Market Size EstimationTAM Scenario Analysis

  Using submarkets such as cloud AI semiconductors, CPU orchestration, and China AI chips as targets, the report constructs base-case and bull-case scenarios.

  The report provides estimates such as US$485bn for cloud AI semiconductors in the 2026e bull case, about US$753bn for global AI semiconductors by 2030, and US$91bn for China AI chips by 2030 to assess long-term growth potential.

• Supply Chain AnalysisCapacity and Customer Mix Breakdown

  The report breaks down areas including TSMC advanced nodes, CoWoS, SoIC, wafer consumption, and NVIDIA GB200/300 rack supply and demand.

  By examining capacity expansion, customer mix, and packaging technology roadmaps, the report judges who can capture the structural demand created by AI infrastructure expansion.

• Cost Efficiency AnalysisTCO and Cost-per-Token Comparison

  The report compares domestic Chinese AI accelerators with NVIDIA processors in terms of total cost of ownership and unit token cost in China's inference scenarios.

  The report argues that domestic chips, because of their significantly lower prices, can deliver better performance per dollar in certain inference scenarios.

• Relative ValuationSemiconductor Subsector Valuation Comparison

  The report compares sectors including fabless, power semiconductors, FPGA, analog, foundry, backend, memory, IDM, and semiconductor equipment.

  Valuation comparisons are used to help identify relative opportunities and risks across the AI theme, memory recovery, test equipment, and mature nodes.

Asset mapping & comparison

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

• TSMC

  Core beneficiary of advanced AI processes and CoWoS/SoIC

  Strengths

  Advanced nodes, CoWoS capacity, AI customer demand, and rising wafer consumption together support revenue and margins.

  Weaknesses

  High capex and capacity expansion requirements, with sensitivity to demand changes from a small number of large AI customers.

  Comparison

  Compared with typical foundries, TSMC occupies a more critical position in the AI GPU, ASIC, and advanced packaging supply chain.

  Risks

  Slowing AI capex, advanced packaging supply-demand mismatch, geopolitics, and customer concentration risk.

• MediaTek

  Potential beneficiary of AI and edge devices, listed as one of the Top Picks

  Strengths

  The report includes it in the AI OW list and mentions that a new Nvidia/MediaTek AI PC could trigger replacement demand in edge devices.

  Weaknesses

  There is uncertainty around the edge AI replacement cycle and the price elasticity of consumer electronics.

  Comparison

  Compared with the pure cloud AI chain, MediaTek is more exposed to the demand elasticity of end-device and edge AI.

  Risks

  AI PC penetration below expectations, weak end demand, and chip cost inflation compressing margins.

• NVIDIA

  The leading reference asset for global AI GPU and server rack demand

  Strengths

  The report says NVIDIA accounts for the majority of AI compute wafer consumption and analyzes GB200/300 rack supply and demand.

  Weaknesses

  CSP self-developed ASICs continue to advance and may divert part of AI accelerator demand over the long term.

  Comparison

  NVIDIA remains the benchmark for AI GPUs, but CSP custom chips and domestic Chinese accelerators face substitution pressure in specific scenarios.

  Risks

  Supply-chain bottlenecks, customer in-house ASICs, export controls, and substitution in the China market.

• Cambricon

  A key China AI accelerator name, with the report expressing an OW view

  Strengths

  The report emphasizes its inference performance and customer anchoring, and expects 2025-2028e revenue CAGR of 90%.

  Weaknesses

  Advanced process access, software ecosystem, and supply-chain capabilities still require continued validation.

  Comparison

  In the comparison among Cambricon, MetaX, and Iluvatar, Cambricon is positioned as stronger in inference performance and customer base.

  Risks

  Order conversion, domestic advanced-node capacity, intensifying competition, and policy/export restrictions.

• Iluvatar

  A key China AI accelerator name, with the report expressing an OW view

  Strengths

  The report emphasizes supply-chain resilience and order visibility, and expects 2025-2028e revenue CAGR of 122%.

