Asia semiconductor sector Report Interpretation

Overview

UBS’s Asia semiconductor sector report links accelerating cloud-AI infrastructure spending to tighter leading-edge foundry and advanced-packaging supply, a broader recovery in mature-node foundries, and new opportunities in custom AI chips, co-packaged optics and related equipment. Its preferred exposure is concentrated in TSMC and selected Taiwanese foundry, fabless, packaging, testing, wafer and equipment companies.

Core views

UBS’s central thesis is that cloud-AI demand is broadening beyond GPUs into custom ASICs, server CPUs, networking and advanced packaging, creating a multi-year capacity constraint across the semiconductor chain. It reiterates Buy on TSMC, expecting N2 foundry demand of about 220kwpm by 2028 and TSMC to retain more than 80% leading-edge foundry dominance. UBS expects N3 capacity to be the larger bottleneck in 2027: even with N3 capacity reaching 190kwpm by end-2027, utilization is projected above 100%, while cloud-AI accelerators, CPUs and networking rise to 72% of N3 demand in 2027E from 35% in 2026E. The report translates announced AI data-centre projects into semiconductor demand. For each incremental 1GW server build-out, UBS estimates roughly 2-5kwpm of N3/N2 capacity, 3-6kwpm of CoWoS capacity, US$1-2bn of logic WFE spending and US$1-2bn of TSMC revenue opportunity, or about 1.0-1.5% of sales. TSMC’s revenue opportunity per GW varies by architecture: US$1.1bn for Blackwell Ultra/Rubin and US$1.4-1.9bn for Rubin Ultra/Feynman; ASICs can require more leading-edge capacity and represent 10-11% of rack value for TSMC versus 4-6% for GPUs. UBS estimates TSMC’s NVIDIA AI-GPU sales at US$19.4bn in 2026E and US$28.8bn in 2027E, with NVIDIA representing about 20.7% and 21.3% of TSMC sales, respectively. Advanced packaging is presented as an essential performance enabler as front-end scaling alone becomes insufficient for cloud AI. UBS forecasts industry CoWoS capacity at 160kwpm at end-2026 and 250kwpm at end-2027, comprising 180kwpm at TSMC and 70kwpm at OSATs in 2027. It forecasts NVIDIA CoWoS demand to rise 56% in 2027 to 1,238k wafers, while TPU requirements grow 142% year on year and AMD demand potentially triples. Intel’s EMIB-T is expected to provide a meaningful second source rather than displace TSMC’s CoPoS: in UBS’s stress case, Intel reaches US$16bn and 13% industry share in 2030E while TSMC retains 65-70%; even with Intel at 18%, TSMC could retain 62% and grow advanced-packaging sales at a 47% 2026-30E CAGR. Custom AI silicon is another major demand driver. UBS expects Google to ramp both Broadcom’s v8i and MediaTek’s v8t through H2 2026 to 2027 because of tight TSMC capacity, with dual tracks retained for TPU v9. It forecasts total TPU shipments of 4.8m in 2026E and 10.7m in 2027E, and MediaTek’s share rising from 9% to 38%; a scenario has MediaTek shipping 4m v8t units in 2027, generating US$18bn of sales and 45% of its business. The report estimates MediaTek’s per-TPU-die service fee at 50% below Broadcom’s, supporting share gains. It also expects server CPU shipments to rise from 23m in 2025 to 63-70m in 2030, with ARM share reaching 42% and AMD 29%; UBS estimates this creates a US$44.1bn TSMC revenue opportunity by 2030. On optical networking, UBS expects fibre to increasingly replace copper as bandwidth and distance requirements rise. It forecasts the optical-transceiver market to grow from US$12bn in 2025E to US$32bn by 2030E, a 22% CAGR, with CPO reaching US$7bn and 20% of the market; its upside scenario assumes a 30% CAGR. Silicon-photonics sales are forecast to grow 35% annually in the base case to US$4-6bn in 2028-30. UBS sees CPO as supportive for TSMC, ASE, substrate makers, server ODMs and testing suppliers, but notes that traditional pluggable transceivers should remain the mainstream solution for the next two to three years. It expects CPO scale-out adoption from late 2026 and potential GPU scale-up adoption around 2029, while testing and optical alignment remain challenges. UBS is also positive on mature foundries. It expects more rational Chinese competition, TSMC’s reallocation of mature resources toward advanced packaging, and rising AI-server power-semiconductor demand to improve supply-demand balance over the next two to three years. It forecasts 8-inch utilization at 84% in 2026E and above 90% in 2027E, from 81% in 2025; 12-inch 28/40nm utilization should remain above 80% in 2026-27E despite consumer weakness. UBS expects selected UMC price increases of 5-10% in H2 2026 and another 10% more broadly in early 2027, while Chinese foundries may raise prices by about 10% or more on some products in 2026. Company implications follow from these supply-chain views. UBS highlights TSMC’s greater AI content, MediaTek’s TPU opportunity, ASE’s advanced-packaging and testing ramp, UMC’s mature-node and silicon-photonics opportunities, and PSMC’s improving memory, logic and 3D-AI-foundry mix. It expects ASE’s leading-edge advanced-packaging and testing revenue to rise from US$1.6bn in 2025 to US$3.67bn in 2026E and US$7.47bn in 2027E. It also identifies Aspeed, Hon Precision, KYEC, All Ring, GPTC, Gudeng, ASMPT and GlobalWafers as beneficiaries of expanding server, packaging, testing, wafer and optical demand, while maintaining Neutral views on several covered names including Hua Hong, SMIC, Vanguard, Faraday, Novatek, Parade, Realtek and Powertech.

