Author: See the Subtle to Understand the Significant
Source: Morgan Stanley Greater China Semiconductors Research
Report Date: May 8, 2026
I. Primary Core Contradiction
Global AI capital expenditures are expanding beyond expectations, but compute supply is evolving from a NVIDIA-dominated model to a three-track system involving GPU, ASIC, and China-native chips. The core issue is not whether demand is sufficient, but who will capture a share of this expansion and how quickly non-AI semiconductors will be marginalized in the process.
II. Key Conclusions (Ranked by Trading Importance)
III. In-depth breakdown by track
3.1 Advanced Packaging (CoWoS / SoIC) — The Strongest Definitive Main Theme
Core contradiction: Demand is exploding, but only TSMC is irreplaceable in production; non-TSMC packaging providers (Amkor/ASE/UMC) are facing shrinking market share.
[Key Driver] The four major cloud providers (AWS, Google, Microsoft, Meta) increased their capital expenditures by 95% year-over-year in Q1 2026, with total annual cloud capital spending projected to reach $685 billion, directly driving demand for CoWoS/SoIC capacity.
Key data and timelines:
NVIDIA accounts for approximately 59% of CoWoS consumption, Broadcom about 20%, and AMD about 9%.
· The total value of AI computing wafers consumed in 2026 is expected to reach approximately $27.2 billion, a historical peak.
TSMC's AI chip revenue is projected to grow at a CAGR of 60% from 2024 to 2029, with AI revenue accounting for over 30% of total revenue by 2026.
[Transmission Path]
Cloud provider Capex → Orders to NVIDIA/Broadcom/Google TPU → CoWoS/SoIC become bottlenecks → TSMC’s bargaining power increases → AI revenue share continues to grow.
[Trading Insights]
TSMC is the core of the core theme; no timing is needed, and the holding rationale is clear. SoIC is the second growth curve starting in 2025; pay attention to opportunities among OSAT suppliers entering SoIC assembly (such as ASE).
3.2 Test Equipment (Handler / Socket / Probe Card) — Lowest Valuation, Most Certain Growth
[Core Contradiction]
Chip complexity has increased, causing test duration to structurally double, but market reassessment of the TAM for test equipment has lagged significantly.
[Key Drivers]
The testing duration per GPU chip generation doubles (Hopper: 350 seconds → Blackwell: 700–1000 seconds → Rubin: 1200–1400 seconds → next generation: 1800–2000 seconds); the number of test socket pins has surged from mobile-level 1,500 to AI/HPC-level 6,000, and beyond to over 10,000 in the next generation.
Data for the three core indicators:
· Global Handler market size: $436 million in 2023 → $6.6 billion in 2027, with a CAGR of over 35%
CPO optical testing requirements will scale significantly starting in 2025, entering the electrical + optical combined testing phase by 2027 (Insertion 4i)
[Transmission Path]
Increased chip size, layer count, and complexity → Longer test durations → Higher demand and prices for handlers and sockets → New CPO optical testing requirements added → Second growth curve initiated.
[Trading Insights]
These three companies represent the segment with the lowest valuation and highest growth certainty within the AI infrastructure chain, making them ideal for medium-term core positioning. Their limited market coverage and undervalued pricing make them the most compelling value opportunity today.
3.3 Chinese AI Chips (Domestic GPU/ASIC) — Long-term Irreversible, Short-term Divergence Pronounced
[Core Contradiction]
Export controls are driving demand for domestic alternatives, but China’s chip technology and mass production maturity vary widely; securing large customer orders is the key differentiator.
[Key Drivers]
DeepSeek validates the feasibility of low-cost inference → Domestic cloud providers accelerate the switch → SMIC's 7nm capacity expansion supports mass production → Domestic chip TCO advantage (30-60% lower than NVIDIA) creates a positive feedback loop.
Market size and landscape:
2026E domestic market share: Huawei 62%, Cambricon 14%, Kunlunxin 5%, T-Head 5%, others 14%.
Among the "Ten Dragons," MS focuses on comparing three targets:
[Transmission Path]
Export controls → Domestic substitution → SMIC 7nm capacity expansion → Increased volumes from Huawei/Cambricon → Local cloud providers (ByteDance/Alibaba/Tencent) switch procurement → Lower inference costs → Surge in applications → New wave of compute demand.
