AI semiconductor upcycle persists, and TSMC advanced process and CoWoS remain core bottleneck assets

Overview

This report is Morgan Stanley's AI semiconductor outlook for the Greater China semiconductor industry, including a TSMC performance preview and a CoWoS supply-demand update. The core themes include TSMC 2Q26 results and 3Q26 guidance, advanced-node pricing and capacity, CoWoS and SoIC expansion, hyperscaler AI capex, AI GPU and ASIC demand, HBM and storage supply-demand, China AI GPU market dynamics, and opportunities in CPO and test equipment. The overall conclusion is constructive: AI-related demand remains the strongest semiconductor cycle driver, and TSMC sits at the center of the bottleneck supply in advanced process and advanced packaging.

Core views

The report's core views are: first, TSMC remains the key beneficiary of advanced logic processes and advanced packaging, supported by technology leadership, tight EUV supply, strong N2/N3 demand, and CoWoS expansion; second, hyperscaler AI capex remains robust, and AI semiconductor market size is expected to reach about US$753bn by 2030; third, NVIDIA is expected to consume most AI wafer, CoWoS and HBM resources, while CSP self-developed ASICs such as Google TPU, AWS Trainium/Inferentia, and Meta MTIA are also likely to continue increasing in volume; fourth, China AI compute demand is being driven by inference applications like DeepSeek, cloud providers and telecom operators, and domestic AI accelerators are becoming more competitive with cost advantages; fifth, storage, test equipment, CPO, ABF substrates and advanced packaging are key bottlenecks supporting AI semiconductor expansion.

Analysis framework

The report applies a combination of top-down TAM sizing and bottom-up supply-chain tracking: it derives AI chip demand from global cloud capex, AI training and inference demand, rack power, and deployment scale, then cross-validates it against TSMC process, CoWoS, SoIC, HBM, ASIC projects, CSP chip roadmaps, and Chinese AI accelerator orders.

Methodology notes

• Industry cycle analysisTAM and bottom-up supply-chain estimation

  Estimate AI semiconductor market size through cloud capex, chip shipments, rack power, wafer and packaging demand.

  The report combines global CSP capex, NVIDIA and CSP custom-chip roadmaps, CoWoS/SoIC/HBM consumption, and TSMC capacity plans to assess long-term AI semiconductor revenue opportunities and supply bottlenecks.

• Company competitiveness analysisAdvanced process and packaging technology roadmap comparison

  Compare TSMC, Intel, and Samsung positions in logic density, process nodes, and packaging solutions such as CoWoS and EMIB.

  The report believes TSMC remains ahead in technology roadmap, logic density, and ecosystem execution. Intel EMIB has strong large-package potential but depends heavily on supply-chain execution.

• Demand decompositionLayered demand decomposition for AI GPU, ASIC, and CPU orchestration

  Break AI semiconductor demand into segments of general AI GPUs, proprietary ASICs, AI server CPUs, HBM, CoWoS, and testing.

  The report emphasizes that as AI transitions from inference toward action and multi-tool orchestration, demand for CPU orchestration, AI GPUs and proprietary ASICs may all move higher.

Asset mapping & comparison

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

• TSMC

  Core beneficiary

  Strengths

  Strong leading position in advanced process, logic density, EUV capacity, and CoWoS/SoIC advanced packaging, with a rising share of AI semiconductor revenue expected.

  Weaknesses

  High capex intensity; execution of capacity expansion and customer demand tempo can impact margins.

  Comparison

  Compared with Intel and Samsung, the report believes TSMC still maintains leadership in technology roadmap and logic density.

  Risks

  AI demand below expectations, EUV or advanced packaging expansion below expectations, and price increases reducing customer demand elasticity.

• AI GPU supply chain

  Strong demand-driven

  Strengths

  NVIDIA-related GPUs, including GB200/300 and Rubin platforms, continue to drive wafer, CoWoS, HBM, and test demand.

  Weaknesses

  The chain is highly concentrated and depends on a limited number of customers and platform cycles.

  Comparison

  Compared with custom ASICs, general AI GPUs remain the main compute supply, but CSP custom chip demand is steadily increasing.

  Risks

  Cloud capex slowdown, power and electricity constraints, export controls, and supply bottlenecks.

• AI ASIC and CSP custom chips

  Incremental opportunity

  Strengths

  Projects such as Google TPU, AWS Trainium/Inferentia, and Meta MTIA are expanding, with CSPs still needing custom chips to optimize costs and specific workloads.

  Weaknesses

  Mass production pace depends on ABF substrate, advanced packaging, and client program timing.

