Author: Miscellaneous Talks on Seeing the Big Picture from Small Details

Source: Morgan Stanley Greater China Semiconductors Research

Report Date: May 8, 2026

I. Core Contradiction

Global AI capital expenditure has expanded beyond expectations, but the supply of computing power is evolving from a "one-man show by NVIDIA" model to a three-pronged approach of "GPU + ASIC + Chinese domestic chips . " The core issue is not whether there is enough demand, but who can capture a share of this expansion, and how quickly non-AI semiconductors will be marginalized in this process.

II. Key Conclusions (Sorted by Transaction Importance)

III. In-depth development of different tracks

3.1 Advanced Packaging (CoWoS / SoIC) — The Strongest Deterministic Theme

[Core contradiction] Demand is exploding, but TSMC is the only irreplaceable supplier in terms of production capacity; non-TSMC packaging (Amkor/ASE/UMC) faces market share squeeze.

[Key Drivers] The capital expenditure of the four major cloud vendors (AWS/Google/Microsoft/Meta) increased by 95% year-on-year in Q1 2026, and the total cloud capital expenditure is expected to reach $685 billion for the whole year. The demand for AI servers directly drives the demand for CoWoS/SoIC queues.

Key data and timelines:

NVIDIA alone accounts for approximately 59% of CoWoS consumption, Broadcom approximately 20%, and AMD approximately 9%.

• The total value of AI computing wafers consumed in 2026 is approximately US$27.2 billion, a historical peak.

TSMC 's AI chip revenue is projected to account for 60% of its total revenue from 2024 to 2029, with AI revenue expected to exceed 30% of total revenue by 2026.

[Transmission Pathway]

Cloud vendors' Capex → NVIDIA/Broadcom/Google TPU orders → CoWoS/SoIC become bottlenecks → TSMC's bargaining power increases → AI revenue continues to expand.

[Trading Insights]

TSMC is the main theme within the main theme, requiring no market timing and having a clear holding logic. SoIC is the second growth curve starting in 2025, so pay attention to opportunities in OSAT suppliers (ASE, etc.) that are entering the SoIC assembly field.

3.2 Testing Equipment (Handler / Socket / Probe Card) — Lowest Valuation, Most Certain Growth

[Core Contradiction]

As chip complexity increases, testing time has structurally doubled, but the market's revaluation of testing equipment (TAM) is severely lagging.

[Key Drivers]

The testing time for each generation of GPU chips has doubled (Hopper 350 seconds → Blackwell 700-1000 seconds → Rubin 1200-1400 seconds → next generation 1800-2000 seconds); the number of test socket pins has jumped from 1500 for mobile phones to 6000 for AI/HPC, and even 10000+ for the next generation.

Data from the three core targets:

Global Handler Market Size: $436 million in 2023 → $6.6 billion in 2027, CAGR of 35%+

• Demand for CPO optical testing will increase significantly starting in 2025, and will enter the stage of combined electrical and optical testing (Insertion 4i) in 2027.

[Transmission Pathway]

Increased chip size/layer count/complexity → Increased testing time → Increased volume and price of Handler/Socket → New demand for CPO optical testing → Start of the second growth curve.

[Trading Insights]

These three companies represent the lowest-valued and most certain-of-growth sub-sectors within the AI ​​infrastructure chain, making them suitable for core mid-term allocation. Their limited market coverage and relatively low pricing make them the most worthwhile value-for-money sector to focus on at present.

3.3 China's AI Chips (Domestic GPUs/ASICs) — Long-term irreversible, short-term differentiation obvious

[Core Contradiction]

Export controls are driving demand for domestic substitution, but the maturity of domestic chip technology and mass production varies greatly; the ability to secure orders from major customers is the key differentiator.

[Key Drivers]

DeepSeek validates the feasibility of low-cost inference → Domestic cloud vendors accelerate their switch → SMIC's 7nm capacity expansion supports mass production → The TCO advantage of domestically produced chips (30-60% lower than NVIDIA) forms a positive feedback loop.

Market size and structure:

2026E domestic market share: Huawei 62%, Cambricon 14%, Kunlun Chip 5%, T-Head 5%, others 14%.