[Trading Insights]
Cambricon offers the highest certainty and is the preferred investment; TianShu Intelligent Chip has the greatest upside potential but is not yet profitable, carrying higher risk. Huawei (unlisted) is the biggest competitive variable—its growing market share indirectly pressures other domestic manufacturers and requires ongoing monitoring. The time window: 2026–2027 is the critical turning point when domestic AI chips transition from substitutes to mainstream solutions.
3.4 Non-AI Semiconductors (Consumer / Automotive / Industrial) — Structurally Bearish, Weak Recovery Not a Strong One
[Core Contradiction]
Supply chain resources are being systematically siphoned off by AI systems, and the recovery pace of the traditional semiconductor industry continues to lag behind expectations, with the market overestimating the resilience of the rebound.
[Key Drivers]
Manufacturing capacity, T-Glass substrates, and storage are all being redirected toward AI; non-AI chips are facing longer wait times; wafer and OSAT costs are rising; chip design companies are under pressure on gross margins.
After excluding NVIDIA AI GPUs and storage, the growth rate of non-AI semiconductors is expected to decline significantly by 2026.
· MCU inventory days remain at historically high levels (flat in 4Q25 after peak in 1Q25); major manufacturers such as STM and GD are experiencing slow inventory digestion.
·Logic foundry utilization is not expected to rebound to 80% until H2 2026, with limited recovery elasticity
· SiC outperforms GaN: Recommend SICC (OW); SiC penetration rate expected to exceed 50% by 2030; Avoid InnoScience (EW), as capacity expansion and depreciation suppress profits
[Trading Insights]
Avoid pure traditional semiconductor exposure; the MCU sector has bottomed out but is experiencing only weak recovery—avoid heavy positioning for a strong rebound. SiC is the only traditional sector sub-segment worth noting.
3.5 Storage (HBM / NAND / DDR4) — Significant internal divergence; signals require careful identification
[Core Contradiction]
AI is clearly driving explosive demand for HBM; the price increases in DDR4 and NAND are due to supply being diverted by AI, not a genuine recovery in demand, resulting in distorted signals and limited price elasticity.
[Trading Insights]
HBM remains strongly bullish, with Hynix most benefited; Macronix (NOR Flash, Top Pick) benefits from shortages and has reasonable valuation; rising NAND/DDR4 prices do not equate to improved demand—avoid chasing price increases.
Four: Macroeconomic and Geopolitical Variables as Explanatory Factors for Sector Analysis
[Geopolitical] Export controls continue to tighten
NVIDIA's export restrictions to China → Certainty of demand for domestic Chinese AI chips is rising; China's cloud capital expenditures are projected to reach $105 billion in 2026, rapidly approaching 14% of global cloud capital expenditures.
[Macro] Energy Constraints (U.S. Side)
Tight power supply for data centers in the United States is a potential ceiling for GPU demand, but it has not yet become a material constraint in the short term (by 2026).
[Industry Structure] AI Cannibalization Effect
The suction effect of AI demand on non-AI supply chains (T-Glass, traditional DRAM, consumer foundry capacity) is the core explanatory variable for the persistent underperformance of non-AI semiconductors, not cyclical factors.
[Cost Side] Tech Inflation
Wafer, OSAT, and memory costs are rising across the board, putting pressure on gross margins for chip design companies (especially those outside the AI sector); foundries like TSMC continue to strengthen their pricing power.
Five. Recommended Portfolios and Trading Frameworks
Based on analysis across all sectors, construct the following trading framework:
Six, a one-sentence summary
Buy packaged chips (TSMC), buy test equipment (Hon Precision / WinWay / MPI), buy China’s leading AI chip company (Cambricon); avoid semiconductors with non-AI-driven strong recovery expectations, favor HBM within memory, remain neutral on traditional DRAM/NAND. Time window: 2026–2027; AI capital expenditure cycle is far from over.
