Cloud AI Expansion Further Tightens Asia Semiconductor Supply-Demand

Overview

The report systematically assesses the impact of cloud AI capital expenditure on the Asian semiconductor value chain, covering advanced nodes, advanced packaging, Google TPU, co-packaged optics, mature nodes, and key companies. The core conclusion is that AI computing demand is expanding from GPUs to custom ASICs, TPUs, server CPUs, and networking chips, driving continued tightness in key capacity such as N3, N2, and CoWoS. With leading process technology, packaging capabilities, and execution track record, TSMC remains in the most favorable position in the value chain; MediaTek, ASE, and related equipment, testing, and materials vendors will also share incremental value.

Core views

First, advanced packaging has become a necessary condition for improving overall AI chip performance, and TSMC CoPoS and Intel EMIB-T are more likely to coexist rather than form a single dominant technology route; CoPoS focuses on the highest performance and integration density, while EMIB-T focuses on cost, ultra-large package scalability, and supply-chain diversification. Second, Google TPU demand is driven simultaneously by internal training and inference, frontier model R&D, external customers, and cloud service penetration, and a dual-supplier route is favorable for MediaTek to gain structural share. Third, each GW of AI server buildout will generate substantial demand for advanced nodes, CoWoS, and wafer fabrication equipment, and next-generation NVIDIA platforms may increase TSMC’s value content per rack. Fourth, TSMC’s market position in N3, N2, and advanced packaging remains solid, and CoWoS tightness may persist into 2027. Fifth, mature-node competition is becoming more rational, and combined with demand recovery and some capacity being repurposed for advanced packaging, supply-demand is expected to improve over the next two to three years.

Analysis framework

The report quantifies industry opportunities from 2026 to 2030 by combining bottom-up chip and rack value-content estimates, per-GW server capacity demand models, advanced packaging market share stress tests, Google TPU shipment scenario analysis, process and packaging technology comparisons, supply-chain capacity tracking, and company valuation and earnings forecasts.

Methodology notes

• Demand quantificationPer-GW AI server demand model

  Converting data center power scale into chip, wafer, and packaging demand

  Based on different GPU, ASIC, and CPU architectures, the report estimates the N3 or N2 wafer capacity, CoWoS capacity, wafer fabrication equipment investment, and TSMC revenue opportunity required for each additional 1GW server project.

• Scenario analysisAdvanced packaging market share stress test

  Testing the impact of competitors’ share gains on the leader’s market position

  By assuming Intel EMIB-T gains 13% or 18% industry share in 2030, respectively, the report assesses the resilience of TSMC’s advanced packaging share and revenue growth.

• Technology comparisonMulti-dimensional comparison of EMIB-T and CoPoS

  Comparing technology routes in terms of performance, cost, yield, scalability, and ecosystem

  The report compares the two platforms in terms of architectural complexity, interconnect density, power integrity, ultra-large package capability, mass production timing, manufacturing risk, and customer fit scenarios.

• Supply-chain analysisTPU dual-route share model

  Breaking down Google TPU shipments and revenue opportunities by supplier

  Combining TSMC capacity constraints, Broadcom and MediaTek design service fees, and Google demand growth, the report estimates the two suppliers’ TPU shipment shares and sales scale in different years.

• Fundamentals and valuationBottom-up earnings and valuation analysis

  Mapping end demand, capacity, value content, and market share to company earnings

  Using company revenue, EPS growth, valuation multiples, target prices, and industry relative valuations, the report screens major beneficiaries of AI growth and supply-demand improvement.

Asset mapping & comparison

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

• TSMC（2330.TW）

  Core beneficiary of cloud AI advanced nodes and advanced packaging

  Strengths

  Has technological leadership, scale advantage, a deep customer ecosystem, and the strongest execution track record in N3, N2, and advanced packaging; leading-node share is expected to remain above 80%.

  Weaknesses

  Large capital expenditure scale, and advanced packaging expansion involves challenges such as warpage control and maturity of new toolchains.

  Comparison

  Compared with Intel EMIB-T, CoPoS has higher interconnect density, better signal and power integrity, and richer integration capability, but higher cost and manufacturing complexity.

  Risks

  Slower AI capital expenditure, lower-than-expected N3 or N2 demand, advanced packaging yields below expectations, competitors’ share gains, and geopolitical risks.

