In-depth semiconductor research for future AI infrastructure: parallel expansion in GPU, ASIC, optical chips, and China chips

Overview

This report is Morgan Stanley’s deep research on the semiconductor industry in Greater China, centered on CPU, GPU, ASIC, optical chips, and China chips for future AI infrastructure. From cloud capex, AI servers, AI semiconductors, advanced packaging, HBM, CPO, test equipment, power semiconductors, and China AI GPU supply chains, the report analyzes how AI infrastructure expansion drives the semiconductor value chain.

Core views

The core view is that AI remains the strongest structural driver for semiconductors. Cloud capex remains robust, and the AI server and AI semiconductor markets continue to expand; while NVIDIA GPUs remain dominant, cloud companies’ in-house ASICs and non-NVIDIA GPUs will also become important incremental contributors. Advanced packaging, CoWoS/SoIC, HBM, CPO, and test consumables benefit from rising packaging complexity. In China, low-cost inference demand driven by DeepSeek, improving domestic supply chain capabilities, and expanding advanced-node capacity are expected to support growth in the local AI GPU market.

Analysis framework

The report uses an industry-chain, data-driven framework: top-down estimation of cloud capex, AI server capex, and AI semiconductor TAM, then bottom-up tracking of TSMC advanced packaging capacity, NVIDIA GB200/300 rack supply-demand, HBM Gb consumption, wafer demand, CSP custom ASIC projects, China advanced-node capacity, and the competitive landscape of domestic AI GPU vendors.

Methodology notes

• Market sizingLayered TAM and capex decomposition

  Break down sequentially from Cloud Capex to AI Server Capex to AI Semi TAM

  The chart shows estimated 2025e cloud capex of about US$400bn, AI server capex of about US$300bn, and AI Semi of about US$235bn, of which cloud AI ASICs and non-NVIDIA GPUs are about US$30bn.

• Supply chain trackingAdvanced packaging and wafer demand tracking

  CoWoS/SoIC, leading-edge wafers, HBM, and rack supply-demand

  By tracking TSMC CoWoS capacity, NVIDIA GB200/300 racks, HBM Gb consumption, and leading-edge wafer demand, the report gauges the constraints on upstream manufacturing and packaging resources from AI infrastructure expansion.

• Competitive landscapeGPU and ASIC substitution/complementarity analysis

  Custom ASIC vs. AI GPU

  The report argues that powerful NVIDIA AI GPUs do not eliminate CSPs’ need for in-house chips; cost, workload optimization, and supply-chain security still drive Trainium, TPU, and more ASIC projects.

• Localization analysisMapping China AI GPU and advanced-node capacity

  Domestic demand, capacity, and vendor landscape

  The report combines DeepSeek-driven inference demand, China’s advanced-node capacity expansion, domestic foundry capabilities, and comparisons among Cambricon, MetaX, and Iluvatar to assess the growth potential of China AI GPUs.

Asset mapping & comparison

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

• AI semiconductor supply chain

  Core beneficiary asset group

  Strengths

  Driven by cloud capex, AI servers, GPUs, ASICs, and HBM demand, with market size expected to expand rapidly.

  Weaknesses

  Technology inflation raises wafer, OSAT, and memory costs, which may compress margins for chip designers.

  Comparison

  Compared with non-AI semiconductors, AI semiconductors have higher priority in supply-chain resource allocation.

  Risks

  Budget constraints, energy limits, capacity bottlenecks, regulation, and export controls.

• TSMC and advanced packaging

  Key manufacturing and packaging segment for AI GPU and ASIC expansion

  Strengths

  CoWoS/SoIC capacity expansion directly benefits from demand for NVIDIA and CSP AI chips.

  Weaknesses

  Capacity build-out takes a long time, and demand volatility may create capex efficiency risk.

  Comparison

  Advanced packaging is closer to the AI compute bottleneck than traditional process nodes.

  Risks

  Uncertainty around equipment, materials, yields, power, and customer demand timing.

• NVIDIA ecosystem

  Main contributor to AI Semi TAM

  Strengths

  AI GPUs remain dominant in training and inference infrastructure, and GB200/300 rack supply-demand is a key tracking indicator.

