AI Computing Expansion Continues to Reshape Global Semiconductors, with Custom ASICs, China Computing Power, and Quantum Security Emerging as New Growth Drivers

Overview

This is an industry outlook covering global AI semiconductors, cloud custom ASICs, China AI computing power, and quantum security. The report believes cloud capex, inference demand, advanced packaging, and domestic computing infrastructure will support long-term growth, but the divergence between AI and non-AI semiconductor conditions will widen, while costs, energy, capacity, and regulation constitute the main constraints.

Core views

The report assigns an Attractive view to the Asia-Pacific semiconductor industry, with a clear allocation preference for the AI supply chain, niche memory, and China's domestic computing ecosystem. Within AI, MediaTek is the top pick, while TSMC, SMIC, Aspeed, Alchip, KYEC, ASE, FOCI, ASMPT, AllRing, and GUC are also favored. Within memory, AP Memory is the top pick, while Macronix, Nanya Tech, Winbond, and GigaDevice are also favored. Within China AI/semiconductors/wafer fab equipment, the report prefers Iluvatar, Cambricon, Hygon Information, NAURA Technology, and AMEC; testing equipment and consumables names include WinWay, MPI, Hon Hai, and Gudeng; and UMC is preferred within mature nodes. The report rates OmniVision Group, Phison, MetaX, Realtek, and GlobalWafers EW, and WIN Semi, Silergy, and ASMedia UW. The overall cycle shows a divergence between strong AI and weak non-AI demand. The report expects the global semiconductor industry to potentially reach US$1.5 trillion by 2030, with AI semiconductors contributing approximately half; the AI semiconductor TAM is estimated at approximately US$753 billion by 2030, while a supply-chain-data-driven bull-case scenario indicates that the cloud AI semiconductor TAM could reach US$485 billion in 2026. Meanwhile, excluding memory and NVIDIA AI GPU revenue, non-AI semiconductor growth is expected to decline in 2026. Logic semiconductor foundry utilization could reach 80% in the second half of 2026. Historically, declining days of inventory have usually coincided with gains in semiconductor stock indices, but the current recovery is not occurring synchronously across all submarkets. Cloud service provider spending remains the core starting point for AI chip demand. Combined capex at the four major cloud service providers—Amazon, Google, Microsoft, and Meta—grew 87% year over year in the second quarter of 2026, and the capex-to-EBITDA ratio has exceeded 70%. Morgan Stanley's tracking model estimates that capex among the world's top 14 listed cloud service providers will approach US$1.4 trillion in 2027, excluding sovereign AI. The report further derives demand for racks, power, GPUs/ASICs, advanced-node wafers, CoWoS, and HBM from capex, concluding that budgets remain robust, although capex intensity, US power availability, and the speed of project deployment will determine whether the forecasts materialize. TSMC is the primary beneficiary of this transmission chain. The report estimates that TSMC produced 5.1 million related chips in 2025 and that full-year GB200 NVL72 rack shipments could reach 30,000 units; AI compute wafer consumption value is expected to exceed US$46 billion in 2027, with NVIDIA still contributing most of the demand. HBM consumption could reach as high as 48 billion Gb in 2027, with NVIDIA again accounting for the majority. Total CoWoS demand is expected to increase from 1.394 million wafers in 2026 to 2.694 million wafers in 2027, leading the report to suggest that TSMC could expand CoWoS capacity to 200,000 wafers per month by 2027. The customer mix will become more diversified: NVIDIA's CoWoS demand is projected at 780,000 wafers in 2026 and 1.222 million wafers in 2027, representing approximately 56% and 45%, respectively; Broadcom's shares are approximately 22% and 18%, AMD's approximately 9% and 20%, and MediaTek's 2027 demand is approximately 180,000 wafers, or about 7%. Advanced-node demand is similarly robust. Information provided by TSMC indicates that N2 capacity could achieve a 70% compound annual growth rate from 2026 to 2028; N5 capacity will decline in 2027 while N3 continues to increase, consistent with the report's customer demand forecasts. Mature-node and specialty-node capacity is expected to expand at a 7% compound annual growth rate from 2024 to 2029. AI semiconductor revenue could account for more than 30% of TSMC's 2026 revenue and drive the margin expansion thesis presented in the report. AI server CPUs and memory constitute two additional growth and constraint chains. NVIDIA's Vera CPU uses a design that