AI Semiconductors: A Long-Term Bull Market—Robust Cloud Capital Expenditures and the Rise of China’s Chip Industry

Overview

This report provides an in-depth analysis of the current state and future trends in global AI infrastructure, covering CPUs, GPUs, ASICs, optical modules, and China’s chip industry chain. Morgan Stanley argues that, despite challenges such as technological inflation and the crowding-out effect of AI on non‑AI demand, robust capital expenditures by cloud service providers will continue to drive strong growth in the AI semiconductor market. The report forecasts that by 2030, the global semiconductor market will reach US$1.5 trillion, with AI semiconductors accounting for half of this total. In addition, the report highlights China’s progress toward self-reliance in the AI chip sector, noting that the emergence of models like DeepSeek is spurring demand on the inference side, while the domestic foundry supply chain is steadily acquiring the capability to produce AI GPUs.

Core views

Cloud Capital Expenditure and Robust AI Semiconductor Demand The research report indicates that by 2026, capital expenditures among the world’s top ten publicly listed cloud service providers are projected to reach approximately $685 billion, significantly exceeding earlier forecasts. Notably, the share of capital spending allocated to AI‑related servers has risen markedly. Against this backdrop, the total addressable market (TAM) for AI semiconductors in the cloud is expected to hit $235 billion by 2025. NVIDIA remains the primary beneficiary, commanding the lion’s share of AI‑compute wafer consumption; however, AMD, Broadcom, and proprietary ASICs developed by major cloud players—such as Google’s TPU and AWS’s Trainium—are also expanding rapidly. Addressing Advanced Packaging Capacity Constraints and Technological Advancements To meet surging demand, TSMC is substantially scaling up its CoWoS capacity, with output projected to more than double from current levels to 165,000 wafers per month by 2027. Additionally, SoIC (System on Integrated Chip) is poised to become a key growth driver, with capacity expected to reach 78,000 wafers per month by 2028. The report underscores that as AI evolves from inference toward action‑oriented “agentic AI,” the CPU:GPU ratio at the cluster level will increase, opening new avenues for CPU‑centric growth. Meanwhile, larger package sizes are becoming an industry trend to accommodate greater numbers of compute cores and memory components. Structural Shortages in the Storage Market AI‑driven data storage demands have triggered shortages in NAND flash memory, a situation the report projects will persist through 2026. NOR flash is likewise experiencing supply constraints. On the DRAM front, DDR4 shortages are expected to continue into the second half of 2026, despite potential downward pressure on spot prices. Meanwhile, demand for HBM (High Bandwidth Memory) is forecast to surge to 32 billion Gb by 2026, primarily driven by NVIDIA, Google, and AMD. China’s Ascendancy in AI Chips and Supply Chain Restructuring The report highlights that DeepSeek’s demonstration of lower inference costs has catalyzed domestic demand for inference‑focused AI solutions in China. The Chinese AI GPU market is anticipated to grow at an extraordinary pace, with a TAM reaching $67 billion by 2030. Despite U.S. export controls, China is mitigating gaps in single‑chip performance by ramping up advanced‑node capacity, refining packaging technologies such as chiplets, and building large‑scale clusters. Companies like Cambricon, Hygon Information (Iluvatar), and Moore Threads (MetaX) are positioned as leading domestic players, with Cambricon receiving an overweight rating due to its superior inference performance and strong customer base. Top Picks Across Sub‑Sectors In terms of individual stock selection, the report favors companies with significant AI exposure and robust technological barriers: - Foundry and Design Services: TSMC, MediaTek (top pick), and SMIC are preferred. MediaTek is particularly favored for its potential in edge‑AI and ASIC design services. - Memory: A selective approach is recommended, with Macronix as the top choice, given its attractive valuation and exposure to NOR flash shortages. - China‑Focused AI/Semiconductors: Cambricon, Hygon Information, NAURA Tech, and AMEC are viewed favorably. - Test Equipment and Consumables: Benefiting from extended testing times and increased complexity associated with AI chips, WinWay, MPI, and Hon Precision are preferred. - Power Semiconductors: SiC is favored over GaN, with Tianyue Advanced (SICC) as the preferred stock.