• MediaTek（2454.TW）

  Beneficiary of Google TPU dual-supplier route and cloud and edge AI growth

  Strengths

  Has a cost advantage in design service fees, is expected to achieve structural share gains in Google TPU projects, and has the opportunity to shift from a traditional fabless chip company to a cloud and edge AI platform supplier.

  Weaknesses

  Customer and project concentration may rise after rapid TPU business scale-up, and revenue scenarios are sensitive to shipment scale assumptions.

  Comparison

  Broadcom remains an important partner for Google, but MediaTek’s service fee per TPU is estimated to be about 50% lower, which is favorable for expanding its share.

  Risks

  Google project schedule delays, TSMC capacity constraints, Broadcom competition, mass production execution below expectations, and TPU demand below forecasts.

• ASE（3711.TW）

  Beneficiary of advanced packaging and next-generation cloud and edge AI assembly and test demand

  Strengths

  The rising importance of advanced packaging helps increase value added per chip, improve the business mix, and strengthen margins; the earnings upgrade cycle is still in its early stage.

  Weaknesses

  The business still has semiconductor cyclicality, and early-stage capacity expansion may face utilization and cost pressures.

  Comparison

  As a leading independent OSAT company, it can capture spillover demand from advanced packaging and complement the integrated packaging ecosystem led by TSMC.

  Risks

  Slower advanced packaging demand, customers building in-house capacity, price competition, and returns on capital expenditure below expectations.

• UMC（2303.TW） and PSMC（6770.TW）

  Major beneficiaries of mature-node supply-demand improvement

  Strengths

  More rational competition in mainland China, demand recovery, and TSMC shifting some mature-node capacity resources toward advanced packaging are expected to improve industry supply-demand and pricing over the next two to three years.

  Weaknesses

  Technology nodes are mature, product differentiation is relatively low, and exposure to consumer electronics and traditional cyclical demand is relatively sensitive.

  Comparison

  Compared with advanced-node companies, growth elasticity comes more from utilization and price recovery rather than structural value-content improvement from process upgrades.

  Risks

  Mainland China capacity additions exceed expectations, consumer demand remains weak, price recovery is below expectations, and capacity utilization declines.

• Aspeed（5274.TWO）

  Beneficiary of server market expansion and higher silicon content per unit

  Strengths

  Growth in AI server volume, strong demand for server CPUs, and rising value content per unit from management chips can expand its potential market.

  Weaknesses

  Valuation is relatively high, and growth expectations already reflect strong server demand.

  Comparison

  Compared with foundry and packaging companies, its opportunity comes more directly from server platform volume and growth in chip content per unit.

  Risks

  AI server buildout delays, customer concentration, intensifying competition, and high-valuation pullback.

• Intel

  Second source for EMIB-T advanced packaging and potential competitor to TSMC

  Strengths

  The EMIB-T architecture is simpler, does not require a large interposer, and has cost potential, ultra-large package scalability, and an earlier target mass production timeline.

  Weaknesses

  Mass production yield for the integration of large substrates and silicon bridges still needs to be validated, and it faces challenges in integrating front-end foundry and back-end packaging.

  Comparison

  EMIB-T is suitable for customers that emphasize cost, ultra-large packages, and supply-chain diversification; CoPoS is more suitable for AI products pursuing the highest performance and integration density.

  Risks

  Substrate yields fail to improve to high levels, commercial mass production delays, limited customer adoption, and insufficient ecosystem maturity.

