Cloud capex, custom ASICs, and China’s domestic compute capacity jointly support strong long-term growth in AI semiconductors

Overview

The report examines six major themes: the global AI semiconductor cycle, custom ASICs from cloud service providers, TSMC’s advanced processes and CoWoS, memory supply and demand, China’s domestic AI compute capacity, and quantum security. Its core conclusion is that AI infrastructure investment can continue driving long-term industry expansion, but the divergence between AI and non-AI semiconductor conditions, rising costs, and energy and capacity bottlenecks will become more pronounced.

Core views

The report first assesses global industry scale and capex, concluding that AI semiconductors remain in a phase of strong expansion. It expects the global semiconductor market to reach US$1.5tn by 2030, with AI semiconductors contributing approximately half; its supply-chain-data-driven bull-case scenario indicates that the cloud AI semiconductor TAM could reach US$485bn in 2026 and the overall AI semiconductor TAM approximately US$753bn by 2030. The demand case is underpinned by cloud service provider investment: capex at Amazon, Google, Microsoft, and Meta increased 87% YoY in the second quarter of 2026, while their capex-to-EBITDA ratio has exceeded 70%. Morgan Stanley’s tracking model further estimates that capex by the world’s top 14 listed cloud service providers will approach US$1.4tn in 2027, excluding sovereign AI. Meanwhile, cloud logic chips account for approximately 20% of global cloud capex, allowing capital investment to flow through the server, GPU/ASIC, wafer, advanced-packaging, and memory value chains. This growth will not be evenly distributed across the semiconductor industry. The report expects logic foundry utilization to reach 80% in the second half of 2026, but excluding memory and NVIDIA AI GPU revenue, non-AI semiconductor growth is expected to decline in 2026. Rising wafer, assembly and testing, and memory costs could suppress demand for technology products through price elasticity and pressure chip-design companies’ margins. The supply chain will also prioritize scarce resources such as T-Glass and memory for AI products, crowding out non-AI semiconductors. The report therefore depicts structurally strong conditions in the AI value chain alongside weak traditional demand, rather than a synchronized industry-wide recovery. Advanced processes, CoWoS, and HBM are the main physical bottlenecks for realizing AI demand. The report estimates that TSMC produced 5.1mn related chips in 2025 and that full-year GB200 NVL72 rack shipments could reach 30k units. The value of wafers consumed by AI compute could exceed US$46bn in 2027, while HBM consumption could reach as much as 48bn Gb, with NVIDIA still accounting for the majority. Global CoWoS demand is expected to increase from 1,394 thousand wafers in 2026 to 2,694 thousand wafers in 2027: NVIDIA demand rises from 780 thousand to 1,222 thousand wafers, while its share declines from 56% to 45%; AMD rises from 130 thousand to 530 thousand wafers, with its share increasing from 9% to 20%; Broadcom rises from 300 thousand to 484 thousand wafers, while its share declines from 22% to 18%; MediaTek is expected to require 180 thousand wafers in 2027, representing 7%. Demand sources are becoming more diversified, but NVIDIA remains the largest customer. TSMC could accordingly expand CoWoS capacity to 200 thousand wafers per month by 2027. Its N2 capacity is expected to achieve a 70% CAGR from 2026 to 2028, N3 capacity will continue increasing, while N5 capacity will decline in 2027. Mature and specialty-process capacity is expected to expand at a 7% CAGR from 2024 to 2029. AI semiconductors are expected to account for more than 30% of TSMC’s revenue in 2026, while the report also states that AI semiconductor revenue growth from 2024 to 2029 “could reach 60%.” In packaging competition, TSMC CoWoS can support up to 9.5x reticle size and approximately four chips per wafer; if supply-chain execution proceeds smoothly, Intel EMIB could more readily support large chips exceeding 12x reticle size. CoWoS therefore has existing scale and ecosystem advantages, but competing technology routes remain for ultra-large chip packaging. The report believes that powerful general-purpose GPUs will not eliminate cloud service providers’ need to develop custom chips. Google TPU, AWS Trainium/Inferentia, Meta MTIA, and other ASIC projects can optimize performance, cost, and supply for specific workloads. Alchip’s estimated Trainium3 volumes are 400k chips in 2026 and 600k in 2027, while Trainium4 is expected to reach 800k chips in 2028. Its Trainium revenue is expected to be US$1.8bn, US$2.8bn, and US$8bn, respectively, accounting for 60%, 63%, and 82% of Alchip’s total revenue; Alchip’s total revenue over the same period is expected to be US$3.1bn, US$4.4bn, and US$9.8bn. Total Google TPU volume is expected to increase from 