Analysis framework

UBS combines bottom-up capacity, wafer-content, packaging and unit-demand models with supply-demand analysis across leading-edge and mature foundries. It translates AI data-centre gigawatts and server architectures into foundry, CoWoS, WFE and revenue requirements; compares competing packaging technologies; uses scenario analysis for TPU, CPO and company opportunities; and applies forward P/E or P/BV multiples to selected company price targets.

Methodology notes

• Industry AnalysisSupply-demand framework

  Foundry, CoWoS, silicon-wafer and optical-market supply-demand modelling.

  UBS compares projected capacity, utilization and end-demand to identify where supply remains constrained or pricing may improve.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  AI infrastructure demand traced through foundry, packaging, testing, optics, substrates and equipment.

  The report connects accelerator and server build-outs to the suppliers that receive wafer, packaging, tool or testing demand.

• Valuation methodsP/E and PEG Valuation

  Forward P/E-based price targets for many covered companies.

  UBS applies specified forward earnings multiples to its company earnings forecasts to derive target prices.

• Valuation methodsPB valuation

  Forward P/BV valuation for GlobalWafers.

  UBS uses a price-to-book multiple where it considers asset and cycle recovery relevant to valuation.

Asset mapping & comparison

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

• TSMC (2330.TW)

  Primary leading-edge foundry and advanced-packaging beneficiary of cloud AI, GPUs, ASICs and server CPUs.

  Strengths

  Expected N2 demand of about 220kwpm by 2028; >80% leading-edge share; tight N3 and CoWoS capacity; growing NVIDIA content.

  Weaknesses

  N2 ramp and overseas expansion could dilute margins into H2 2026.

  Comparison

  UBS expects TSMC to retain advanced-packaging leadership despite Intel EMIB-T becoming a second source.

  Risks

  Capacity expansion, technology migration, customer demand and competitor execution.

• MediaTek (2454.TW)

  Covered fabless beneficiary of Google TPU design-service share gains and edge AI.

  Strengths

  UBS forecasts US$18bn TPU sales in 2027E and strong operating leverage.

  Weaknesses

  TPU v9 execution and supply-chain capacity remain relevant constraints.

  Comparison

  MediaTek’s per-die TPU service fee is estimated at 50% below Broadcom’s.

  Risks

  TPU ramp, customer adoption and smartphone-market execution.