Risk Disclaimer: This note is compiled from publicly available research reports by Morgan Stanley and is intended solely for internal research purposes; it does not constitute any investment advice. The market is subject to uncertainty, and actual results may differ significantly from projections. Investors should exercise caution in making decisions.
Building the Future AI Infrastructure—CPU, GPU, ASIC, Optical Modules, and China’s Chips
Strong outlook for AI semiconductors
Morgan Stanley has characterized the AI semiconductor outlook as "Strong," with demand driven by three forces: the continued explosion of AI killer applications, the computing power arms race among tech giants, and sovereign AI infrastructure demands worldwide. Meanwhile, this report identifies four growth constraints—budgets, U.S. energy bottlenecks, Chinese chip production capacity, and regulation—whose essence is supply failing to keep pace with demand, rather than demand itself faltering.
In the long term, there are three structural variables to watch out for:
1) Tech inflation (rising wafer, packaging, and testing costs compress profits for chip design companies);
2) AI cannibalization effect (supply chain resources are being redirected toward AI, marginalizing non-AI semiconductors);
3) The DeepSeek effect (cost-effective inference has been validated, domestic demand for inference in China is accelerating, and local contract manufacturing supply chains are simultaneously enhancing AI GPU production capacity). The combination of these three factors forms the foundational logical framework for all subsequent market segment assessments in the report.
Valuation Comparison: Foundry, Backend, Storage, IDM (Integrated Device Manufacturer), and Semiconductor Equipment
Valuation comparison: Fabless, power semiconductors, FPGA, and analog chips
Semiconductor mega-cycle
The key takeaway is sectoral divergence rather than broad recovery: logic foundry utilization is expected to rebound to 80% in H2 2026, but non-AI semiconductor growth, excluding NVIDIA AI GPUs and memory, is projected to decline significantly in 2026; the decline in inventory days from peak levels is a positive signal, as historical data shows that inventory correction cycles often coincide with rising semiconductor indices, but the structural divergence in this recovery far exceeds previous cycles.
AI semiconductor supply chain and niche memory
By 2030, the global semiconductor industry market size could reach $1.5 trillion, with half of it coming from AI semiconductors.
Key long-term anchor points: The global semiconductor market is projected to reach $1.5 trillion by 2030, with AI semiconductors contributing approximately $753 billion; under a bull case scenario, the total addressable market (TAM) for cloud AI semiconductors is assumed to reach $235 billion by 2025 (primarily driven by NVIDIA AI GPUs), with a CAGR of 38% from 2023 to 2030, providing a foundational market space framework for valuing all subsequent sectors.
Cloud Semiconductor: A Brighter Outlook
The four major cloud providers (AWS, Google, Microsoft, Meta) saw a 95% year-over-year increase in capital expenditures in Q1 2026, representing the strongest single demand-side data point in the article; the Capex/EBITDA ratio is expected to remain stable at around 50%, indicating that the cloud providers’ expansion plans are financially sustainable; Aspeed’s profit forecasts have been consistently revised upward, and as the market leader in BMC chips for cloud AI servers, its upward revision trend confirms the authenticity of cloud demand.
Cloud capital expenditures by major cloud service providers remain strong
Microsoft's Cloud CapEx Tracker forecasts that global top 10 cloud providers' capital expenditures will reach $685 billion by 2026, approximately 10% higher than market consensus; historical charts showing synchronized growth between global cloud CapEx and TSMC's capital expenditures serve as the core visual evidence supporting the judgment that "this cycle is not short-term"; with short-life-cycle assets accounting for about 65%, cloud providers must continuously make purchases each year, ensuring demand is rigid.
The power deployment impact announced by TSMC
Estimate CoWoS wafer demand bottom-up based on rack specifications and deployment power from four major customers: NVIDIA, AMD, Broadcom, and AWS; NVIDIA’s Rubin NVL144 rack has a power rating of 220 kW and 45,000 racks, implying an annual CoWoS demand of 136k wafers in 2027, serving as the core quantitative basis for the assessment of CoWoS supply-demand tightness throughout the report.
Due to sustained strong demand for AI, TSMC may increase its CoWoS capacity to 165,000 wafers per month before 2027.