  Weaknesses

  Gross margin is described as moderate, while commercial scale and product roadmap are still being validated.

  Comparison

  Compared with Cambricon, Iluvatar's strengths are more concentrated in supply-chain resilience, order visibility, and an aggressive product roadmap.

  Risks

  Gross margin pressure, product iteration missing expectations, customer concentration, and capacity constraints.

• Hon Precision

  Beneficiary of longer AI testing time and handler market expansion

  Strengths

  The report expects its handler market share to rise from 28% in 2024 to 42% in 2027e, with 2025-2028 revenue CAGR of 67%.

  Weaknesses

  Sensitive to AI/HPC testing demand and handler market expansion.

  Comparison

  Within the test equipment chain, Hon Precision benefits more directly from longer testing time and stronger handler demand.

  Risks

  Volatility in AI testing capex, delayed customer procurement, and competition causing market-share gains to fall short of expectations.

• MPI

  Technological leader in probe cards with optionality in CPO

  Strengths

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

  Weaknesses

  Growth depends on penetration of high-end probe cards and volume ramp in CPO equipment.

  Comparison

  Compared with typical test consumables companies, MPI has stronger probe card technology and CPO equipment-driven upside.

  Risks

  Delayed CPO adoption, failed technology iteration, and volatile AI customer demand.

• WinWay

  Beneficiary of sockets and probe cards as AI packaging complexity rises

  Strengths

  The report expects its market share to rise from 8.6% in 2024 to 20% in 2028, while pin capacity expands from 3.5mn per month in 2025 to 9mn per month in 2026.

  Weaknesses

  Capacity expansion requires execution, and the company is sensitive to rising AI/HPC packaging complexity.

  Comparison

  Compared with a pure socket vendor, WinWay emphasizes an integrated probe card and socket solution.

  Risks

  Capacity utilization, customer qualification, and AI packaging complexity rising less than expected.

• Macronix、AP Memory、Nanya Tech、Winbond、GigaDevice

  Beneficiary chain of memory shortages and the price cycle

  Strengths

  The report lists Macronix as the Memory Top Pick and discusses tight supply-demand conditions in NAND, NOR, and DDR4.

  Weaknesses

  Memory is highly cyclical, and price increases may be affected by demand elasticity and customer inventories.

  Comparison

  Compared with non-AI logic chips, the memory chain is more clearly supported by AI memory demand and supply constraints in DDR4/NOR.

  Risks

  Price declines, overly rapid supply recovery, and weak end demand.

• SICC、InnoScience

  China compound semiconductor and power semiconductor-related names

  Strengths

  The report prefers SiC over GaN and mentions that SiC penetration could exceed 50% by 2030.

  Weaknesses

  Capacity expansion brings depreciation pressure that may weigh on profitability.

  Comparison

  SICC is more supported by rising SiC penetration, while InnoScience is tied to GaN market opportunities.

  Risks

  EV demand volatility, overly rapid capacity expansion, price competition, and depreciation pressure.