Analysis framework

Supply–Demand Framework and Industry-Chain Transmission Analysis The research report employs a classic semiconductor supply–demand analytical framework, beginning with the capital expenditures (Capex) of downstream cloud service providers (CSPs) to derive, in a top-down manner, the demand for upstream servers, chips (GPUs/ASICs/CPU), and components (HBM/packaging/optical modules). By tracking the budget allocations and technology roadmaps of major CSPs—such as NVIDIA’s GB200/300 rack shipment plans—the report quantifies the specific requirements for TSMC’s CoWoS capacity and HBM. Technology Paths and Competitive Landscape Comparison The report provides a detailed comparison of the strengths and weaknesses of various technological approaches, including the competition between TSMC’s CoWoS and Intel’s EMIB, as well as the cost‑effectiveness of general‑purpose GPUs versus custom ASICs across different application scenarios. Furthermore, by contrasting China and the United States across nine dimensions spanning AI chips, systems, and infrastructure, the analysis examines how China seeks to narrow its single‑point technology gaps through “stacking” strategies—deploying more chips, larger clusters, and greater production capacity. Valuation and Catalysts Combined In its stock‑picking logic, the report integrates relative valuation metrics (P/E, P/B) with business catalysts, such as the progress of capacity expansion, new‑product launches, and the accelerating pace of domestic substitution. For Chinese semiconductor companies, it particularly underscores the additional growth drivers stemming from supply‑chain security and rising localization rates in the context of geopolitical dynamics.

Methodology notes

• Industry/Industrial Analysis FrameworkSupply-and-Demand Framework

  By analyzing the capital expenditures of downstream cloud service providers, we can infer upstream chip demand.

  The research report begins with macro-level cloud‑capital‑expenditure data, progressively disaggregating it to specific demands for AI servers, GPU/ASIC chips, and even wafer‑fabrication and packaging capacity—representing a quintessential top‑down supply‑and‑demand analysis that helps readers grasp the root drivers of industry prosperity.

• Industry/Industrial Analysis FrameworkTransmission across the upstream, midstream, and downstream segments of the industrial chain

  AI Infrastructure Value Chain Transmission

  The report maps out the entire value chain, from cloud service providers (CSPs) to server assembly, and on to chip design, manufacturing, packaging and testing, as well as raw materials, clearly delineating the order of beneficiaries and the degree of elasticity at each stage in the AI-driven boom.

• Competition and Strategic FrameworkMoat / competitive advantage

  TSMC’s technological barriers in advanced packaging

  The research report underscores TSMC’s leading position and near-exclusive capacity in advanced packaging technologies such as CoWoS and SoIC, characterizing these as its core competitive moat and explaining why TSMC enjoys formidable bargaining power during periods of AI chip shortages.

• Cyclical and Business Cycle FrameworkInventory cycle (Kitchin)

  Semiconductor Inventory Levels and Price Elasticity

  The report discusses changes in inventory days across the semiconductor supply chain and the resulting price elasticity—known as “Tech Inflation”—highlighting that, during periods of strong demand, rising wafer and packaging/testing costs can erode chip design firms’ profit margins. This metric serves as a key indicator for assessing the current phase of the industry cycle.

Asset mapping & comparison

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

• Taiwan Semiconductor Manufacturing Company (TSM)

  Key beneficiaries: the undisputed leader in AI chip manufacturing and advanced packaging (CoWoS/SoIC)

  Strengths

  Technologically advanced, with a highly exclusive production capacity and strong pricing power.

  Weaknesses

  Geopolitical risks and substantial capital expenditures

  Comparison

  Compared with Samsung and Intel, it enjoys a distinct advantage in yield rates and customer trust.

  Risks

  The global macroeconomic downturn has led to weak non-AI demand.

• MediaTek (2454.TW)

  Top Pick: Benefiting from the growth of edge AI and ASIC design services.

  Strengths

  A solid foothold in the mobile SoC market, with proactive expansion into AIoT and automotive chips.

  Weaknesses

  The high-end market is highly competitive.

  Comparison

  Compared with Qualcomm, it enjoys a stronger competitive edge in terms of cost-effectiveness and the mid-range market.

