Asia semiconductors: Cloud AI demand is broadening semiconductor supply constraints beyond accelerators into foundry, packaging, optics and mature nodes

Overview

This sector deep dive examines how expanding Cloud AI infrastructure demand affects Asian semiconductor supply, from advanced foundry and packaging through ASIC design, optical interconnects, testing, equipment, wafers and mature-node foundries. UBS’s central conclusion is that the AI build-out is creating multi-year capacity and content opportunities that extend well beyond GPU makers.

Core views

UBS argues that Cloud AI demand is moving semiconductor constraints from a narrow accelerator issue into a broader foundry-and-packaging cycle. It expects TSMC’s N3 capacity to be the larger bottleneck than CoWoS in 2027, with N3 loading above 100% in 2026-27 even as installed capacity reaches 210kwpm by end-2027. Server accelerators, CPUs and networking are forecast to represent 73% of N3 demand in 2027E, versus 36% in 2026E. UBS also raises its 2028 N2 foundry TAM estimate to 312kwpm from 230kwpm, citing faster adoption in server CPUs and AI accelerators. It forecasts TSMC N2 installed capacity of 90kwpm by end-2026 and 140kwpm by end-2027, while maintaining that TSMC can preserve more than 80% leading-edge foundry dominance through technology leadership and customer breadth. The report translates AI data-centre construction into semiconductor demand. For every incremental 1GW server build-out, UBS estimates approximately 2-5kwpm of N3/N2 capacity, 3-6kwpm of CoWoS capacity, US$1-2bn of logic WFE spending, and US$1-2bn of potential TSMC revenue, equivalent to about 1.0-1.5% sales upside. TSMC’s revenue opportunity per GW rises from about US$1.1bn for Blackwell Ultra/Rubin to US$1.4-1.9bn for Rubin Ultra/Feynman as leading-edge process content, GPU count per rack and advanced packaging increase. ASICs can consume more leading-edge capacity per GW than GPUs: Google’s Broadcom-designed TPU v7p is estimated to require about 5kwpm of N3 per GW, while AMD’s MI455 may require around 1kwpm of N2. UBS estimates TSMC content at roughly 5% of a GB200 or VR200 rack and expects Nvidia to account for 20.3% and 21.6% of TSMC sales in 2026E and 2027E, respectively. Advanced packaging remains a central beneficiary. UBS forecasts industry CoWoS capacity to expand from 160kwpm at end-2026E to 270kwpm at end-2027E, comprising 180kwpm at TSMC and 90kwpm at OSATs. Demand is supported by Nvidia, whose CoWoS demand is forecast to grow 61% in 2027 to 1,290k wafers, as well as Google TPU demand, expected to rise 185% year on year in 2027E, and AMD demand, which could triple with MI450 and Venice CPU ramps. TSMC’s advanced-packaging sales are forecast to rise from US$16.2bn in 2026E to US$87.0bn in 2030E in UBS’s stress-test case. Even if Intel EMIB-T reaches US$16bn of sales and 12% industry share by 2030E, UBS estimates TSMC could retain 65-70% packaging share; in a lower case where Intel gains 17%, TSMC would still hold 63% and grow packaging sales at a 49% CAGR from 2026-30E. UBS views Intel EMIB-T and TSMC CoPoS as likely to coexist rather than create a winner-takes-all outcome. EMIB-T uses embedded silicon bridges and removes the interposer, offering a simpler architecture, potential cost advantage and scalability to very large packages. However, the report flags manufacturing challenges: substrate yield is estimated at around 50% versus 80% for HPC ABF substrates, and Intel must improve yields