does not split cores across chiplets, shortening the connection paths between cores; the report states that its performance could reach 1.8 times that of the highest-performance x86 CPU. The bull-case scenario corresponds to a US$238 billion CPU orchestration TAM, while the top-down model projects a 251% compound annual growth rate for the Agentic CPU TAM from F26 to F30. In memory, AI storage demand is causing NAND shortages, tight NOR Flash supply is expected to persist through 2026, and the DDR4 shortage could continue into the second half of 2026. Meanwhile, rising wafer, assembly and testing, and memory costs will place margin pressure on chip design companies in 2026, while price elasticity could suppress demand for end-market technology products. Custom ASICs will not become unnecessary because of improvements in NVIDIA GPU performance. The report believes large cloud service providers will still require chips optimized for their proprietary workloads and expects more ASIC projects to proceed according to each cloud service provider's roadmap. Alchip's model shows 800,000 confirmed Trainium4 units in 2028, corresponding to US$8 billion in revenue; Alchip's total revenue that year is projected at US$9.8 billion, with Trainium contributing 82%. Google TPU forecasts show total volume rising from 3.7 million units in 2026 to 7.35 million units in 2027 and remaining above 6.5 million units in 2028, although upside could be constrained by ABF substrate supply. MediaTek is the most prominent custom ASIC beneficiary highlighted in the report. Its TPU revenue is expected to rise from US$13.5 billion in 2027 to US$43.5 billion in 2028 and reach US$70 billion in 2029. Even if the project adopts a customer-owned tooling model, the report still believes MediaTek will at least be responsible for TPU v10 packaging and the I/O die. KYEC is considered a beneficiary of simultaneous growth in AI GPUs, TPUs, and CPUs. In advanced packaging competition, TSMC's CoWoS can support up to approximately 9.5 times reticle size, or four chips per wafer; if its supply chain executes smoothly, Intel EMIB can more readily support larger chips exceeding 12 times reticle size. Therefore, the final competitive outcome depends not only on design capabilities but also on supply chain delivery. China AI computing power is an independent long-term growth theme. DeepSeek has demonstrated a lower-cost approach to inference, which the report believes will stimulate inference demand; the domestic foundry supply chain is also becoming increasingly capable of producing AI GPUs. China's AI chip/GPU TAM is expected to reach US$91 billion by 2030, with the market size and share of domestic accelerators continuing to expand. Demand-side evidence includes persistently rising NVIDIA 5090 prices in China, surging token usage on ByteDance's Volcano Engine/Doubao, and Chinese cloud capex trends. The report's inference economics comparison shows that domestic chips have lower TCO and comparable per-token costs relative to NVIDIA processors available in China, while delivering better performance per unit cost due to significantly lower prices. China's infrastructure capabilities are also narrowing the market's perceived technology gap. The 2026 World Artificial Intelligence Conference showcased multiple SuperPod solutions: Huawei Atlas 950 is configured with 1,024 Ascend 950DT NPUs, 16 compute cabinets, and four UnifiedBus cabinets, uses hybrid copper-optical interconnects, and provides up to 256 TB of pooled memory; Moore Threads' solution covers 128 to 256 S-series GPGPUs, while the Enflame-ZTE and MetaX-ZTE solutions are each configured with 64 accelerators and use a midplane-free orthogonal architecture. Hygon Information's integrated CPU and GPU platform received an OW view; the report expects China's CPU TAM to rise to US$42 billion by 2030 and Hygon's share of China's server CPU market to reach 18% by 2028. The quantum security section views hardware security as foundational to the global quantum supply chain. The report notes that modern electronic systems primarily rely on asymmetric cryptography, while quantum computing is prompting countries to formulate post-quantum cryptography migration roadmaps. Timelines vary by region, but 2030 is a key inflection point, and major economies are broadly converging toward NIST's PQC standards. The report analyzes quantum security demand across different connectivity layers and expects the TAM for physically unclonable functions, or PUFs, to reach approximately US$310 million by 2030, while assessing market pricing through the P/E ranges of relevant semiconductor/PUF IP companies and eMemory.