Among the "Ten Dragons," MS focuses on a comparison of three key stocks:

[Transmission Pathway]

Export controls → Domestic substitution → SMIC 7nm capacity expansion → Huawei/Cambricon ramp-up → Local cloud vendors (ByteDance/Alibaba/Tencent) switch procurement → Inference costs decrease → More applications explode → A new round of computing power demand.

[Trading Insights]

Cambricon offers the strongest certainty and is the preferred target; Tianshu Intelligent Chip offers the greatest potential but is not yet profitable, posing a higher risk. Huawei (unlisted) is the biggest competitor, and its market share growth indirectly puts pressure on other domestic manufacturers, requiring continuous monitoring. Time window: 2026–2027 is a crucial turning point for domestic AI chips, transitioning from substitutes to mainstays.

3.4 Non-AI Semiconductors (Consumer/Automotive/Industrial Control) — Structurally bearish; weak recovery is not a strong recovery.

[Core Contradiction]

Supply chain resources are being systematically drained by AI, the recovery of traditional semiconductors continues to be slower than expected, and the market has overestimated the rebound potential.

[Key Drivers]

Manufacturing capacity, T-Glass substrates, and memory are all shifting towards AI; non-AI chips are lagging behind in the supply chain, and wafer and OSAT costs are rising; chip design companies are under pressure on their gross margins.

Excluding NVIDIA AI GPUs and storage, the growth rate of non-AI semiconductors is expected to decline significantly in 2026.

• MCU inventory days remain at historically high levels (peak in Q1 2025, flat in Q4 2025); major manufacturers such as STM/GD are experiencing slow inventory digestion.

• Logic foundry utilization is expected to recover to 80% only in the second half of 2026, with limited upside potential.

SiC is superior to GaN: SICC (OW) is recommended, with SiC penetration expected to exceed 50% by 2030; InnoScience (EW) should be avoided, as expansion and depreciation will suppress profits.

[Trading Insights]

Avoid exposure to purely traditional semiconductors. The MCU sector has bottomed out but is experiencing a weak recovery; heavy betting on a strong rebound is not recommended. SiC is the only traditional sub-sector worth paying attention to.

3.5 Storage (HBM / NAND / DDR4) — Significant internal differentiation, signal identification required.

[Core Contradiction]

AI is clearly driving a surge in HBM demand; however, the price increase of DDR4/NAND is due to supply being squeezed by AI rather than a genuine recovery in demand, resulting in distorted signals and limited price elasticity.

[Trading Insights]

HBM remains bullish, with Hynix being the biggest beneficiary; Macronix (NOR Flash, Top Pick) benefits from shortages and has a reasonable valuation; however, rising NAND/DDR4 prices do not necessarily indicate improved demand, so be wary of chasing the rally.

IV. Macroeconomic and Geopolitical Variables: Explanatory Variables for Track Judgment

[Geographical] Export controls continue to tighten

NVIDIA's exports to China are restricted → Demand for domestically produced AI chips to replace them is rising; China's cloud capital expenditure will reach $105 billion in 2026E, rapidly approaching 14% of global cloud capital expenditure.

[Macroeconomic] Energy Constraints (US Side)

Tight power supplies at U.S. data centers pose a potential ceiling to GPU demand growth, but it has not yet become a substantial constraint in the short term (2026).

[Industry Structure] AI's Encroaching Effect

The siphon effect of AI demand on the non-AI supply chain (T-Glass, traditional DRAM, consumer foundry capacity) is the core explanatory variable for the continued weak performance of non-AI semiconductors compared to cyclical factors.

[Cost Side] Technology Inflation

The rising costs across wafer fabrication, OSAT (Outsourced Automation and Testing), and storage technologies are putting pressure on the gross margins of chip design companies (especially those in non-AI sectors); foundries like TSMC are gaining increasing bargaining power.

V. Recommended Portfolio and Trading Framework

Based on assessments of various sectors, the following trading framework is constructed:

VI. Summary in one sentence

Buy packaging (TSMC), testing equipment (Hon Precision / WinWay / MPI), and China's leading AI chip manufacturer ( Cambricon ); avoid the strong recovery expectations for non-AI semiconductors, focus on HBM in the memory sector, and remain neutral on traditional DRAM/NAND. The time window is 2026-2027, and the AI ​​capital expenditure cycle is far from over.