1.75mn chips in 2025 to 3.70mn in 2026 and 7.35mn in 2027, remaining above 6.50mn in 2028, although volume growth for AI chips such as TPUs may depend on ABF substrate supply. MediaTek is one of the report’s top AI picks. Even under a customer-owned-tooling COT model, the report believes MediaTek could at least be responsible for TPU v10 packaging and the I/O die. Its TPU revenue is expected to rise from US$13.5bn in 2027 to US$43.5bn in 2028 and reach US$70bn in 2029. The business is expected to account for 38% of MediaTek’s US$35bn total revenue in 2027 and 65% of its US$67bn total revenue in 2028. By project, TPU v8t is expected to reach 3mn chips and US$13.5bn in revenue in 2027; TPU v9 is expected to reach 3mn chips and US$39bn in revenue in 2028 and 4mn chips and US$52bn in revenue in 2029; TPU v10 is expected to reach 1mn chips and US$18bn in revenue in 2029. This growth will also flow through to packaging, I/O dies, and chip testing, with the report specifically noting that KYEC benefits from growth in AI GPUs, TPUs, and CPUs. In addition to GPUs and ASICs, the report views Agentic AI as a new source of server CPU growth. NVIDIA’s Vera CPU could deliver 1.8x the performance of the highest-performance x86 CPU, and its cores are not separated across different chiplets, enabling faster inter-core connectivity. Morgan Stanley’s top-down model expects the Agentic CPU TAM to expand at a 251% CAGR from F26 to F30, with the bull-case scenario implying a US$238bn CPU orchestration TAM. This view extends AI compute demand beyond accelerators to CPUs responsible for data processing, task coordination, and system orchestration. Memory is likewise affected by AI infrastructure demand and shifting supply priorities. The report expects AI storage demand to cause NAND shortages, with NOR Flash undersupply persisting through 2026. The DDR4 shortage is expected to continue into the second half of 2026, although upside in spot prices is limited. The rationale is that AI servers require more high-capacity and advanced memory, while capacity and supply-chain resources are shifting toward AI products, creating divergent degrees of supply-demand tightness and price performance across memory categories. China AI compute is another independent growth theme. The report expects China’s AI chip TAM to grow to US$91bn by 2030. Low-cost inference demonstrated by DeepSeek is stimulating inference demand, while the local foundry chain’s AI GPU production capability is improving. Surging Token volumes at ByteDance’s Volcano Engine and Doubao are used as indicators of demand intensity. Compared with NVIDIA processors in the Chinese market, domestic chips offer a lower total cost of ownership and similar per-Token costs, demonstrating stronger cost-adjusted performance because of their significantly lower prices. The SuperPod ecosystem presented at WAIC 2026 further illustrates improving domestic system capabilities: Huawei Atlas 950 can scale to 1,024 Ascend 950DT NPUs, using 16 compute cabinets and 4 UnifiedBus cabinets, hybrid copper-optical interconnects, and up to 256TB of pooled memory. Moore Threads, Enflame with ZTE, and MetaX with ZTE also demonstrated solutions at scales of 128 to 256 accelerators or 64 accelerators. The report believes infrastructure and system-integration capabilities are narrowing the market-perceived technology gap between China and the United States. In China’s CPU market, the report expects the TAM to rise to US$42bn by 2030 and expects Hygon Information’s share of China’s server CPU market to reach 18% by 2028; its integrated CPU and GPU computing platform receives an OW view. For domestic AI accelerators, the report compares vendors including Cambricon, MetaX, and Iluvatar CoreX and incorporates chips, systems, and infrastructure into a nine-factor China-US comparison. Growth nevertheless remains constrained by US energy supply, China’s chip capacity, budgets, and regulation; supply realization cannot be inferred solely from end-market inference demand. Finally, the report views quantum security as a new semiconductor demand theme around 2030. Modern electronic systems primarily rely on asymmetric cryptography, while quantum computing may drive countries to migrate to post-quantum cryptographic systems. Major economies are broadly aligning with NIST post-quantum cryptography standards, although migration timelines differ by region, with 2030 viewed as a key inflection point. Security demand will span different connectivity layers and make hardware security foundational to the quantum supply chain. The report expects the TAM for physical unclonable functions, or PUFs, to reach approximately US$310mn by 2030 and assesses related asset valuations using 2027 P/E ratios for global PUF/IP-related companies and eMemory’s historical P/E range. Overall, this theme is far smaller than AI chips but has a clear policy-driven migration timetable and a logical basis in hardware-security applications.

