Morgan Stanley: AI-Driven Long Bull Market in Semiconductors; TSMC and Chinese Computing Chain Benefit

Overview

This research report systematically reviews the latest trends in the global and Greater China AI cloud semiconductor industry chain. The core conclusion is that AI demand will continue to drive structural growth in the semiconductor industry, with TSMC holding strong bargaining power as the core manufacturing link; meanwhile, China's AI computing infrastructure is accelerating domestic substitution, forming a computing ecosystem independent of the US. The report provides detailed forecasts for cloud capex, advanced packaging capacity, storage supply/demand, and the China AI chip market size, and updates ratings and investment themes for key companies.

Core views

The global AI semiconductor market has entered a long-term expansion cycle. The report predicts the global AI semiconductor Total Addressable Market (TAM) will reach approximately $753 billion by 2030, accounting for half of the then-$1.5 trillion global semiconductor market. The core driver of this growth is the continued aggressive capital expenditure by cloud service providers (CSPs): the top four CSPs (Amazon, Google, Microsoft, Meta) saw their capex increase by 87% YoY in Q2 2026, with capex-to-EBITDA ratios exceeding 70%. Based on optimistic assumptions driven by supply chain data, the cloud AI semiconductor TAM in 2026 can reach $485 billion. Logic chips are expected to account for about 20% of global cloud capex, and power consumption growth remains robust. TSMC's monopoly position in advanced nodes and packaging gives it significant pricing power. Due to its leading technology roadmap and tight EUV equipment supply, TSMC is the 'only choice' for high-end AI chip manufacturing. The report expects TSMC to be able to raise advanced node prices by 5-10% in 2027 to reflect its value to customers. In terms of capacity, given sustained strong AI demand, TSMC may expand CoWoS capacity to 200k wafers per month (kwpm) by 2027. Demand for 2nm process is very strong, with an expected CAGR of 70% in capacity from 2026-2028. Financially, the proportion of AI semiconductor revenue in TSMC's total revenue is expected to rise from 2024 levels to 60% by 2029, with this ratio expected to exceed 30% in 2026. Additionally, TSMC's capex in 2027 is expected to reach $75 billion, with a gross margin of 60% being considered the floor. China's AI computing ecosystem is maturing rapidly, forming a unique 'decoupled' development path. The report expects China's AI GPU market size to reach $91 billion by 2030, and the CPU market size to reach $42 billion. Despite technological gaps, domestic chips achieve better 'performance per dollar' and inference economics (TCO and cost per token are superior to NVIDIA's products sold to China) thanks to lower prices. The emergence of low-cost inference models like DeepSeek further stimulates domestic demand for inference-side AI. Regarding specific names, Hygon receives an Overweight (OW) rating for providing an integrated CPU+GPU computing platform, with its share in China's server CPU market expected to reach 18% by 2028. Domestic cluster solutions such as Huawei Atlas 950 SuperPod also demonstrated thousand-card-level scalability at WAIC 2026. Storage and advanced packaging have become key bottlenecks and growth points in the AI industry chain. AI's massive consumption of storage has led to NAND shortages, and NOR Flash is expected to be in short supply until 2026, with DDR4 shortages potentially continuing until H2 2026. HBM consumption is expected to reach 48 billion Gb in 2027, with NVIDIA still occupying the vast majority of the supply. On the packaging side, besides TSMC dominating CoWoS, Intel's EMIB technology, if executed smoothly within the supply chain, can support larger chips. Back-end equipment maker ASMPT benefits from advanced packaging expansion, with its expected market share in TCB equipment reaching about 44% by 2028; among mainland Chinese OSAT firms, KYEC is favored due to strong growth in AI GPU/TPU/CPU testing business, while ASE is expected to double its LEAP (AI) revenue by 2027. Custom ASICs and Agentic CPUs constitute new sources of incremental growth. Even though NVIDIA GPUs are powerful, CSPs still need custom chips to optimize costs and efficiency. Google TPU has entered its sixth generation, with projects like AWS Trainium and Meta MTIA progressing continuously. MS Asia team predicts rapid growth in TPU shipments, although volume release may be limited by ABF substrate supply. Additionally, NVIDIA's latest Vera CPU is designed specifically for Agentic AI, with faster inter-core connections. In an optimistic scenario, it could support a $238 billion CPU orchestration TAM, with the Agentic CPU TAM expected to grow at a CAGR of 251% from 2026-2030.