Risk Warning: This note is compiled based on a publicly available research report by Morgan Stanley and is for internal research reference only. It does not constitute any investment advice. Market uncertainties exist, and actual results may differ significantly from forecasts. Investors are advised to make decisions with caution.

Building the Future AI Infrastructure: CPUs, GPUs, ASICs, Optical Modules, and Chinese Chips

Strong Outlook for Artificial Intelligence Semiconductors

Morgan Stanley characterizes the AI ​​semiconductor outlook as "Strong," driven by three forces: the continued explosion of killer AI applications, the computing power arms race among tech giants, and the sovereign AI development needs of various countries. Meanwhile, this report identifies four growth constraints—budget, US energy bottlenecks, Chinese chip production capacity, and regulation—these constraints are essentially due to supply falling short of demand, rather than demand itself faltering.

In the long run, there are three structural variables that warrant our attention:

1) Technological inflation (rising costs in wafer fabrication/packaging/memory testing squeeze chip design company profits);

2) AI cannibalization effect (supply chain resources are tilted towards AI, and non-AI semiconductors are marginalized);

3) The DeepSeek effect (low-cost inference has been validated, the demand for inference in China is accelerating, and the production capacity of AI GPUs in the domestic OEM supply chain is improving simultaneously). These three factors combined form the underlying logical framework for all subsequent sector assessments in the report.

Valuation Comparison: Foundry, Back-end, Storage, IDM (Integrated Device Manufacturer), and Semiconductor Equipment

Valuation comparison: Fabless, power semiconductors, FPGAs and analog chips

Semiconductor Big Cycle

The core conclusion is cyclical differentiation rather than overall recovery: Logic foundry utilization is expected to rebound to 80% in 2H26, but the growth rate of non-AI semiconductors, excluding NVIDIA AI GPUs and memory, is expected to decline significantly in 2026; the decline in inventory days from the peak is a positive signal. Historical data shows that inventory decline cycles often correspond to semiconductor stock index rises, but the degree of structural differentiation in this round of recovery is far greater than in the past.

Artificial intelligence semiconductor supply chain and niche memory

By 2030, the global semiconductor industry market size may reach $1.5 trillion, half of which will come from AI semiconductors.

Important long-term anchor: The global semiconductor market is expected to reach $1.5 trillion by 2030, of which AI semiconductors will contribute about $753 billion; the bull market scenario for cloud AI semiconductor TAM assumes that it will reach $235 billion in 2025 (mainly from NVIDIA AI GPUs), with a CAGR of 38% from 2023 to 2030, providing a top-level market space basis for the valuation of all subsequent sectors.

Cloud-based semiconductors: A brighter future

The four major cloud vendors (AWS/Google/Microsoft/Meta) saw their capital expenditures increase by 95% year-on-year in Q1 2026, making it the strongest single data point on the demand side in the entire report; the Capex/EBITDA ratio is expected to remain stable at around 50%, indicating that the expansion intentions of cloud vendors are financially sustainable; Aspeed's profit forecast continues to be revised upwards, and as the leading BMC chip provider for cloud AI servers, its revision trend confirms the authenticity of cloud demand.

Major cloud service providers maintain strong cloud capital expenditures

MS Cloud Capex Tracker predicts that the global top 10 cloud vendors' capital expenditures will reach $685 billion in 2026, about 10% higher than the market consensus. The historical chart showing that global cloud Capex and TSMC capital expenditures have climbed in close tandem is the core visual evidence supporting the judgment that "this round is not a short cycle". With short-lifecycle assets accounting for about 65%, it means that cloud vendors must continue to purchase them every year, and the demand is rigid.

TSMC has announced the impact of power deployment.

By analyzing the rack specifications and deployment power of four major customers—NVIDIA, AMD, Broadcom, and AWS—the CoWoS wafer demand is estimated from the bottom up. The NVIDIA Rubin NVL144 rack has a power of 220kW and 45k racks, implying an annual CoWoS demand of 136k wafers in 2027, which is the core quantitative basis for the judgment of tight CoWoS supply and demand in the entire article.