Analysis framework

The report first assesses aggregate demand using the global semiconductor cycle, inventories, and cloud capex, and then decomposes cloud spending layer by layer into power, server racks, GPUs/ASICs, advanced-node wafers, CoWoS, and HBM. For Trainium and TPU projects, the report estimates supplier revenue by multiplying chip volumes by recognized unit value; for China AI computing power, it compares market size, token demand, TCO, per-token cost, and SuperPod architectures. The quantum security section derives the PUF market opportunity from cryptographic migration timelines, connectivity layers, and hardware security components, supplemented by comparisons of P/E ranges.

Methodology notes

• Industry/Sector Analysis FrameworkSupply-demand framework

  Supply-Demand Model for AI Chips, CoWoS, HBM, and Memory

  The report separately estimates demand and supply capacity for chips, wafers, advanced packaging, and memory to assess capacity expansion, the duration of shortages, and customer shares.

• Industry/Sector Analysis FrameworkUpstream-Midstream-Downstream Industry Chain Transmission

  Transmission from Cloud Capex to Semiconductor Demand

  Starting from cloud service provider capex and power deployment, the research sequentially derives demand for racks, GPUs/ASICs, advanced-node wafers, CoWoS, HBM, and testing.

• Industry/Sector Analysis FrameworkVolume-price decomposition

  ASIC Project Volume and Unit-Price Revenue Model

  The Alchip Trainium and MediaTek TPU forecasts multiply the volume of each chip generation by the supplier's recognized per-chip value to derive annual project revenue and its share of revenue.

• Cycle and Business Conditions FrameworkInventory cycle (Kitchin)

  Semiconductor Days of Inventory and Business Conditions Assessment

  The report examines supply chain days of inventory and their historical changes, noting that periods of declining days of inventory have generally coincided with rising semiconductor stock indices.

• Valuation MethodPE/PEG valuation

  Comparison of P/E Ratios and Historical P/E Ranges

  The report compares valuations across semiconductor subsectors and assesses market pricing using the P/E ranges of PUF/IP-related companies and eMemory.

• Industry/Sector Analysis Framework

  Comparison of Inference TCO, Per-Token Cost, and Performance per Unit Cost

  The report combines hardware procurement and operating costs with model inference output to compare domestic chips and NVIDIA processors in China in terms of total cost of ownership, per-token cost, and price-performance.

Asset mapping & comparison

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

• MediaTek (2454.TW)

  The report names it the top pick in AI and expects it to capture packaging, I/O die, and related chip value from the Google TPU project.

  Strengths

  TPU revenue is expected to rise from US$13.5 billion in 2027 to US$70 billion in 2029; even under a customer-owned tooling model, it would still handle at least TPU v10 packaging and the I/O die.

  Comparison

  The report places MediaTek at the center of the global cloud custom ASIC supply chain.

  Risks

  Upside to Google TPU volume growth could be constrained by ABF substrate supply.

• TSMC

  Benefits from rising demand for AI advanced-node wafers, CoWoS, and multiple GPU/ASIC customers.

  Strengths

  N2 capacity is expected to achieve 70% compound growth from 2026 to 2028, while AI semiconductor revenue could exceed 30% of 2026 revenue.

  Weaknesses

  N5 capacity is expected to decline in 2027.

  Comparison

  CoWoS can support up to approximately 9.5 times reticle size; Intel EMIB can support more than 12 times reticle size if supply chain execution proceeds smoothly.

  Risks

  Cloud capex, project power deployment, and the pace at which customer demand materializes affect capacity expansion utilization.

• Alchip

  Benefits from growth in AWS Trainium custom AI training chips.

  Strengths

  Trainium revenue is expected to reach US$8 billion in 2028, with total revenue projected at US$9.8 billion.

  Weaknesses

  Trainium revenue is expected to account for 82% of total revenue in 2028.

  Comparison

  The report identifies it as a direct beneficiary of major cloud service providers' custom ASIC projects.

• KYEC

  The report believes it benefits from simultaneous growth in testing demand for AI GPUs, TPUs, and CPUs.

  Strengths

  Demand sources span general-purpose GPUs, cloud custom ASICs, and AI server CPUs.

• NVIDIA (NVDA.US)

  It remains the largest contributor to AI wafer, CoWoS, and HBM demand and is expanding into Agentic CPUs through Vera CPU.

  Strengths

  Vera performance could reach 1.8 times that of the highest-performance x86 CPU; 2027 CoWoS demand is expected to be 1.222 million wafers.

  Comparison

  Domestic chips achieve lower TCO and stronger performance per unit cost in China through lower prices.

  Risks

  US energy constraints, cloud customer budgets, and regulation are limiting factors for AI expansion.

• SMIC (00981.HK)

  Included in the OW list for China AI/semiconductors and benefits from improving domestic Chinese AI GPU production capabilities.

  Strengths

  The report believes China's domestic foundry supply chain is becoming more capable in AI GPU production.

  Comparison

  Together with other Chinese domestic computing and wafer fab equipment companies, it forms the domestic AI supply chain.

  Risks

  China's chip capacity is an explicitly identified constraint on growth in the report.

• Cambricon (688256.SH)

  Included in the OW list for China AI/semiconductors and covered in the horizontal comparison of domestic AI accelerators.

  Strengths

  Benefits from the expansion of China's AI chip/GPU TAM and the rising share of domestic accelerators.

  Comparison

  The report compares it with MetaX and Iluvatar.

  Risks

  Chip capacity and regulation could constrain growth in China's AI market.

• Hygon Information

  Its integrated CPU and GPU computing platform received an OW view.

  Strengths

  The report expects its share of China's server CPU market to reach 18% by 2028.