  Comparison

  Compared with NVIDIA GPUs, custom ASICs may have cost advantages in certain inference or cloud-internal workloads.

  Risks

  Design iteration failures, immature ecosystem, and supply-chain delivery below expectations.

• HBM and storage

  AI compute bottleneck

  Strengths

  The report expects NVIDIA to still absorb most HBM supply in 2027e; AI storage demand could also create tightness in NAND and NOR.

  Weaknesses

  Storage pricing and supply-demand cycles can be volatile.

  Comparison

  HBM benefits more directly than traditional DRAM from AI accelerator demand.

  Risks

  New supply coming online faster than demand, customer spec changes, and constrained spot DDR4 pricing.

• Chinese AI accelerators

  Domestic substitution and inference-demand beneficiaries

  Strengths

  China AI chip TAM is expected to rise to US$91bn, and domestic chips are seen to have lower TCO and stronger performance per dollar in Chinese inference scenarios.

  Weaknesses

  Advanced process and high-end supply chain remain constrained.

  Comparison

  Compared with NVIDIA processors, domestic chips may win in China on price and local supply, but still lag in high-end performance and ecosystem.

  Risks

  Regulatory restrictions, capacity bottlenecks, uncertain customer order conversion, and insufficient technology iteration speed.

Key data

• TSMC 2026e capexUS$56bnThe report's estimate of TSMC 2026e capital expenditure.
• TSMC 2027e capexUS$75bnThe report tracks TSMC's 2027 capital expenditure at US$75bn.
• TSMC leading-node price increase potential5%-10% in 2027The report believes TSMC can pass through value to customers through price increases.
• TSMC N2 capacity growth70% CAGR from 2026-2028eTSMC guidance implies N2 capacity grows at a high rate from 2026 to 2028e.
• AI semiconductor TAM~US$753bn by 2030The report expects AI semiconductor TAM to reach about US$753bn by 2030.
• Global semiconductor market sizeUS$1.5tn by 2030The report expects the global semiconductor industry market to possibly reach US$1.5tn by 2030, of which AI semiconductors contribute about half.
• Top 4 CSP capex growth+95% Y/Y in 1Q26CYCapital expenditure rose sharply year-over-year for Amazon, Google, Microsoft, and Meta.
• Top 14 listed global CSP cloud capexnearly US$1.3tn in 2027Morgan Stanley's cloud capex tracker estimate, excluding sovereign AI.
• Orchestration CPU TAM base caseUS$79bn by 2030The report raises the base case from US$60bn to US$79bn.
• Orchestration CPU TAM bull caseUS$238bn by 2030Top-down bull-case estimate.
• China AI chip TAMUS$91bn by 2030The report expects the China AI chip market to grow to US$91bn by 2030.
• Test equipment market growth35% CAGR during 2024-27eAI and HPC test demand is driving expansion in test equipment and components.

Impact & implications

From an investment perspective, the report reinforces structural strength across the AI semiconductor chain: TSMC, advanced packaging, HBM, test, ASIC design services, CPO, and domestic China AI accelerators could all benefit. At the same time, non-AI semiconductor demand may be weak in 2026, and rising wafer, OSAT, and storage costs may also compress semiconductor design company margins, so allocations should tilt toward segments with bottleneck supply positions, technological leadership, or clear AI order leverage.

Risks

• Hyperscaler capex or AI training and inference demand falls short of expectations.
• TSMC's expansion of advanced processes, CoWoS, or SoIC does not meet expectations.
• AI semiconductor supply chain faces bottlenecks from EUV, HBM, ABF substrate, T-Glass, test equipment, or power constraints.
• Weak non-AI semiconductor demand and rising chip costs compress design company margins.
• China AI chips are affected by export controls, advanced-process capacity limits, and the regulatory environment.
• CSP custom ASIC projects miss mass-production timing or performance expectations.

What to watch

• TSMC 2Q26 actual results, 3Q26 guidance, and 2026/2027 capex updates.
• TSMC advanced process capacity expansion for N2, N3, A16, A14 and customer demand mix.
• Whether CoWoS expansion is near the 2027 200kwpm target and SoIC capacity planning.
• Supply and demand timing for NVIDIA GB200/300, Rubin, Vera CPU, and AI PC RTX Spark/N1X.
• Changes in shipment forecasts for CSP self-developed ASICs such as Google TPU, AWS Trainium/Inferentia, and Meta MTIA.
• HBM, NAND, NOR, and DDR4 supply-demand and pricing trends.
• China AI GPU/accelerator orders, TCO, TPS performance, and cloud-provider inference demand.
• Orders in the CPO, PIC, test equipment, test sockets, and FOCI optical/ test chain.