Direct CoWoS supply-side data: TSMC's capacity will increase from 120 kwpm at the end of 2025 to 165 kwpm at the end of 2027, while Non-TSMC (Amkor/UMC/ASE) capacity will simultaneously rise from 23 kwpm to 80 kwpm; on the demand side, NVIDIA accounts for approximately 59% of total CoWoS consumption, and Broadcom about 20%—high concentration means demand fluctuations from a few key customers significantly impact TSMC.
SoIC (System-on-Integrated-Chips) expansion will be a key focus for TSMC in the coming years.
SoIC is positioned as a key strategic direction for TSMC over the coming years: capacity is set to expand from 45 kWPM by end-2025 to 78 kWPM by end-2027, with demand driven by NVIDIA, AMD, Apple, and Qualcomm/Broadcom; compared to CoWoS, SoIC offers higher integration and greater technical barriers, making it TSMC’s second growth curve in advanced packaging after CoWoS, with rapid scaling expected from 2026 to 2027.
TSMC may double its CoWoS and SoIC capacity by 2025, and we expect this trend to continue through 2026.
AI wafer consumption in 2026 could reach up to $27.2 billion, with NVIDIA accounting for the majority.
Listed from bottom to top, the CoWoS capacity allocation, chip shipments, wafer consumption, and wafer value for all major AI chips in 2026 (NVIDIA B300/Rubin/H200, Google TPU, AWS Trainium3, Microsoft Maia, OpenAI Nexus); aggregated, the total wafer value consumed by AI chips in 2026 is approximately $27.2 billion, with NVIDIA dominating—this forms the most compelling foundational estimate for TSMC’s AI revenue scale.
2026 HBM (High Bandwidth Memory) consumption—up to 32 billion Gb
In 2026, total HBM demand is approximately 32,279 mn GB, with NVIDIA accounting for about 58% of consumption; list the HBM specifications (capacity, generation, supplier) for each AI chip model: the Google TPU series primarily consumes HBM3e 12hi, while AWS and Microsoft consume HBM3/HBM4; Hynix, Samsung, and Micron share the supply, with Hynix benefiting the most due to its leadership in HBM technology.
Estimated rack production of NVIDIA GB200/300
NVIDIA GB200/300 server rack supply and demand assumptions
TSMC's AI semiconductor revenue share may reach 60% between 2024 and 2029.
TSMC's AI chip revenue is projected to grow at a CAGR of 60% from 2024 to 2029, with AI revenue accounting for over 30% of total revenue by 2026; revenue streams include general-purpose AI chips, custom ASICs, CoWoS packaging and testing, and AI server CPUs, with customers Apple (19%), NVIDIA (21%), and Broadcom (11%); gross margin and EBITDA margin continue to expand, confirming the positive impact of AI business on TSMC's overall profitability.
TSMC's advanced wafer demand segmentation
Agentic AI — Expanding CPU Opportunities
AI transitions from the reasoning phase to the "action" phase, shifting the CPU/GPU ratio from GPU-heavy (1:12) to CPU-heavy (≥1:1), driven by tool-oriented tasks such as API calls, code execution, and multi-agent concurrency; MS estimates that agentic AI could create an additional $32.5–60 billion in CPU market space by 2030, with MediaTek, as an AI server CPU designer, named as a beneficiary in the report.
AI storage is causing NAND shortages; we expect the supply-demand imbalance for NOR Flash to persist until 2026.
DDR4 shortages will persist until the second half of 2026; spot prices are capped.
AI ASIC, CPO, and chip testing
AI Semiconductors: Now and Beyond — «Key Drivers»
Present the four dimensions—drivers, constraints, technical solutions, and growth perspectives—of AI semiconductors in parallel; specifically highlight three sets of comparative perspectives—inference vs. training, edge vs. cloud, and custom ASIC vs. AI GPU—these three comparisons serve as the mental map for understanding all subsequent divergences in sector assessments within the report.
Cloud service providers (CSPs) still require custom chips, even with NVIDIA's powerful AI GPUs.
According to plans by various cloud service providers (CSPs), more ASIC projects are coming soon.
How does TSMC's CoWoS compete with Intel's EMIB?