  Comparison

  Corresponds to the growth trajectory of China's CPU TAM rising to US$42 billion by 2030.

  Risks

  China's chip capacity constraints could affect market expansion.

• AP Memory

  Named the top pick in memory and participates in the advanced packaging technology chain related to Intel EMIB.

  Strengths

  Positioned at the intersection of niche memory and advanced packaging, both favored by the report.

  Comparison

  It has a higher priority than the other named companies in the recommended memory portfolio.

Key data

• Global Semiconductor Market SizeUS$1.5tn by 2030eThe report expects AI semiconductors to contribute approximately half by then
• AI Semiconductor TAM~US$753bn by 2030eLong-term global AI semiconductor market forecast
• Cloud AI Semiconductor TAM Bull-Case ScenarioUS$485bn in 2026eBull-case scenario assumptions driven by supply chain data
• Capex Growth of the Four Major Cloud Service Providers87% Y/Y in 2Q26CYCombined total for Amazon, Google, Microsoft, and Meta
• Cloud Service Provider Capex Intensityover 70%Capex-to-EBITDA ratio
• Global Cloud Capexnearly US$1.4tn in 2027Top 14 listed global cloud service providers, excluding sovereign AI
• Total CoWoS Demand1,394k wafers in 2026e; 2,694k wafers in 2027eGlobal customer demand forecast
• Potential TSMC CoWoS Capacity200kwpm by 2027Capacity expansion assessment based on persistently strong AI demand
• 2027 AI Compute Wafer Consumption>US$46bnNVIDIA is expected to contribute the majority of demand
• 2027 HBM Consumptionup to 48bn GbNVIDIA is expected to continue consuming most of the supply
• TSMC N2 Capacity Growth70% CAGR from 2026-2028eCapacity growth trajectory indicated by TSMC
• TSMC AI Semiconductor Revenue Share>30% of 2026e revenueForecast share of AI-related revenue in TSMC's total revenue
• Vera CPU Performance1.8xRelative to the highest-performance x86 CPU
• Agentic CPU TAM Growth251% CAGR from F26 to F30Morgan Stanley top-down model
• Alchip Trainium RevenueUS$8bn in 2028eExpected to account for 82% of Alchip's US$9.8 billion total revenue that year
• MediaTek TPU RevenueUS$70bn in 2029eThe 2027 and 2028 forecasts are US$13.5 billion and US$43.5 billion, respectively
• China AI Chip/GPU TAMUS$91bn by 2030eLong-term forecast for China's AI accelerator market
• China CPU TAMUS$42bn in 2030Hygon's share of China's server CPU market is expected to reach 18% by 2028
• PUF TAMUS$310mn by 2030Market forecast related to quantum hardware security

Impact & implications

The report believes semiconductor growth will become increasingly concentrated in AI chips, advanced nodes, CoWoS, HBM, testing, and custom ASICs, allowing TSMC, MediaTek, Alchip, and related assembly and testing companies to capture greater value. Non-AI chip design companies may simultaneously face slowing demand and rising wafer, assembly and testing, and memory costs. China's domestic chips partly offset the single-chip performance gap through lower prices, lower TCO, and SuperPod infrastructure; post-quantum cryptography migration also creates a new long-term market for hardware security components such as PUFs.

Risks

• AI growth could be constrained by cloud service provider budgets and capex intensity.
• US power supply is an explicit constraint on AI infrastructure expansion.
• China's chip capacity could limit the realization of domestic AI computing demand.
• Regulatory changes could affect AI semiconductor projects and market expansion.
• Rising wafer, assembly and testing, and memory costs will place margin pressure on chip design companies in 2026 and could weaken end demand through price elasticity.
• Supply chain prioritization of AI semiconductors could crowd out non-AI semiconductors and cause T-Glass and memory shortages.
• Upside to volume growth for custom chips such as Google TPUs could be constrained by ABF substrate supply.
• Whether Intel EMIB can realize its large-chip advantage depends on its supply chain execution.

What to watch

• Track whether capex growth at major cloud service providers and the capex-to-EBITDA ratio can remain elevated.
• Monitor power deployment, GB200/300 rack shipments, and their conversion into demand for CoWoS, advanced-node wafers, and HBM.
• Watch the progress of TSMC's CoWoS capacity expansion, N2 ramp-up, and the decline in N5 and increase in N3 during 2027.
• Track near-term AI demand indicators such as NVIDIA 5090 prices in China, token prices for mainstream large models, and Doubao token usage.
• Monitor whether shortages of NAND, NOR Flash, and DDR4 persist through 2026 and the second half of 2026, respectively, as described in the report.
• Watch regional post-quantum cryptography migration roadmaps and implementation milestones around 2030.