  Risks

  The smartphone market recovery has fallen short of expectations.

• Cambricon (688256.SS)

  China’s leading AI chip company: benefiting from the explosive growth in domestic computing power demand

  Strengths

  It boasts superior inference performance and high customer stickiness.

  Weaknesses

  Persistent losses, reliance on government subsidies and ecological conservation projects

  Comparison

  One of the few domestic companies capable of providing end-to-end AI chip solutions

  Risks

  U.S. sanctions are intensifying, heightening the risk of supply-chain disruptions.

• WinWay (6515.TW)

  Test Socket Leaders: Benefiting from the Increasing Complexity of AI Chip Testing

  Strengths

  Vertical Integration Strategy: Synergies Between Probe Cards and Test Socket Adapters

  Weaknesses

  High customer concentration

  Comparison

  It possesses technological barriers in the high-end test interface sector.

  Risks

  Fluctuations in the semiconductor testing capital expenditure cycle

• Tianyue Advanced (SICC, 688234.SS)

  Silicon Carbide Substrates: Benefiting from Demand in Electric Vehicles and Energy Infrastructure

  Strengths

  Market share is expanding, and production capacity is growing rapidly.

  Weaknesses

  Intensified industry competition has led to pricing pressures.

  Comparison

  Compared with international giants such as Wolfspeed, it enjoys cost and delivery advantages.

  Risks

  Risks associated with technological advancements and a slowdown in downstream electric vehicle sales.

Key data

• Global Semiconductor Market Size Forecast for 2030USD 1.5 trillionAmong them, AI semiconductors are expected to account for 50%.
• Global Top 10 CSP Capital Expenditures in 2026~US$685 billionMorgan Stanley’s forecast exceeds the market consensus.
• 2025 Cloud-Based AI Semiconductor TAMUS$235 billionForecasts under a bull market scenario
• TSMC CoWoS Capacity Forecast (2027)165,000 units/monthA substantial expansion from current levels
• Total HBM Demand in 202632 billion GbPrimarily driven by NVIDIA (58%) and Google (24%)
• China’s AI GPU Total Addressable Market (TAM) in 2030US$67 billionReflecting the rapid growth in domestic computing power demand

Impact & implications

For global semiconductor giants, AI is not only a short-term driver of earnings but also the cornerstone of long-term valuation re‑pricing. TSMC, benefiting from its dominant position in advanced packaging, could see AI‑related revenue account for as much as 60% by 2029. Meanwhile, memory manufacturers are likely to maintain robust profitability amid shortages of HBM and high‑end NAND, though they should remain vigilant about cyclical volatility. For China’s semiconductor industry, external restrictions have instead accelerated the maturation of the domestic supply chain—particularly in mature process technologies, packaging and testing, and specialized AI inference chips—enabling Chinese firms to capture greater market share and secure enhanced policy support. Investors should focus on companies that can integrate into the global AI supply chain or assume a leading role in China’s domestic substitution efforts.

Risks

• Techflation: Rising costs for wafers, packaging and testing, and memory could squeeze the profit margins of chip design firms.
• AI-induced crowding-out effect: Prioritizing resource allocation to AI chips may lead to supply shortages or suppressed demand for non-AI chips, such as traditional MCUs and analog chips.
• Geopolitics and Export Controls: The U.S. further tightening of semiconductor export restrictions on China could impede the development of China’s AI chip industry and undermine global supply-chain stability.
• Energy Bottleneck: The high energy consumption of AI data centers may face electricity supply constraints in the United States and other regions, thereby limiting deployment speed.
• Regulatory Risk: Uncertainty surrounding AI-related regulatory policies worldwide.

What to watch

• The actual progress of TSMC’s capacity expansion for CoWoS and SoIC, as well as their utilization rates.
• Changes in AI capital expenditure guidance from major cloud service providers (Microsoft, Amazon, Google, Meta).
• The shipment cadence and yield rates of NVIDIA’s GB200/300 series rack-mounted systems.
• The declining inference costs of China’s homegrown large AI models (such as subsequent versions of DeepSeek) are driving increased demand for hardware.
• The degree of supply constraints and price trends for HBM3e and HBM4.