Larger package sizes are becoming a key industry trend.
Test duration has surged from 350 seconds for Hopper to 1,800–2,000 seconds for the next-generation GPU, representing the most critical structural driver in the test equipment sector; the number of test socket pins has jumped from 1,500 for mobile/PC levels to 6,000 for AI/HPC levels, and is expected to exceed 10,000 in the next generation; the global test equipment market is projected to grow at a CAGR of 35% from 2024 to 2027, while TSMC’s packaging roadmap shows continuous expansion of interposers—both factors jointly support a long-term bullish outlook for test equipment.
Describe the role divisions of Honghai Precision, WinWay Technology, and MPI in the semiconductor supply chain.
New advancements in test equipment and components: Co-Packaged Optics (CPO)
Hon Hai Precision: A key beneficiary of the structural trend of extended testing times; Morgan Stanley rating: Overweight (OW)
MPI: Market leader in probe card technology with CPO options; Morgan Stanley rating: Overweight (OW)
Yingwei Technology: Leading test socket provider with a competitive advantage in AI packaging complexity; Rating: Overweight (OW)
China's semiconductor industry: OSAT, compound semiconductors, MCU, and AI GPU
Positive on backend equipment (ASMP), but neutral on Chinese OSATs.
Favor SiC (silicon carbide) over GaN (gallium nitride): SICC (Buy) vs. InnoScience (Sell)
MCU: Hit bottom but not yet recovered
The domestic AI semiconductor market size and share continue to grow.
The domestic Chinese AI accelerator market has a clear landscape: Huawei dominates with 62%, Cambricon holds 14%, and all other players are below 10%; Chinese AI GPU companies continue to see growing market valuations, with more IPOs pending; the expansion of the market size and increased capital market activity provide the foundational context for subsequent analysis of key targets.
We expect China’s total addressable market (TAM) for AI GPUs to grow to $67 billion by 2030.
China is expanding its advanced process capacity to meet domestic AI GPU production demands.
Recent market tracking of China's AI GPU demand
AI chip value chain — China and the United States — Decoupling in AI computing
China's infrastructure strength is narrowing the perceived technology gap.
Compare the AI infrastructure capabilities between China and the U.S. across nine dimensions using a radar chart: China scores close to the U.S. in policy support, AI data center space, and software optimization (LLMs), with primary gaps concentrated in front-end wafer processing, HBM memory, and optical networking; propose a three-step strategy for China to compensate for insufficient single-chip computing power—multi-die packaging → larger racks and clusters → expanded manufacturing capacity—with Huawei CloudMatrix 384 A3 SuperPod serving as a real-world validation of this strategy.
Economics of Reasoning: Total Cost of Ownership (TCO) vs. Per-Token Cost
The total cost of ownership (TCO) for domestic Chinese AI chips is 30-60% lower than NVIDIA’s, and the top domestic accelerators can match or even outperform NVIDIA in per-token inference cost; this conclusion serves as the core evidence for the claim that “domestic substitution in China is not merely a political imperative but also an economic rationality,” directly supporting the report’s long-term bullish outlook on China’s AI chip sector.
Order placement status and potential orders for domestic AI accelerator developers
TPS (Tokens Per Second) — Performance Analysis
Due to significant price reductions, domestic chips have achieved stronger performance per dollar.
China's "Ten Dragons" of AI GPGPU manufacturers. We focus primarily on Cambricon,沐曦, and TianShu Intelligence.
Comparison of Cambricon, Muxi, and Iluvatar
A comparative analysis of China’s three most prominent AI chip companies: Cambricon (SMIC 7nm ASIC, locked-in major clients, the only profitable one), MetaX Muxi (SMIC 12nm GPGPU, backed by sovereign funds, clear technological gap), and Iluvatar (TSMC 7nm GPGPU, strong supply chain resilience); across the three dimensions of profitability, customer structure, and process node, Cambricon emerges as the most certain choice—this is the report’s implied conclusion.
Cambricon: Leading in inference performance (TFLOPS) and customer commitment; Outperform rating (OW)
TianShu Intelligence (Iluvatar): Supported by strong order visibility and supply chain resilience; Outperform rating (OW)