toward the high 90% range. CoPoS is expected to target FY28 mass production, with Nvidia Feynman as an anticipated first major customer. UBS considers CoPoS better positioned for maximum performance because of higher interconnect density, signal and power integrity, and a deeper manufacturing ecosystem, while noting warpage and tool-chain readiness risks. Google TPU demand is another major source of supply-chain expansion. UBS expects Google to run Broadcom v8i and MediaTek v8t projects simultaneously from H226 through 2027 amid tight TSMC capacity, and to retain dual tracks into the TPU v9 ramp in 2028. Total TPU shipments are forecast at 4.9m units in 2026E and 10.1m in 2027E. Broadcom’s volume share is estimated to decline from about 90% in 2026E to 60% in 2027E, while MediaTek’s rises from 10% to 40%. UBS estimates MediaTek could ship 4.5m TPU v8t units in 2027, producing US$18bn in sales and representing 46% of its business. Its case rests on a lower per-die design-service fee than Broadcom, Google’s internal TPU use, frontier-model training demand, external demand and possible cloud penetration. Co-packaged optics is presented as a longer-duration extension of AI infrastructure demand. Fibre is replacing copper over longer data-centre connections because it offers more bandwidth, faster transmission, longer reach and better resistance to environmental factors and interference. CPO brings optical engines close to the switch or xPU to improve bandwidth, signal integrity and power efficiency. UBS expects pluggable transceivers to remain mainstream for the next two to three years, with broader GPU scale-up CPO adoption potentially emerging in 2029E. Its base case forecasts optical-transceiver sales rising from US$12bn in 2025E to US$32.3bn in 2030E, a 22% CAGR, with CPO reaching US$6.5bn, or 20% of the market. Silicon-photonics sales are forecast to grow at a 35% CAGR to US$6.1bn by 2030E. TSMC, ASE/SPIL, KYEC, All Ring, Hon Precision, Chroma and equipment providers are positioned at different stages of the value chain. The report also turns constructive on mature foundry conditions over the next two to three years. UBS forecasts 8-inch utilisation to rise to 95% in 2027 from 85% in 2026, supported by server PMIC demand, while 12-inch 28/40nm utilisation improves to 88% from 84%, despite near-term consumer and smartphone headwinds. It attributes the improvement to slower mature-node expansion, more rational Chinese competition as capacity spending shifts to leading edge, TSMC reallocating mature resources toward advanced packaging, and rising AI-server power-semiconductor demand. This supports its positive views on UMC and PSMC. Among company-specific conclusions, UBS reiterates Buy on TSMC with a NT$3,650 target, citing Cloud AI demand, N2 growth and a valuation discount to the SOX. It highlights MediaTek as a key Buy at NT$7,300 because of expanding TPU design-service share and edge-AI opportunities. ASE is favored for CoWoS outsourcing, full-process packaging and testing, with leading-edge packaging and testing revenue forecast to rise from US$3.67bn in 2026E to US$7.47bn in 2027E. UMC is expected to benefit from mature-node recovery, pricing, advanced packaging, an Intel collaboration and silicon photonics. The report also identifies opportunities for Aspeed in server BMCs, KYEC in Cloud AI testing, GlobalWafers in tighter wafer supply, and All Ring, GPTC, Gudeng and ASMPT in advanced-packaging equipment and materials.