Analysis framework

The report adopts a typical 'top-down TAM modeling + bottom-up supply chain verification' analytical framework. First, it establishes a cloud AI semiconductor total demand model by tracking the capex guidance and financial report data of major global cloud service providers (CSPs); then it breaks this down into physical indicators such as wafer consumption, CoWoS/HBM demand, and cross-validates them with capacity plans from suppliers like TSMC, thereby judging supply-demand gaps and price trends. For the Chinese market, the report introduces a comparative analysis method of 'inference economics', not simply comparing peak chip compute power, but combining price, power consumption, and cluster interconnect efficiency to calculate TCO and cost per token for inference, thereby assessing the true competitiveness of domestic chips. Meanwhile, through field research at industry exhibitions like WAIC, it verifies the engineering implementation progress of domestic SuperPods, compensating for the lag in public financial data.

Methodology notes

• Industry/Industrial Analysis FrameworkSupply and Demand Framework

  Linkage Analysis of Cloud Capex and Semiconductor TAM

  Using downstream cloud vendors' capital expenditure (Capex) as a leading indicator for upstream semiconductor demand, deriving AI chip TAM through Capex-to-EBITDA ratios and historical elasticity coefficients. This method assumes that cloud vendors' investments will directly translate into chip purchase orders, which is the core logic for predicting semiconductor prosperity.

• Competition and Strategy FrameworkMoat / competitive advantage

  Technological Monopoly Premium of Advanced Nodes and Packaging

  The report emphasizes that TSMC has become the 'Only Game in Town' due to EUV equipment restrictions and leading technological density. This rigid constraint on the supply side endows the leading enterprise with the ability to raise prices (5-10%) when demand is strong, which is a key moat indicator for judging its profitability resilience.

• Industry/Industrial Analysis FrameworkVolume and Price Breakdown

  Physical Consumption Measurement of AI Storage and Packaging

  Not just looking at monetary values, but breaking down AI demand into specific physical units (e.g., CoWoS thousands of wafers per month, HBM billions of Gb, NVL72 rack counts). This physical volume analysis eliminates the interference of price fluctuations, more accurately identifying real bottlenecks in the industry chain (such as ABF substrates, CoWoS capacity) and investment opportunities.

• Company Fundamentals and Financial FrameworkProfitability Quality Analysis

  Inference Economics (TCO and Cost per Token)

  When evaluating the competitiveness of China's AI chips, the report does not adopt simple theoretical compute comparisons, but uses total lifecycle cost (TCO) and unit inference output cost as measurement standards. This reflects the industry common sense that in the AI commercialization stage, customers pay more attention to actual deployment benefits rather than paper parameters.

Asset mapping & comparison

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

• TSMC (Taiwan Semiconductor Manufacturing Company)

  Core monopolist in AI semiconductor manufacturing and packaging, directly benefiting from global AI Capex expansion

  Strengths

  Leading technology roadmap; uniqueness under tight EUV capacity; strong demand for 2nm; rapidly increasing AI revenue share

  Weaknesses

  Geopolitical risks; high costs of overseas factory construction

  Comparison

  Maintains generational lead in logic density and yield compared to Intel/Samsung

  Risks

  Global macroeconomic slowdown leading to lower-than-expected cloud Capex; power supply constraints limiting capacity release

• MediaTek (MediaTek Inc.)