Analysis framework

The report first estimates total AI infrastructure demand using cloud service provider capex, power deployment, and server plans, then converts this demand into GPU/ASIC volumes, wafer consumption, CoWoS packaging volume, and HBM capacity. It subsequently derives revenue for companies such as Alchip and MediaTek based on chip volumes and per-chip ASPs for each cloud service provider project. The China section compares Token demand, chip price-performance, total cost of ownership, and SuperPod system capabilities. The quantum-security section derives the PUF market from regional migration schedules, changes in cryptographic systems, and hardware-security segments, supplemented by P/E ranges for valuation assessment.

Methodology notes

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

  Transmission from cloud capex to the AI chip supply chain

  Starting with cloud service provider capex and data-center power demand, the report sequentially derives demand for servers, GPUs/ASICs, wafers, advanced packaging, HBM, and testing to identify where growth accrues across the value chain.

• Industry/Sector Analysis FrameworkSupply-demand framework

  Supply-demand estimates for wafers, CoWoS, and memory

  The report compares customer demand, existing capacity, and expansion plans to assess supply-demand gaps, customer shares, and shortage durations for CoWoS, HBM, NAND, NOR, and DDR4.

• Industry/Sector Analysis FrameworkVolume-price decomposition

  Revenue model based on ASIC project volumes multiplied by per-chip ASP

  The revenue estimates for Alchip Trainium and MediaTek TPU use project chip volumes and identifiable per-chip ASPs as core variables, then calculate project revenue and its share of total company revenue.

• Industry/Sector Analysis FrameworkCost curve analysis

  Comparison of total cost of ownership, per-Token cost, and cost-adjusted performance

  The China AI compute section compares hardware prices, total inference costs, and per-Token costs to explain why domestic chips may achieve strong cost-adjusted performance at lower prices.

• Valuation MethodPE/PEG valuation

  Cross-sectional P/E comparison and historical ranges

  The report compares P/E ratios across foundries, back-end companies, memory, IDMs, equipment, fabless companies, and quantum-security-related companies, while referencing 2027 P/E ratios for PUF/IP companies and eMemory’s historical P/E range.

• Industry/Sector Analysis Framework

  Top-down TAM and bull-case estimates

  The report uses aggregate capital investment, compute demand, and application assumptions to derive potential market sizes for cloud AI semiconductors and Agentic CPUs, presenting a separate bull-case scenario rather than treating it as the sole outcome.

Asset mapping & comparison

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

• MediaTek（2454.TW）

  The report names it as a Top Pick in AI and expects it to undertake design, packaging, or I/O die work for Google TPUs.

  Strengths

  TPU revenue is expected to increase from US$13.5bn in 2027 to US$70bn in 2029; even under a COT model, the report believes it could at least be responsible for TPU v10 packaging and the I/O die.

  Weaknesses

  TPU operations are expected to account for 65% of total revenue in 2028, significantly increasing the importance of project revenue.

  Comparison

  Compared with providing only a single design function, the report emphasizes that it may cover more value-chain segments, including chip design, packaging, and I/O dies.

  Risks

  TPU volume growth may be constrained by ABF substrate supply.

• TSMC

  Global AI GPU and ASIC demand flows through to TSMC via advanced processes, CoWoS, and wafer testing.

  Strengths

  CoWoS capacity could expand to 200kwpm by 2027, N2 capacity is expected to achieve a 70% CAGR from 2026 to 2028, and AI semiconductors are expected to account for more than 30% of revenue in 2026.

  Weaknesses

  N5 capacity is expected to decline in 2027, while the business mix becomes more concentrated on AI demand.

  Comparison

  CoWoS can support up to 9.5x reticle size, while Intel EMIB could support more than 12x reticle size if its supply chain executes smoothly.

  Risks

  Energy, customer budgets, project execution, and advanced-packaging supply capacity could all affect demand conversion.

• NVIDIA（NVDA.US）

  The report expects it to continue accounting for the largest share of AI wafer, CoWoS, and HBM consumption.

  Strengths

  CoWoS demand is expected to reach 1,222 thousand wafers in 2027; Vera CPU performance could be 1.8x that of the highest-performance x86 CPU, and it could benefit from Agentic CPU demand.