• ASE (3711.TW)

  Covered advanced-packaging and testing beneficiary.

  Strengths

  LEAP and testing revenue forecast at US$3.67bn/US$7.47bn in 2026E/2027E; margin expansion potential.

  Weaknesses

  Requires elevated capex and successful capacity ramp.

  Comparison

  Benefits as CoWoS supply diversifies beyond TSMC.

  Risks

  Advanced-packaging demand, execution and capex returns.

• UMC (2303.TW)

  Covered mature-foundry recovery beneficiary.

  Strengths

  Potential pricing gains, Intel partnership, differentiated 22nm and silicon-photonics optionality.

  Weaknesses

  Exposure to mature-node and consumer demand.

  Comparison

  UBS favors UMC within mature foundries because of improving fundamentals and supply-demand.

  Risks

  Consumer weakness, Chinese competition and timing of new opportunities.

• PSMC (6770.TW)

  Covered memory and mature/3D-AI-foundry beneficiary.

  Strengths

  DRAM price increases, improving mature-foundry dynamics and 3D AI foundry growth.

  Weaknesses

  DRAM-price realization has a four-to-five-month lag.

  Comparison

  UBS recently upgraded PSMC on better fundamentals and supply-demand.

  Risks

  Memory-cycle and 3D-AI-foundry execution.

• Aspeed (5274.TWO)

  Covered server-management-chip beneficiary.

  Strengths

  Server BMC TAM forecast to rise from 21m units in 2025 to 77m in 2030; new AST2700 ramp.

  Weaknesses

  Dependent on server CPU, accelerator and ASIC deployment timing.

  Comparison

  UBS sees accelerating TAM and silicon-content expansion.

  Risks

  Platform rollout and AI-server demand.

Key data

• TSMC N3 utilization100%+ in 2026-27EUBS expects N3 to remain a larger bottleneck than CoWoS despite expansion to 190kwpm by end-2027.
• AI share of N3 demand72% in 2027EUp from 35% in 2026E for server accelerators, CPUs and networking.
• CoWoS industry capacity160kwpm end-2026E; 250kwpm end-2027ETSMC/OSAT capacity is forecast at 180/70kwpm in 2027E.
• TSMC NVIDIA AI-GPU salesUS$19.4bn in 2026E; US$28.8bn in 2027ENVIDIA is estimated to account for 20.7%/21.3% of TSMC sales.
• Google TPU shipments4.8m in 2026E; 10.7m in 2027EMediaTek share is forecast to increase from 9% to 38%.
• Optical transceiver marketUS$12bn in 2025E to US$32bn in 2030EBase-case 22% CAGR; CPO could reach US$7bn and 20% of the market in 2030E.
• Mature-foundry utilization8-inch 84% in 2026E and >90% in 2027EVersus 81% in 2025; UBS expects firmer pricing and improving supply-demand.

Impact & implications

UBS views cloud-AI spending as a broader semiconductor-cycle driver rather than a GPU-only theme. The report’s implications favor leading-edge foundry, advanced packaging, custom ASIC design, testing, CPO-related equipment and selected mature-node producers, while recognizing that adoption timing, capacity execution and consumer demand still differentiate outcomes across companies.

Risks

• Technology-company operating models and financial results can be highly volatile and difficult to forecast.
• The report’s forecasts depend on numerous assumptions; different assumptions could produce materially different results.
• CPO remains early stage, with uncertain individual-company revenue timing and significant testing and optical-alignment challenges.
• Consumer weakness could affect parts of mature-node demand.

What to watch

• N3/N2 and CoWoS capacity additions and utilization through 2027.
• Google TPU v8i/v8t ramps, MediaTek share gains and TPU v9 execution.
• NVIDIA Rubin, Rubin Ultra and Feynman production ramps and their packaging content.
• CPO deployment timing in AI data centres, particularly scale-out from late 2026 and scale-up around 2029.
• Mature-foundry price increases, utilization and the pace of Chinese capacity competition.
• Advanced-packaging capex and capacity expansion at TSMC, ASE and other OSATs.