Key data

• TSMC target price and reference share priceNT$3,650 / NT$2,365The target price corresponds to a Buy rating; the reference share price is as of August 6, 2026, implying upside of approximately 54%.
• Advanced-node capacity demand per 1GWApproximately 2-5kwpmCovers N3 or N2 capacity required for AI accelerators, CPUs, and networking chips.
• CoWoS capacity demand per 1GWApproximately 3-6kwpmSpecific demand varies with NVIDIA, AMD, and custom ASIC architectures.
• Logic wafer fabrication equipment investment per 1GWUS$1-2bnIncludes front-end process and CoWoS-related capacity investment.
• Revenue opportunity for TSMC per 1GWUS$1-2bnEquivalent to approximately 1.0-1.5% potential incremental sales.
• TSMC value content for next-generation NVIDIA platformsUS$1.4-1.9bn/GWPotential revenue corresponding to Rubin Ultra and Feynman, higher than approximately US$1.1bn/GW for Blackwell Ultra and Rubin.
• N2 foundry demandApproximately 220kwpmUBS expects this level to be reached in 2028, higher than foundry demand in the third year of N3 mass production.
• TSMC advanced packaging stress test2030E share of approximately 65-70%Assumes Intel EMIB-T reaches US$16bn in sales and a 13% industry share.
• TSMC advanced packaging low scenario2030E share of 62%Even if Intel gains an 18% share, TSMC’s advanced packaging business could still achieve a 47% CAGR from 2026 to 2030.
• Total TPU shipment forecast2026E at 4.8m; 2027E at 10.7mDemand comes from Google internal workloads, frontier model development, external customers, and cloud services.
• MediaTek TPU opportunity2027E sales of US$18bnScenario analysis assumes MediaTek ships approximately 4m TPU v8t units in 2027, accounting for about 45% of its business.
• MediaTek design service cost advantageApproximately 50% lower service fee per TPUThe report estimates its design service fee per TPU is about half that of Broadcom.
• Year-to-date market performanceSOX up 70%, covered portfolio up 68% on averageOver the same period, the Dow rose 12% and TAIEX rose 53%; TSMC, MediaTek, and UMC rose 53%, 174%, and 147%, respectively.

Impact & implications

Expansion in cloud AI investment will simultaneously drive demand for advanced logic nodes, advanced packaging, testing, substrates, equipment, materials, and high-speed interconnects, and may extend the industry upcycle. TSMC’s revenue elasticity comes not only from wafer volume growth, but also from process upgrades, an increase in the number of GPUs per rack, and higher value content from CoWoS and future panel-level packaging. Custom ASICs and TPUs have higher power efficiency per unit, but each GW may require more chips, thereby similarly amplifying wafer and packaging demand. For investors, opportunities will spread from core foundry leaders to design services, OSAT, equipment, and mature-node companies, but differentiation is needed by technology execution capability, yield ramp, and customer concentration.

Risks

• Hyperscale cloud service providers and AI developers reduce or delay data center capital expenditure.
• N3, N2, CoWoS, and CoPoS capacity expansion is faster than end demand, causing supply-demand to shift from tightness to oversupply.
• Yield ramp for large substrates, warpage, silicon bridge integration, and new equipment toolchains in advanced packaging falls short of expectations.
• Intel, Samsung Foundry, or other competitors gain greater-than-expected share in leading nodes and advanced packaging.
• Key products such as Google TPU, NVIDIA Rubin, or Feynman are delayed, reducing the pace at which related supply-chain revenue is realized.
• Mature-node demand recovery falls short of expectations, or mainland Chinese manufacturers resume aggressive capacity expansion and pricing strategies.
• Customer concentration, export controls, geopolitical risks, and supply-chain disruption risks.
• Valuations of some beneficiary names are already at relatively high levels, and earnings misses could trigger significant pullbacks.

What to watch

• N3 demand in 2H26 and the mass production progress of NVIDIA Rubin, Google TPU v8i, and v8t.
• TSMC CoWoS capacity expansion, order visibility, and whether supply tightness extends into 2027.
• Whether N2 demand can reach approximately 220kwpm in 2028, and whether TSMC’s leading-node share can remain above 80%.
• Intel EMIB-T commercialization progress in 2H27, substrate mass production yields, and Google TPU v9 adoption.
• TSMC CoPoS mass production readiness in 2028, panel warpage control, and maturity of new equipment toolchains.
• MediaTek’s actual shipment share, revenue recognition, and long-term customer expansion in Google TPU projects.
• Actual power consumption, chip architecture, and changes in TSMC value content per rack for each GW AI server project.
• The maturity of co-packaged optics over the next three years and the inflection point for optical fiber replacing copper interconnects.
• Mature-node pricing, capacity utilization, discipline in new supply from mainland China, and recovery in consumer electronics demand.
• Earnings forecast revisions and valuation changes for TSMC, MediaTek, ASE, UMC, PSMC, and Aspeed.