Strong outlook for AI semiconductors, with cloud capex, advanced packaging, and China AI GPUs as the main themes

Overview

This is a Greater China and Asia-Pacific semiconductor industry outlook report. Its core theme is how future AI infrastructure buildout will drive CPUs, GPUs, ASICs, optical communications, advanced packaging, HBM, and China's domestic AI chips. Through perspectives including global cloud capex, NVIDIA platform upgrades, TSMC advanced packaging and 2nm demand, China AI GPU TAM, domestic advanced-process capacity, back-end equipment, and power semiconductors, the report argues that AI semiconductors remain in a phase of high-growth expansion.

Core views

The core views include: first, the global semiconductor market may reach US$1.5 trillion by 2030, with AI semiconductors contributing about half of the incremental growth; second, major CSP cloud capex continues to support demand for AI GPUs, ASICs, and the data center supply chain; third, CoWoS, SoIC, HBM, T-Glass, and advanced-process capacity remain key constraints on AI compute expansion; fourth, even though NVIDIA GPUs are highly capable, CSPs still need self-developed or customized chips to optimize cost, power consumption, and workloads; fifth, China's AI inference demand, the low-cost inference model demonstrated by DeepSeek, and improving domestic supply-chain capabilities will drive expansion of China's AI GPU and advanced-process ecosystem.

Analysis framework

The report uses top-down cloud capex and TAM estimates, combined with bottom-up Asian supply-chain checks, assumptions for NVIDIA GB200/GB300 rack output, tracking of TSMC CoWoS and SoIC capacity, HBM consumption, wafer consumption, customer demand breakdowns, and China's advanced-process capacity constraints, to form supply-demand judgments across AI semiconductor segments.

Methodology notes

• 市场规模预测TAM and long-term demand driver framework

  Use the 2030e global semiconductor market, cloud AI semiconductor TAM, and China AI GPU TAM to measure the industry's upside potential.

  The report incorporates AI training, inference, cloud and edge demand, CSP custom chips, sovereign AI, and application diffusion into its long-term demand assumptions, and uses TAM figures to express potential market size.

• 供应链验证Bottom-up supply chain checks

  Validate AI semiconductor demand through wafer consumption, HBM consumption, rack shipments, CoWoS/SoIC capacity, and customer demand breakdowns.

  The report frequently cites Morgan Stanley Research estimates, company data, FactSet, and Asian supply-chain checks to translate cloud capex into demand for wafers, packaging, memory, and server racks.

• 瓶颈约束分析Capacity, cost, energy, and regulatory constraint framework

  Treat budgets, energy, chip capacity, regulation, wafer and OSAT costs, and HBM and T-Glass shortages as limiting factors on growth.

  The report is not simply bullish on AI demand; it also emphasizes that chip inflation will compress the margins of design companies, non-AI semiconductors may be crowded out by AI semiconductors, and China's advanced-process constraints and U.S. energy constraints will affect the pace of deployment.

Asset mapping & comparison

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

• AI GPU/CPU/ASIC supply chain

  Core compute assets of AI infrastructure

  Strengths

  Growth is supported by sustained cloud capex, NVIDIA platform upgrades, CSP custom-chip demand, and inference demand.

  Weaknesses

  Highly dependent on supply of advanced processes, CoWoS/SoIC, HBM, and high-end packaging materials.

  Comparison

  NVIDIA GPUs remain the dominant direction, but CSP self-developed ASICs are still necessary for cost, power, and specific workloads.

  Risks

  A slowdown in cloud capex, weaker-than-expected application monetization, rising chip costs, and regulatory restrictions.

• TSMC and the advanced packaging chain

  Key manufacturing and packaging bottleneck in AI chip capacity expansion

  Strengths

  The report expects TSMC CoWoS capacity may expand to 200kwpm by 2027e, while 2nm customer demand and the trend toward larger packaging support long-term demand.

  Weaknesses

  Capacity expansion pace, equipment delivery, materials, and yield may constrain near-term supply.

  Comparison

  Compared with standard foundry services, advanced packaging and leading-edge processes benefit more directly from rising AI accelerator complexity.

  Risks

  Customer demand volatility, overheated capex, technology transitions, and supply-chain bottlenecks.

• HBM and the memory chain

  Key supporting assets for AI servers and AI accelerators

  Strengths

  HBM consumption could reach as high as 32bn Gb in 2026e, and rising AI priority may support memory demand and pricing.

  Weaknesses

  Tight supply-demand conditions may raise costs and compress downstream chip design company margins.

  Comparison

  HBM is more leveraged to AI training and high-performance inference demand than traditional memory.

  Risks

  Supply being released too quickly, a slowdown in AI demand, and weak non-AI end-market demand.

• China AI GPUs and domestic advanced processes

  Core assets for China's AI infrastructure and import substitution

  Strengths

  After DeepSeek demonstrated lower-cost inference, inference demand was further stimulated; the report expects China AI GPU TAM to reach US$91bn by 2030e.