Key data

• Global AI Semiconductor TAM~US$753bn by 2030The report expects AI semiconductor TAM to reach about US$753bn by 2030.
• Cloud AI Semiconductor Bull-Case ScenarioUS$485bn in 2026eIn a bull-case scenario driven by supply-chain data, cloud AI semiconductor TAM could reach US$485bn in 2026e.
• Top 4 CSP Capex+95% YoY in 1Q26Amazon, Google, Microsoft, and Meta posted 95% YoY capex growth in 1Q26.
• Top 11 Listed Global CSP Cloud CapexUS$811bn in 2026The report estimates cloud capex for major listed global CSPs at about US$811bn in 2026, excluding sovereign AI.
• TSMC AI Semiconductor Revenue Mix>30% of 2026e revenueThe report expects TSMC's AI semiconductor revenue to account for more than 30% of revenue in 2026e.
• CPU Orchestration TAM Bull-Case ScenarioUS$238bnMorgan Stanley's bull-case scenario implies CPU orchestration TAM of US$238bn.
• Agentic CPU TAM Growth Rate251% CAGR over 2026-30The report's top-down model estimates Agentic CPU TAM CAGR at 251% during 2026-2030.
• Test Equipment Market Growth35% CAGR during 2024-27The report expects the test equipment market to grow at a 35% CAGR from 2024 to 2027.
• Handler Market SizeUS$436mn in 2023 to US$6.6bn in 2027eThe report expects the overall handler market to grow from US$436mn in 2023 to US$6.6bn in 2027e.
• China AI Chip TAMUS$91bn by 2030The report expects China AI chip TAM to grow to US$91bn by 2030.
• China AI Chip Self-Sufficiency70% in 2030eThe report expects China AI chip self-sufficiency to reach 70% by 2030e, corresponding to US$60bn/US$85bn.
• Cambricon Revenue Growth90% CAGR, 2025-28eThe report expects Cambricon revenue to grow at a 90% CAGR during 2025-2028e.
• Iluvatar Revenue Growth122% CAGR, 2025-28eThe report expects Iluvatar revenue to grow at a 122% CAGR during 2025-2028e.

Impact & implications

The investment implication is that AI infrastructure no longer benefits only a single GPU supplier, but instead extends across advanced foundry, advanced packaging, servers/racks, ASIC design, CPU orchestration, memory, test equipment, consumables, and domestic Chinese AI accelerators. Within Greater China's semiconductor chain, companies with advanced process capacity, packaging bottleneck resources, customer anchoring, cost advantages, or leverage to rising testing complexity are more likely to benefit; by contrast, weak non-AI demand, cost inflation, and regulatory restrictions may weigh on some traditional or highly externally dependent businesses.

Risks

• Chip inflation: rising wafer, OSAT, and memory costs may compress margins for chip design companies.
• AI crowding-out effect: AI semiconductors preferentially consume resources such as T-Glass, memory, and advanced packaging, potentially squeezing non-AI semiconductors.
• Growth pressure in non-AI semiconductors: excluding memory and NVIDIA AI GPU revenue, non-AI semiconductor growth may decline in 2026.
• Slowing AI capex: if CSP cloud capex comes in below expectations, demand across GPUs, ASICs, advanced packaging, and the server chain will be affected.
• Regulation and export controls: the report discloses compliance risks related to U.S. Executive Orders 14032 and 14105 and export control restrictions, which may affect related transactions and investments.
• China advanced-node capacity constraints: domestic AI GPU output depends on advanced processes, packaging, and supply-chain execution capability.
• Intensifying competition: competition among CSP self-developed ASICs, NVIDIA GPUs, domestic Chinese accelerators, and different packaging roadmaps may change profit allocation.
• Cyclical risks in memory and power semiconductors: changes in pricing, inventory, EV demand, and industrial demand may lead to a cycle reversal.

What to watch

• Whether cloud capex growth remains strong among the top 4 and top 11 CSPs.
• TSMC CoWoS, SoIC, and 2nm/3nm/4-5nm customer demand and capacity expansion progress.
• Changes in NVIDIA GB200/300 rack supply, demand share, and customer mix.
• Mass production and performance evolution of CSP custom ASIC projects such as AWS Trainium and Google TPU.
• China AI accelerator orders, IPO progress, domestic advanced-node capacity, and the path of self-sufficiency.
• Whether token demand, TCO, and cost per token in China's inference market continue to support substitution by domestic chips.
• Supply-demand gaps and price trends for NAND, NOR, and DDR4.
• Whether AI/HPC testing time and demand for probe cards, sockets, handlers, and CPO equipment continue rising.
• Changes in U.S. export controls, investment restrictions, and related entity lists.