  Weaknesses

  Its share of CoWoS demand is expected to decline from 56% in 2026 to 45% in 2027, indicating faster growth among other ASIC and GPU customers.

  Comparison

  Domestic Chinese chips are significantly cheaper and can achieve similar per-Token costs and strong cost-adjusted performance in Chinese inference scenarios.

  Risks

  US energy supply, customer budgets, and regulation are among the AI growth constraints identified in the report.

• Alchip

  AWS Trainium3 and Trainium4 projects are the primary drivers of its revenue growth.

  Strengths

  Trainium revenue is expected to increase from US$1.8bn in 2026 to US$8bn in 2028, while the company’s total revenue is expected to rise from US$3.1bn to US$9.8bn over the same period.

  Weaknesses

  Trainium’s revenue share is expected to rise from 60% in 2026 to 82% in 2028.

  Comparison

  It primarily serves AWS custom training chips, while MediaTek’s incremental growth mainly comes from Google TPU projects.

  Risks

  The report does not separately identify other asset-specific risks.

• Cambricon（688256.SH）、MetaX-U（688802.SH）、Iluvatar CoreX（09903.HK）

  The report includes these companies in its comparison of China’s domestic AI accelerator and SuperPod ecosystems.

  Strengths

  Relying on significantly lower prices, domestic chips demonstrate lower total cost of ownership, similar per-Token costs, and strong cost-adjusted performance.

  Weaknesses

  China’s AI chip supply remains constrained by domestic capacity.

  Comparison

  The report compares China and US AI capabilities across nine factors, including chips, systems, and infrastructure, and separately compares Cambricon, MetaX, and Iluvatar CoreX.

  Risks

  Chip capacity and regulation are explicitly identified in the report as constraints on China’s AI growth.

• Hygon Information

  Its integrated CPU and GPU computing platform is positioned to address demand for Chinese server CPUs and AI compute, and the report assigns it an OW view.

  Strengths

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

  Comparison

  It is positioned as an integrated CPU and GPU computing platform rather than solely providing a standalone AI accelerator.

  Risks

  China’s chip capacity and regulatory constraints may affect the realization of industry demand.

Key data

• Global semiconductor market sizeUS$1.5tn，2030eThe report expects AI semiconductors to contribute approximately half.
• Cloud AI semiconductor TAMUS$485bn，2026eSupply-chain-data-driven bull-case scenario.
• Global AI semiconductor TAMApproximately US$753bn，2030eThe report’s long-term market-size forecast.
• Capex growth of the four largest cloud service providersYoY+87%Second quarter of 2026; includes Amazon, Google, Microsoft, and Meta.
• Cloud service provider capex-to-EBITDA ratioAbove 70%Indicates that AI infrastructure investment intensity is already elevated.
• Global cloud capexNearly US$1.4tn，2027eTop 14 listed global cloud service providers, excluding sovereign AI.
• Share of capex allocated to cloud logic chipsApproximately 20%Estimated share of global cloud capex.
• Logic foundry utilization80%Expected to be reached in the second half of 2026.
• TSMC related-chip production5.1mn chips，2025Used to support supply estimates for GB200 and related AI systems.
• GB200 NVL72 rack shipments30k units，2025The report’s full-year shipment estimate.
• Total global CoWoS demand1,394 thousand wafers，2026e；2,694 thousand wafers，2027eDemand is expected to nearly double.
• NVIDIA CoWoS demand780 thousand wafers，2026e；1,222 thousand wafers，2027eCorresponding shares are 56% and 45%, respectively, while it remains the largest customer.
• AMD CoWoS demand130 thousand wafers，2026e；530 thousand wafers，2027eCorresponding share rises from 9% to 20%.
• Broadcom CoWoS demand300 thousand wafers，2026e；484 thousand wafers，2027eCorresponding share declines from 22% to 18%.
• TSMC CoWoS capacity target200kwpm，2027eEquivalent to potential capacity of approximately 200 thousand wafers per month.
• Value of wafers consumed by AI computeAbove US$46bn，2027eNVIDIA is expected to account for the majority.
• HBM consumptionUp to 48bn Gb，2027eNVIDIA is expected to continue consuming most of the supply.
• TSMC N2 capacity growth70% CAGR，2026-2028eReflects strong 2nm customer demand.
• TSMC mature and specialty-process capacity growth7% CAGR，2024-2029eRuns in parallel with advanced-process expansion.
• TSMC AI semiconductor revenue shareAbove 30%，2026eThe importance of AI to company revenue continues to increase.
• Vera CPU relative performance1.8xThe report’s estimate relative to the highest-performance x86 CPU.
• Agentic CPU TAM growth251% CAGR，F26-F30Estimated using Morgan Stanley’s top-down model.
• CPU orchestration TAMUS$238bnThe report’s bull-case scenario.
• Alchip Trainium revenueUS$1.8bn / US$2.8bn / US$8bnCorresponding to 2026e, 2027e, and 2028e, respectively.
• Alchip Trainium revenue share60% / 63% / 82%Corresponding to 2026e, 2027e, and 2028e, respectively.
• Total Google TPU volume1.75mn / 3.70mn / 7.35mn / >6.50mnCorresponding to 2025e, 2026e, 2027e, and 2028e, respectively.
• MediaTek TPU revenueUS$13.5bn / US$43.5bn / US$70bnCorresponding to 2027e, 2028e, and 2029e, respectively.
• MediaTek TPU revenue share38%，2027e；65%，2028eCorresponding to company total-revenue forecasts of US$35bn and US$67bn.
• China AI chip TAMUS$91bn，2030eThe domestic AI semiconductor share and market size are expected to grow in tandem.
• China CPU TAMUS$42bn，2030eThe report’s long-term forecast for China’s server CPU market.
• Hygon server CPU share18%，2028eForecast share of China’s server CPU market.
• Atlas 950 SuperPod scale1,024 Ascend 950DT NPUsConfigured with 16 compute cabinets, 4 UnifiedBus cabinets, and up to 256TB of pooled memory.
• PUF market sizeApproximately US$310mn，2030eTAM forecast related to quantum hardware security.