Given the continued strong demand for AI, TSMC may expand its CoWoS capacity to 165,000 wafers per month by 2027.

Here are the supply-side data for CoWoS: TSMC's capacity will expand from 120kwpm at the end of 2025 to 165kwpm at the end of 2027, while Non-TSMC (Amkor/UMC/ASE) capacity will expand from 23kwpm to 80kwpm. On the consumer side, NVIDIA accounts for about 59% of the total consumption of CoWoS, and Broadcom accounts for about 20%. This high concentration means that changes in the demand of a few customers have a huge impact on TSMC.

Expanding SOIC (System-on-a-Chip) will be a key focus for TSMC in the coming years.

SoIC has been identified as a key strategic direction for TSMC in the coming years: capacity will expand from 45kwpm at the end of 2025 to 78kwpm at the end of 2027, with demand from NVIDIA, AMD, Apple, and Qualcomm/Broadcom all included; SoIC has higher integration and deeper technological barriers than CoWoS, and is TSMC's second growth curve in advanced packaging after CoWoS, entering a period of rapid volume expansion in 2026-2027.

TSMC may double its CoWoS and SoIC capacity by 2025, and we expect this trend to continue into 2026.

AI computing wafer consumption may reach $27.2 billion in 2026, with Nvidia accounting for the majority.

The data is presented from bottom to top, listing the CoWoS capacity allocation, chip shipments, wafer consumption, and wafer value for all major AI chips (NVIDIA B300/Rubin/H200, Google TPU, AWS Trainium3, Microsoft Maia, and OpenAI Nexus) in 2026. The total value of AI chip wafer consumption in 2026 is estimated at approximately $27.2 billion, with NVIDIA dominating. This is the most convincing underlying calculation of TSMC's AI revenue scale in the entire article.

HBM (High Bandwidth Memory) consumption in 2026 – a staggering 32 billion Gb

In 2026, the total demand for HBM is approximately 32,279 mn Gb, with NVIDIA consuming about 58%. The HBM specifications (capacity, generation, and supplier) of each AI chip are listed below. Google's TPU series mainly consumes HBM3e 12hi, while AWS/Microsoft consumes HBM3/HBM4. Hynix, Samsung, and Micron share the supply, with Hynix benefiting the most due to its leading HBM technology.

NVIDIA GB200/300 rack production estimates

NVIDIA GB200/300 server rack supply and demand assumptions

TSMC's AI semiconductor revenue is expected to account for 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. The revenue structure covers four segments: general-purpose AI chips, custom ASICs, CoWoS packaging and testing, and AI server CPUs. Apple accounts for 19% of the customer base, NVIDIA for 21%, and Broadcom for 11%. The continued expansion of gross margin and EBITDA margin confirms the positive impact of the AI ​​business on TSMC's overall profitability.

TSMC's advanced wafer demand segmentation

Agentic AI – Expanding CPU Opportunities

AI is moving from the reasoning stage to the "action" stage, with the CPU/GPU ratio shifting from GPU-heavy (1:12) to CPU-heavy (≥1:1). The driving force is tool-type tasks such as API calls, code execution, and multi-agent concurrency. MS estimates that Agentic AI could add $32.5-60 billion to the CPU market (by 2030), and MediaTek, as an AI server CPU designer, is the beneficiary mentioned in the report.

AI storage is causing a NAND shortage; we expect the NOR Flash supply shortage to continue until 2026.

The DDR4 shortage will continue until the second half of 2026; and spot prices have an upper limit.

AI ASIC, CPO and chip testing

AI Semiconductors: Present and Future – Key Drivers

The report presents AI semiconductors across four dimensions: drivers, constraints, technological solutions, and growth perspectives. It also includes three sets of growth perspective comparisons: inference vs. training, edge vs. cloud, and custom ASIC vs. AI GPU. These three sets of comparisons serve as a mind map for understanding the points of divergence in all subsequent tracks of the report.

Even with NVIDIA's powerful AI GPUs, cloud service providers (CSPs) still require custom chips.

According to various cloud service providers (CSPs), more ASIC projects are on the horizon.

How does the competition between TSMC's CoWoS and Intel's EMIB fare?