  Weaknesses

  Advanced-node capacity, EDA/IP, packaging, yield, and export controls remain constraints.

  Comparison

  Compared with globally leading AI GPUs, China's domestic chips place more emphasis on availability, supply security, and local ecosystem adaptation.

  Risks

  Escalating regulation, insufficient capacity, performance gaps, and slower-than-expected demand realization.

• Back-end equipment, test consumables, and OSAT

  Supporting beneficiaries of rising advanced packaging complexity

  Strengths

  The report is positive on back-end equipment and test consumables such as ASMPT and Winway, and expects OSAT capex to grow about 35% in 2026e.

  Weaknesses

  Competition in traditional OSAT remains intense, and some segments may be affected by cyclicality.

  Comparison

  Compared with ordinary packaging and testing, higher AI packaging complexity brings greater value per unit for equipment, sockets, and consumables.

  Risks

  Packaging expansion falling short of expectations, high customer concentration, and price competition.

• SiC/GaN power semiconductors

  Assets related to AI data center power infrastructure

  Strengths

  The report prefers SiC over GaN, and mentions SICC as OW and InnoScience as EW.

  Weaknesses

  Application scenarios, cost curves, and customer adoption timing differ across material paths.

  Comparison

  In the report, SiC is relatively more preferred than GaN, while GaN HVDC 800V is still listed as one of the semiconductor solutions.

  Risks

  Technology-path shifts, price declines, and slower-than-expected evolution of data center power architectures.

Key data

• Global semiconductor market sizeabout US$1.5 trillion in 2030eThe report says about half of this may come from AI semiconductors.
• Cloud AI semiconductor TAM, bull-case scenarioabout US$235bn in 2025eBased on a bull-case assumption driven by supply-chain data.
• TSMC CoWoS capacitymay expand to 200kwpm in 2027eSustained strong AI demand is driving advanced packaging capacity expansion.
• AI compute wafer consumptionabout US$27bn in 2026eThe report says NVIDIA accounts for the majority share.
• HBM consumptionup to 32bn Gb in 2026eHBM is one of the key bottlenecks in the AI compute supply chain.
• NVIDIA Vera CPU performanceabout 1.8x that of the highest-performance x86 CPUThe report believes the Vera CPU may support agentic CPU demand.
• China AI GPU TAMabout US$91bn in 2030eDriven by inference demand, domestic supply-chain capability, and the local AI chip ecosystem.
• China advanced-process support capacitymay support about US$58bn of AI accelerator revenue in 2030eThe report views domestic advanced-node capacity as a key variable in China's AI GPU production.
• OSAT capexexpected to grow about 35% in 2026eDriven jointly by mainstream semiconductors and advanced packaging.

Impact & implications

The investment implication is that AI infrastructure buildout continues to reshape the semiconductor industry's profit pool: upstream advanced processes, advanced packaging, HBM, CPO, custom ASICs, test consumables, and the China AI GPU supply chain offer stronger structural opportunities; meanwhile, non-AI semiconductors, chip design companies with weaker margins, and segments affected by cost inflation may come under pressure. The report also notes that improving domestic supply-chain capabilities in China will narrow the perceived technology gap, but capacity, regulation, and geopolitical restrictions will still determine the pace.

Risks

• Major CSP cloud capex comes in below expectations, leading to downward revisions in demand for AI GPUs, ASICs, servers, and packaging.
• AI applications, killer apps, or inference monetization underperform expectations, weakening long-term TAM assumptions.
• Budgets, energy, regulation, and chip capacity become growth constraints, especially U.S. energy constraints and China's advanced-process capacity constraints.
• Rising wafer, OSAT, HBM, and packaging material costs create margin pressure for chip design companies.
• Higher priority for AI semiconductors may crowd out demand for non-AI semiconductors, putting pressure on traditional semiconductor value chains.
• Supply bottlenecks in HBM, T-Glass, CoWoS/SoIC, or advanced nodes may affect shipment timing.
• Export controls, geopolitics, and changes in the pace of domestic supply-chain substitution may alter the path of China's AI GPU market.
• The report discloses that Morgan Stanley has business relationships with multiple covered companies, and investors should treat this research as one factor in decision-making rather than the sole basis.

What to watch

• Major CSPs' quarterly cloud capex guidance and changes in AI server orders.
• NVIDIA GB200/GB300 NVL72 rack output, supply-demand gaps, and changes in supply share.
• TSMC CoWoS, SoIC, and 2nm expansion progress, as well as the large-size interposer roadmap.
• HBM supply-demand conditions, pricing, capacity releases, and customer consumption pace.
• Progress of CSP in-house ASIC projects, including cloud AI custom chips such as AWS Trainium.
• China AI GPU orders, IPOs, localization progress, and advanced-node capacity releases.
• Delivery of capex by OSAT, back-end equipment, test sockets, and consumables.
• The marginal impact of energy, regulation, and export controls on AI infrastructure buildout.