  Weaknesses

  High costs and supply constraints push CSPs to keep developing in-house ASICs and non-NVIDIA GPUs.

  Comparison

  GPUs are more general-purpose, while ASICs may offer cost and energy-efficiency advantages for specific workloads.

  Risks

  Customer in-house substitution, export controls, and supply-chain bottlenecks.

• CSP in-house ASICs

  An incremental substitute and complement for AI infrastructure

  Strengths

  Can be optimized for internal workloads in terms of cost, energy efficiency, and supply security; Trainium and TPU forecasts are tracked separately.

  Weaknesses

  Long development cycles and high software/ecosystem adaptation difficulty.

  Comparison

  Compared with NVIDIA GPUs, ASICs rely more on a single customer scale and specific use cases.

  Risks

  Project delays, performance shortfalls, and forecast errors.

• China AI GPUs and domestic semiconductors

  A beneficiary direction for localization substitution and domestic AI inference demand

  Strengths

  DeepSeek is driving low-cost inference demand, domestic foundry supply-chain capabilities are improving, and vendors such as Cambricon, MetaX, and Iluvatar are drawing attention.

  Weaknesses

  Advanced nodes, software ecosystems, and high-end manufacturing capabilities remain constrained.

  Comparison

  Compared with global leading GPUs, China’s supply chain is narrowing the perceived technology gap through multi-chip packaging, cluster expansion, and manufacturing capacity growth.

  Risks

  Export controls, restrictions related to U.S. Executive Orders, insufficient advanced-node capacity, and regulatory uncertainty.

Key data

• 2025e cloud AI Semi TAM bull-case assumptionUS$235bnBull-case scenario assumption driven by supply-chain data, mainly contributed by NVIDIA AI GPUs.
• 2025e cloud capexUS$400bnOuter-layer Cloud Capex estimate in the chart.
• 2025e AI server capexUS$300bnAI Server Capex estimate in the chart.
• Cloud AI ASICs and non-NVIDIA GPUsabout US$30bnShown in the chart as a subset within AI Semi.
• AI semiconductor revenue growth2023-30e CAGR 38%Chart label for AI semi revenue breakdown by application.
• Potential TSMC CoWoS capacityMay reach 165kwpm in 2027The report heading indicates TSMC may expand CoWoS capacity further as AI demand remains strong.
• China AI GPU TAMUS$67bn by 2030eThe report heading shows Morgan Stanley expects China AI GPU TAM to grow to US$67bn by 2030e.
• AI accelerator revenue supported by China advanced-node capacityUS$58bn in 2030The report body title mentions that China advanced-node capacity may support US$58bn of AI accelerator revenue in 2030.

Impact & implications

For investors, AI infrastructure expansion benefits not only GPUs and leading foundries, but also advanced packaging, HBM, back-end equipment, test consumables, CPO, power semiconductors, and domestic AI GPUs. Bottlenecks extend from a single compute chip to packaging capacity, memory bandwidth, rack power delivery, test time, and regional regulation, so companies with scarce capacity, complex packaging capabilities, or local substitution positions are more likely to see valuation and earnings upside.

Risks

• Technology inflation raises wafer, OSAT, and memory costs, compressing chip designers’ margins.
• AI may crowd out some non-AI demand, and the supply chain may prioritize resources for AI semiconductors.
• Growth constraints include budgets, U.S. energy limits, China chip capacity constraints, and regulation.
• DDR4 shortages may continue into 2H26, but spot price upside may be limited.
• China-related entities and securities may be affected by U.S. executive orders, export controls, and investment restrictions; investors should confirm compliance on their own.

What to watch

• Whether major CSP cloud capex remains strong.
• TSMC CoWoS/SoIC expansion pace and 2026-2027 capacity utilization.
• NVIDIA GB200/300 rack supply-demand, shipments, and supply-share changes.
• HBM consumption and memory price trends, especially 2026e HBM Gb demand.
• Progress on Trainium, TPU, and other CSP ASIC projects.
• China AI GPU demand, advanced-node capacity expansion, and product progress from Cambricon, MetaX, Iluvatar, and others.
• Order upside for CPO, test equipment, and test consumables from increased AI packaging complexity.