Larger package sizes are becoming a key industry trend.

Chip testing time has surged from 350 seconds for Hopper to 1800-2000 seconds for next-generation GPUs, representing the most critical structural driver in the testing equipment sector. The number of pins in test sockets has jumped from 1500 for mobile phones/PCs to 6000 for AI/HPCs and even 10000+ for the next generation. The global testing equipment market is projected to achieve a CAGR of 35% from 2024 to 2027, and TSMC's package size roadmap also shows a continuous expansion of interposers. Both factors support the long-term positive outlook for the testing equipment market.

Describing the roles and responsibilities of Hon Hai Precision Industry, WinWay Technology, and MPI in the semiconductor supply chain.

New Evolution of Test Equipment and Components: Co-packaged Optics (CPO)

HTC: A key winner benefiting from the structural trend of extended testing time; Morgan Stanley rating: Overweight (OW)

MPI: A leader in probe card technology with CPO options; Morgan Stanley rating: Overweight (OW)

Yingwei Technology: A leading test socket manufacturer leveraging its advantage in AI packaging complexity; Rating: Overweight (OW)

China's semiconductor industry: OSAT, compound semiconductors, MCUs, and AI GPUs

We are optimistic about back-end equipment (ASMP), but hold a neutral view on Chinese OSAT.

We are bullish on SiC (Silicon Carbide) over GaN (Gallium Nitride): SICC (Overweight) and InnoScience (Underweight).

MCU: Bottomed out but not yet recovered

The market size and share of domestically produced AI semiconductors continue to grow.

The landscape of China's domestic AI accelerator market is clear: Huawei dominates with 62%, Cambricon with 14%, and the rest of the players are all below 10%. The market value of Chinese AI GPU companies continues to grow and more IPOs are pending. The expansion of market size and the increase in capital market activity are the background for the subsequent analysis of key targets.

We project that the total accessible market (TAM) for AI GPUs in China will grow to $67 billion by 2030.

China expands advanced process technology capacity to meet domestic AI GPU production needs.

Recent Market Tracking of AI GPU Demand in China

AI Chip Value Chain – China and the United States – Decoupling in AI Computing

China's infrastructure capabilities are narrowing the perception technology gap.

Using radar charts, the gap in AI infrastructure capabilities between China and the US is compared across nine dimensions: China scores close to the US in policy support, AI data center space, and software optimization (LLM), while the main gaps are concentrated in wafer front-end, HBM memory, and optical networks; a three-step strategy is proposed for China to make up for the lack of single-chip computing power: multi-die packaging → larger racks and clusters → expanding manufacturing capacity. Huawei CloudMatrix 384 A3 SuperPod is a real-world validation of this strategy.

Inference Economics: Total Cost of Ownership (TCO) and Cost per Token

The total cost of ownership (TCO) of domestically produced AI chips is 30-60% lower than that of NVIDIA, and the inference cost per token of top-tier domestic accelerators can be on par with or even better than that of NVIDIA. This conclusion is the core evidence that "domestic substitution in China is not only a political necessity, but also an economic rationale," and directly supports the report's long-term bullish view on the Chinese AI chip industry.

Order placement status and potential orders of domestic AI accelerator developers

TPS (Tokens Output Per Second) - Performance Analysis

Due to the significant price reduction, domestically produced chips have achieved stronger performance per dollar.

The "Ten Dragons" of Chinese AI GPGPU manufacturers. We are particularly focused on Cambricon, Muxi, and Tianshu Zhixin.

Comparison of Cambricon, Muxi, and Iluvatar

A horizontal comparison of three of the most watched Chinese AI chip companies: Cambricon (SMIC 7nm ASIC, locked in major customers, the only profitable one), MetaX (SMIC 12nm GPGPU, held by sovereign wealth funds, with a significant technological gap), and Iluvatar (TSMC 7nm GPGPU, with a strong supply chain resilience). Considering profitability, customer structure, and process node, the report implicitly concludes that Cambricon has the strongest certainty.

Cambricon: Leading inference performance (TFLOPS) and customer loyalty; Overweight rating (OW)

Iluvatar: Leveraging strong order visibility and supply chain resilience; Overweight rating (OW)