  Core designer of AI ASICs and edge AI chips, listed as a Top Pick

  Strengths

  Participates in Google TPU v8t/v9/v10 design; rich reserve of custom ASIC projects

  Weaknesses

  High dependence on major customers

  Comparison

  Competes with Alchip/GUC in AI ASIC design services, but with stronger economies of scale

  Risks

  ABF substrate supply bottlenecks limiting shipments; rising proportion of client self-research

• Hygon Information Technology

  Leader in China's domestic x86 CPU+GPU platform, rated Overweight (OW)

  Strengths

  Integrated computing platform aligns with domestic IT application innovation and AI needs; server CPU share expected to reach 18% by 2028

  Weaknesses

  Limited access to advanced nodes; ecosystem maturity weaker than international giants

  Comparison

  Leading in commercial implementation and revenue scale compared to Cambricon/MetaX

  Risks

  Escalation of entity list sanctions; intensifying domestic competition

• KYEC (Kyunghwa Electronics Corp.)

  Core beneficiary of AI chip testing, covering GPU/TPU/CPU

  Strengths

  Rapid growth in AI testing revenue; deeply bound with NVIDIA/AMD, etc.

  Weaknesses

  High capital expenditure pressure

  Comparison

  Higher AI exposure than mainland Chinese OSAT peers

  Risks

  Fluctuation in AI chip demand; risk of testing technology iteration

• ASMPT

  Leader in advanced packaging equipment, benefiting from CoWoS/TCB expansion

  Strengths

  Expected TCB equipment market share to reach 44% by 2028; leading layout in CPO assembly equipment

  Weaknesses

  Cyclical drag from traditional packaging business

  Comparison

  Superior to Chinese OSAT equipment makers in AI packaging equipment

  Risks

  Risk of TSMC replacing with self-developed equipment; slowdown in downstream expansion pace

Key data

• 2030 Global AI Semiconductor TAM~$753 billionExpected to contribute half of the $1.5 trillion global semiconductor market
• 2027 Global Cloud CapexNear $1.4 trillionCovering top 14 listed global CSPs, excluding sovereign AI
• Top 4 CSP 2Q26 Capex Growth+87% YoYCapex/EBITDA ratio exceeds 70%
• TSMC 2027 CoWoS Capacity Target200 kwpmTo meet sustained strong AI demand
• TSMC 2027 Advanced Node Price Increase5-10%Reflecting technological value and customer dependency
• 2030 China AI GPU TAM$91 billionDomestic share continues to rise
• 2027 HBM Consumption48 billion GbNVIDIA occupies most of the supply
• Agentic CPU TAM CAGR (2026-2030)251%Driven by new architectures like NVIDIA Vera

Impact & implications

For the semiconductor industry chain, AI is not just a short-term hotspot, but a long-term structural force reshaping the industry landscape. TSMC and its advanced packaging supply chain (such as ASMPT, KYEC, ASE) will be the most certain beneficiaries, with their earnings visibility extending to 2027-2029. The storage sector faces structural shortages due to AI displacing capacity, and related manufacturers are likely to see both volume and price increases. For the Chinese market, the 'decoupling' of AI computing power has instead spawned a massive internal circulation market. Domestic chips' cost-performance advantage on the inference side allows them to gain substantial shares in bidding by internet giants and operators. Leading enterprises like Hygon and Cambricon are moving from 'usable' to 'good'. Investors should focus on the actual deployment progress of domestic SuperPod clusters and breakthroughs in autonomous processes/packaging links, which are key signals to verify the revaluation of China's AI industry chain value.

Risks

• Global cloud service provider capex falls short of expectations due to budget or energy constraints
• Slow commercialization of AI applications leads to demand disproof
• Escalation of geopolitical tensions and export controls impacts China's AI supply chain
• Expansion of advanced node and packaging capacity is hindered (e.g., power supply, equipment delivery delays)
• Storage and substrate supply bottlenecks constrain AI chip shipments
• Weak non-AI semiconductor demand drags down overall industry recovery

What to watch

• Quarterly capex guidance and actual execution of major global CSPs
• TSMC monthly revenue and progress of CoWoS/2nm capacity ramp-up
• Token call volumes of China's AI large models and winning bids for domestic chips
• Shipment rhythm of NVIDIA GB200/300 NVL72 racks
• Spot price and contract price trends for storage
• Iteration of domestic SuperPod technical solutions at industry exhibitions like WAIC
• Mass production progress of custom ASICs such as Google TPU/AWS Trainium