Global I/O Tech Hardware & Semis: AI supply-chain demand remains intact, while memory inflation deepens the consumer-device downturn

Overview

Based on meetings with 40 companies during its APAC technology tour, UBS argues that AI infrastructure demand remains structurally strong across foundry, memory, advanced packaging, power, testing and semiconductor equipment. Consumer electronics is the principal offset: memory-price inflation and allocation constraints are weakening PC and smartphone unit demand into 2027.

Core views

UBS finds no evidence of the widely discussed AI slowdown in upstream supply chains. Foundry and memory suppliers continue to see demand strength extending for years, while hyperscalers are increasing requests for 2027 server DDR5 and enterprise SSDs. The report raises its 2026E NVIDIA rack forecast from 77,500 to 80,000, retains 2027E at 95,000, and identifies a further potential 20,000-30,000-rack upside if data-center readiness and hyperscaler or neocloud capital budgets improve. The 2027 base case combines roughly 20% rack-unit growth with Rubin rack ASPs of US$7-8 million, about 70% above Blackwell, supporting the report's view that NVIDIA sales could potentially double in 2027. Rubin volume is expected to begin ramping from November, while AMD Helios is expected to reach a higher-volume inflection in 2Q27. The AI build-out is increasingly limited by components rather than demand. UBS projects total accelerator units to increase 32% year on year in 2026E to 17.5 million and 40% in 2027E to 24.9 million; total die count is forecast to rise 44% and 65%, respectively, to 28.3 million and 46.6 million. Constraints span HBM, advanced foundry capacity, high-end PCBs and substrates, MLCCs and AI power ICs. AI power is a particularly important structural theme: the report estimates a US$91.1 billion AI-power market by 2030E, representing a 69% CAGR from 2025-30. It expects migration toward dedicated power sidecars and 400/800VDC architectures as rack power rises, although GPU platform timing remains a key swing factor. AI demand also supports conventional servers. UBS forecasts total server shipments of 23.8 million in 2026E and 27.7 million in 2027E, with conventional-server shipments growing 21% and 15% year on year to 22.6 million and 26.0 million. Memory availability may delay some near-term builds, but order books remain strong, and new AMD and Intel platforms are expected to support a further pickup in the first half of 2027. The report highlights rising demand for BMC chips, server DRAM modules and memory interfaces as agentic AI increases compute requirements. Memory remains a central transmission mechanism across the technology cycle. UBS expects Samsung's blended 3Q26E DRAM ASP to rise 27% quarter on quarter and SK Hynix's by 23%, followed by 4Q26E increases of 10% and 7%. It forecasts blended 2027E HBM ASP growth of 62% for SK Hynix and 93% for Samsung. Hyperscaler demand for DDR5 and enterprise SSDs remains above available supply, and the report estimates unconstrained enterprise-SSD bit demand could grow 100% year on year in 2027E, versus constrained growth nearer 40%, after more than 90% growth in 2026E. HBM supply is a constraint for China's AI-accelerator ecosystem; UBS notes that local HBM ramp timing, production yields and export controls remain uncertain. The same memory inflation is driving UBS's more cautious consumer-device outlook. It retains a 2026E PC-unit decline of 11% but cuts its 2027E forecast from 2% growth to a 4% decline, or 231 million units. The report cites DRAM contract prices up 766% and NAND prices up 471% since mid-2025, plus double-digit CPU price increases, as pressures on affordability and PC demand. UBS estimates 2026E PC units of 241 million, but higher ASPs of 14% support industry revenue growth of 1.4% to US$202 billion. It also lowered annual notebook-ODM shipments to 109.7 million in 2026E, down 15% year on year, after reducing its 3Q26E forecast to 25.1 million units, down 17% quarter on quarter and 26% year on year. UBS similarly cuts its 2027E global smartphone unit forecast to a 3% decline from prior 2% growth, after a projected 10% decline in 2026E; it forecasts 2% growth in 2028E. The report sees relative resilience in the iPhone supply chain, with 2026 builds estimated at roughly 240-260 million units and initial indications of growth in 2027. By contrast, Android demand is more exposed to memory inflation. Samsung is expected to benefit from captive memory supply, while Xiaomi and Transsion seek to manage costs through pricing and supply agreements. The semiconductor-equipment outlook remains notably bullish. Tokyo Electron reiterated a CY2026 WFE forecast of US$150 billion and CY2027 of US$190 billion, with possible upside and another growth year expected in CY2028. UBS notes stronger orders since June, robust demand for advanced logic from TSMC and Intel, and sustained memory-equipment spending as Samsung and SK Hynix may fill a clean-room phase every six months. China WFE demand is also supported by memory and logic expansion; UBS estimates combined CXMT and YMTC WFE spend of US$14.0 billion in 2026E and US$22.5 billion in 2027E. Equipment-component supply is a near-term constraint, though Naura expects the tightest period to be 4Q26 and easing from 1Q27. Advanced packaging, testing and supporting components are further beneficiaries of AI complexity. UBS forecasts CoWoS capacity of 270,000 wafers per month by end-2027E versus 160,000 at end-2026E. It expects ASE's LEAP revenue to exceed US$7.5 billion in 2027E, despite potential additional outsourcing competition. Testing demand benefits from chiplets, higher HBM integration and more wafer-level probing; Advantest raised its CY2026 SoC tester TAM midpoint to US$11.0 billion from US$9.1 billion. In passive components, Murata expects AI-server MLCC revenue to double year on year in FY27 and again in FY28, while it plans ¥80 billion of FY27/28 investment to expand MLCC capacity by 30-35%. UBS's preferred exposure remains concentrated in Memory, Semicap and MLCCs. It added Delta Electronics, D.I Corp and King Yuan to its APAC Tech Most Preferred list and removed GUC, Hansol Chem and Wiwynn. It removed Hua Hong from Least Preferred and added Realtek. The report's broader conclusion is that AI-led demand and supply bottlenecks support selected infrastructure and semiconductor supply-chain participants, whereas consumer-exposed hardware faces a more difficult volume environment.