Impact & implications

The report believes AI infrastructure investment will continue shifting value toward advanced processes, CoWoS, HBM, custom ASICs, I/O dies, and testing. NVIDIA still accounts for the largest share of demand, but projects involving AMD, Broadcom, Google, AWS, and MediaTek will diversify the customer mix. In China, low-cost inference, domestic chips’ price-performance, and improving SuperPod system capabilities could expand demand for domestic AI chips and drive the joint development of CPUs, GPUs, interconnects, and system integration. Conversely, high AI investment will crowd out resources for non-AI semiconductors, while rising wafer, assembly and testing, and memory costs may compress chip-design companies’ margins. The quantum-security market is currently small, but regional post-quantum cryptography migration roadmaps provide relatively clear timing for hardware-security demand around 2030.

Risks

• Rising wafer, assembly and testing, and memory costs could suppress demand for technology products through price elasticity and pressure chip-design companies’ margins in 2026.
• Supply-chain prioritization of AI semiconductors could cause shortages of T-Glass and memory and crowd out resources for non-AI products.
• Excluding memory and NVIDIA AI GPU revenue, non-AI semiconductor growth is expected to decline in 2026.
• AI infrastructure growth is constrained by customer budgets.
• US AI data-center expansion is constrained by energy supply.
• China’s AI chip growth is constrained by domestic capacity.
• Regulatory changes could restrict AI semiconductor demand, supply, or project deployment.
• Volume growth for AI chips such as Google TPUs may depend on ABF substrate supply.
• Intel EMIB’s potential to support ultra-large chips depends on successful supply-chain execution.

What to watch

• Track capex growth among major cloud service providers and whether the capex-to-EBITDA ratio continues rising.
• Monitor whether TSMC can expand CoWoS capacity to 200kwpm by 2027 and whether N2 capacity can achieve a 70% CAGR from 2026 to 2028.
• Watch whether CoWoS demand can reach 2,694 thousand wafers in 2027 and how customer shares change among NVIDIA, AMD, Broadcom, and MediaTek.
• Track Google TPU and AWS Trainium project volumes, ABF substrate supply, and the revenue-recognition progress of Alchip and MediaTek.
• Monitor NAND, NOR, and DDR4 supply and demand, with NOR tightness expected to persist through 2026 and the DDR4 shortage expected to continue into the second half of 2026.
• Watch Token demand from China’s leading large models, the total cost of ownership and per-Token costs of domestic chips, and SuperPod deployment progress.
• Track the four constraints on AI growth: US energy supply, China’s chip capacity, budgets, and regulation.
• Monitor regional post-quantum cryptography migration roadmaps, particularly key implementation milestones around 2030.