Analysis framework

UBS combines bottom-up capacity, wafer-content, rack-build and supply-chain estimates with company guidance and scenario analysis. It maps AI data-centre power deployment into leading-edge wafer, CoWoS, WFE and revenue requirements; compares packaging architectures; models supply-demand balances for foundry and wafers; and applies forward valuation multiples to company targets.

Methodology notes

• Industry AnalysisSupply-demand framework

  Foundry, CoWoS and wafer supply-demand modelling

  UBS compares projected capacity additions with demand from AI accelerators, CPUs, networking and mature-node applications to assess utilisation, tightness and pricing.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  AI infrastructure value-chain mapping

  The report traces Cloud AI spending from data-centre build-outs through foundry, packaging, testing, substrates, optics and equipment suppliers.

• Valuation methodsP/E and PEG Valuation

  Forward P/E valuation

  UBS bases several company price targets on specified forward P/E multiples, generally using 2027E or 2027-28E earnings.

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 infrastructure.

  Strengths

  Expected N3/N2 demand, technology leadership, CoWoS and CoPoS roadmap, and rising content in Nvidia platforms.

  Weaknesses

  N2 ramp and overseas expansion may dilute margins in H226.

  Comparison

  UBS sees TSMC retaining advanced-packaging leadership even as Intel EMIB-T gains share.

  Risks

  Capacity constraints and execution of new packaging and node ramps.

• MediaTek (2454.TW)

  Potential share gainer in Google TPU design services and edge AI.

  Strengths

  Dual-track TPU ramp, lower service fee per TPU die than Broadcom, and flagship smartphone SoC growth.

  Weaknesses

  TPU opportunity is dependent on program execution and capacity availability.

  Comparison

  MediaTek is forecast to gain TPU volume share while Broadcom remains a major partner.

  Risks

  TPU program timing, customer validation and supply-chain capacity.

• ASE (3711.TW)

  Advanced packaging and testing beneficiary.

  Strengths

  CoWoS outsourcing, full-process services, testing expansion and margin improvement potential.

  Weaknesses

  Requires elevated capital expenditure and successful capacity ramp.

  Comparison

  Can gain opportunity as CoWoS supply diversifies beyond TSMC.

  Risks

  Slower advanced-packaging capacity expansion.

• UMC (2303.TW)

  Mature-foundry beneficiary.

  Strengths

  Improving utilisation, potential pricing increases, Intel collaboration and silicon-photonics opportunity.

  Weaknesses

  Near-term consumer and smartphone demand remains a headwind for some mature nodes.

  Comparison

  UBS expects improving dynamics as Chinese peers focus more on leading edge.

  Risks

  Mature-node demand recovery and partnership execution.

• GPTC (3131.TWO)

  Advanced-packaging wet-process equipment supplier and UBS top pick in the space.

  Strengths

  High ASE share, rising tool shipments and higher-value panel-level, SoIC and CPO tools.

  Weaknesses

  Growth depends on customer capacity plans.

  Comparison

  Benefits from both TSMC and OSAT expansion, with especially high share at ASE.

  Risks

  Slower CoWoS expansion, technological competition, geopolitical uncertainty and weaker AI growth.

Key data

• TSMC N3 loading100%+ in 2026-27EExpected despite capacity expansion to 210kwpm by end-2027.
• N2 foundry TAM312kwpm in 2028ERaised from 230kwpm on stronger server CPU and AI-accelerator adoption.
• AI server build-out impactUS$1-2bn TSMC revenue per 1GWUBS also estimates 2-5kwpm of leading-edge and 3-6kwpm of CoWoS capacity per GW.
• CoWoS industry capacity160kwpm at end-2026E; 270kwpm at end-2027ETSMC/OSAT capacity projected at 180/90kwpm by end-2027.
• Google TPU shipments4.9m units in 2026E; 10.1m in 2027EMediaTek share is forecast to rise from 10% to 40%.
• Optical transceiver marketUS$12.0bn in 2025E to US$32.3bn in 2030EBase-case 22% CAGR; CPO reaches US$6.5bn and 20% of the market in 2030E.
• Mature foundry utilisation8-inch 95%; 12-inch 28/40nm 88% in 2027EVersus 85% and 84%, respectively, in 2026E.

Impact & implications

UBS sees Cloud AI creating a multi-year expansion cycle for leading-edge foundry, advanced packaging, testing, optical interconnects and related equipment. It also expects the AI-driven demand mix and more disciplined mature-node supply to improve utilisation and pricing across selected mature foundries and component suppliers.

Risks

• UBS identifies slower-than-expected CoWoS capacity expansion as a downside risk for GPTC.
• UBS identifies competitors achieving technological advances as a downside risk for GPTC.
• UBS identifies geopolitical uncertainty as a downside risk for GPTC.
• UBS identifies worse-than-expected AI growth as a downside risk for GPTC.

What to watch

• TPU v9 “Humufish” tape-out expected by November 2026.
• Google system-level validation in H127 and visibility into TPU v9 performance and 2028 deployment volumes.
• Potential upward revisions to TPU-related guidance in October 2026 and January/February 2027 earnings calls.
• Intel EMIB-T execution and yield improvement.
• The pace of N3, N2, CoWoS and OSAT capacity expansion.
• The timing of CPO adoption in Nvidia networking and scale-up architectures.