Analysis framework

UBS synthesizes feedback from 40 APAC technology-company meetings with its supply-chain checks, company data and forecast models. It assesses demand by end market, traces capacity and component bottlenecks through the supply chain, revises unit, pricing, ASP, capacity and capital-expenditure assumptions, and then identifies preferred and least-preferred sector exposures.

Methodology notes

• Industry AnalysisSupply-demand framework

  Supply-demand analysis across AI infrastructure, memory, equipment and consumer devices.

  The report compares demand requests with available capacity and component supply to explain pricing, shipment constraints and the relative outlook for AI infrastructure versus PCs and smartphones.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  AI supply-chain transmission from hyperscaler demand to foundry, memory, packaging, testing, power and component suppliers.

  UBS uses supplier feedback to link higher accelerator and server demand to downstream requirements for HBM, DDR5, SSDs, CoWoS, PCBs, MLCCs, power systems and manufacturing tools.

• Industry AnalysisVolume-price decomposition

  Separating unit demand from ASP and content growth.

  The report explains revenue implications through changes in rack units, device units, rack ASPs, memory ASPs and rising component content per server or rack.

Asset mapping & comparison

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

• Delta Electronics (2308.TW)

  Most Preferred; AI server-power beneficiary

  Strengths

  Server-power business is expected to grow more than 70% in 2026, supported by AI GPU and ASIC launches and higher power content.

  Weaknesses

  Near-term sales are constrained by power-IC and capacitor lead times.

  Comparison

  UBS says Delta is outpacing Lite-On but is more exposed to supply constraints because of its larger revenue scale.

  Risks

  Supply-demand imbalance may constrain deliveries before capacity catches up.

• Samsung Electronics (005930.KS) and SK Hynix (000660.KS)

  Most Preferred; memory-cycle beneficiaries

  Strengths

  Rising DRAM and HBM pricing, strong server DDR5 and enterprise SSD demand, and elevated capex plans support the outlook.

  Comparison

  UBS forecasts 2027E blended HBM ASP growth of 93% for Samsung and 62% for SK Hynix.

  Risks

  Memory-cycle assumptions depend on sustained AI and hyperscaler demand.

• TSMC (2330.TW)

  Most Preferred; leading-edge AI/HPC foundry beneficiary

  Strengths

  UBS sees robust AI/HPC demand visibility through 2029-30 and strong demand for advanced-node equipment.

  Comparison

  Leading-edge capacity pressure is also directing some demand toward mature foundries.

  Risks

  Data-center construction readiness could affect near-term demand timing.

• Tokyo Electron (8035.T)

  Most Preferred; semiconductor-capital-equipment beneficiary

  Strengths

  The company sees stronger orders, visibility into 2028 and demand across DRAM, China and advanced logic.

  Weaknesses

  Equipment supply chains are the main scaling constraint.

  Comparison

  UBS highlights broad WFE strength across Japanese, US, Korean and Chinese equipment suppliers.

  Risks

  Execution depends on suppliers' ability to expand component and subsystem capacity.

• ASE (3711.TW)

  Most Preferred; advanced-packaging beneficiary

  Strengths

  UBS expects LEAP revenue above US$7.5 billion in 2027E as full-process CoWoS and probing outsourcing ramp.

  Comparison

  A further OSAT wafer-on-substrate outsourcing partner could increase competition.

  Risks

  Additional TSMC outsourcing could dilute competitive positioning.

• MediaTek (2454.TW)

  Most Preferred; cloud-ASIC beneficiary

  Strengths

  Google TPU v8t is expected to contribute more than US$2 billion of revenue in 2026 and US$12-16 billion in 2027.

  Weaknesses

  Smartphone demand remains pressured by memory costs.

  Comparison

  UBS views cloud ASICs as increasingly more important than the smartphone business.

  Risks

  Future Google ASIC supplier mix may affect content and share.

• Murata Manufacturing (JP6981)

  Most Preferred; AI-server MLCC beneficiary

  Strengths

  UBS expects AI-server MLCC revenue to double in FY27 and again in FY28, alongside accelerated capacity investment.

  Weaknesses

  Product mix between general-purpose and high-end MLCCs remains variable.

  Comparison

  Murata and SEMCO both indicate a potentially improving MLCC-pricing environment.

  Risks

  Demand assumptions depend on continued AI-server infrastructure expansion.

• Realtek (2379.TW)

  Least Preferred; consumer-semiconductor exposure

  Strengths

  UBS notes sales could hold up better than feared through price increases and networking, server and automotive initiatives.

  Weaknesses

  Consumer and PC demand is weakening in the second half of 2026.

  Comparison

  The report contrasts stronger networking and automotive segments with softer PC and consumer semiconductors.

  Risks

  Sustained memory inflation could further pressure consumer-device demand.

Key data

• NVIDIA rack forecast80,000 in 2026E; 95,000 in 2027E2026E increased from 77,500; 2027E maintained.
• AI accelerator-unit growth+32% in 2026E to 17.5 million; +40% in 2027E to 24.9 millionUBS expects total die count to grow faster, at +44% and +65%.
• AI power marketUS$91.1 billion in 2030EUBS forecasts a 69% CAGR from 2025-30.
• PC unit outlook-11% YoY in 2026E; -4% YoY in 2027E to 231 million units2027E was cut from prior +2% growth.
• Smartphone unit outlook-10% YoY in 2026E; -3% YoY in 2027E; +2% YoY in 2028E2027E was cut from prior +2% growth.
• CoWoS capacity270,000 wafers per month by end-2027EVersus 160,000 wafers per month by end-2026E.
• Tokyo Electron WFE forecastUS$150 billion in CY2026; US$190 billion in CY2027The company sees possible upside and another growth year in CY2028.

Impact & implications

UBS sees the strongest sector support in AI infrastructure beneficiaries, especially memory, semiconductor equipment, advanced packaging, testing, power, MLCCs, substrates and PCBs. Memory scarcity is simultaneously a positive pricing driver for suppliers and a headwind for consumer-device volumes, creating a split outlook within APAC technology hardware.

Risks

• The semiconductor industry is cyclical, so both upside and downside depend on macroeconomic conditions.
• Semiconductor-capital-equipment companies face risks of market-share loss from product timing, performance, design or distribution.
• Handset supply-chain companies face cyclical macro and inventory risks as well as intense competition requiring continued product innovation and R&D investment.
• AI infrastructure growth can be constrained by memory, advanced-foundry, packaging, PCB, MLCC and power-component availability.

What to watch

• Whether hyperscalers continue to increase 2027 AI-server, DDR5 and enterprise-SSD demand requests.
• Data-center readiness and hyperscaler or neocloud capex budgets, which determine potential upside to the 2027 NVIDIA rack forecast.
• HBM availability, Chinese local-HBM qualification and yield progress, and the effect of export controls on domestic AI accelerators.
• Memory-price inflation and allocation trends affecting PC and smartphone demand in 2027.
• CoWoS capacity expansion, WFE order momentum and equipment supply-chain constraints.