APAC technology and semiconductor supply chain: UBS expects sustained AI demand to extend the APAC memory and semiconductor-equipment upcycle

Overview

The report examines whether AI infrastructure demand and the memory cycle can remain sustainable, then translates that view across APAC semiconductor, equipment, component and hardware stocks. UBS concludes that AI demand remains strong, memory undersupply should persist into 2028, and the most attractive sector exposures are memory, semiconductor equipment, MLCCs and selected Taiwan semiconductor companies.

Core views

UBS answers its central AI-cycle question affirmatively and says visibility is improving into 2028. TSMC, Samsung Electronics and SK Hynix continue to report medium-term demand visibility, while agentic AI is increasingly viewed as the next major adoption inflection. UBS expects additional 2027 upside for CPUs, including AMD Venice and Nvidia Vera, and for AI accelerators, particularly custom ASICs. It also raises 2027 operating-cash-flow forecasts for hyperscalers and neoclouds by 30%, which supports the view that the widening gap between cash generation and capital expenditure should remain manageable, although that funding gap is still the report's leading financial risk. The scale of infrastructure spending remains exceptional. UBS forecasts top-12 hyperscaler capex of US$1.0086tn in 2026E, US$1.4475tn in 2027E and US$1.6186tn in 2028E, representing growth of 98.5%, 43.5% and 11.8%, respectively. Total memory spending by hyperscalers is estimated at US$368.9bn in 2026E and US$933.2bn in 2027E, equal to 37% and 64% of base-case capex. The report therefore identifies affordability as the key limit to the thesis: memory-industry revenue is expected to approach US$1.64tn in 2027E, with more than half consumed by hyperscalers and neoclouds. In an upside capex scenario of US$1.8tn, memory would still represent 52% of 2027 capex. The memory conclusion is that near-term softness in non-AI demand has not broken the upcycle. OEMs still need to accept higher prices to secure supply, with mobile DRAM contract pricing estimated to rise about 40% QoQ in 3Q26E. UBS expects blended DRAM contract ASPs to increase 25% QoQ in 3Q26 and 9% in 4Q26, while NAND rises 20% and 8%, respectively. Current DRAM fulfillment is only 60%. Against constrained supply, UBS forecasts 2027 server DDR bit demand to rise 81% YoY and server SSD demand to rise 50%; unconstrained SSD demand could be twice the supplied amount. DRAM end-consumption growth is expected to accelerate from 21% in 2026E to 39% in 2027E, while NAND bit demand grows 24% in 2027E. UBS consequently extends the DRAM upcycle toward 2Q28 and the NAND upcycle toward 4Q27. HBM de-specification does not change the positive demand conclusion. Nvidia's VR300 move from HBM4E 12-Hi to HBM4 8-Hi is estimated to leave its 2027 HBM bit demand slightly higher. At the same time, stronger DDR demand is encouraging suppliers to reallocate wafers away from HBM, so UBS forecasts 2027 HBM bit supply growth of only 47%, below demand. Additional de-specification for ASIC products is possible, but higher volumes, better yields and potential custom HBM4E testing complexity make the effect on semiconductor-test demand neutral to slightly positive. UBS expects Samsung to reach 41% HBM bit share in 2027, ahead of SK Hynix at 37% and Micron at 22%; SK Hynix is not expected to chase share through aggressive pricing. Long-term agreements should make memory earnings less volatile without eliminating the cycle. UBS expects LTA coverage to reach about 50% of total industry bits by year-end, with 40% on fixed volume and pricing. Prepayments and callable provisions are expected to improve enforceability, while AI customers face a high opportunity cost if they abandon contracted capacity. UBS argues that LTAs should narrow pricing bands, lift through-cycle margins and add more than two percentage points to through-cycle return on equity. As cash generation rises, investor attention should increasingly shift from pricing alone toward shareholder returns at Samsung Electronics and SK Hynix. Beyond memory, UBS's preference order is semiconductor equipment, MLCCs and selected Taiwan semiconductor names. Advanced-logic wafer-fabrication-equipment demand is improving, Korean capex forecasts have risen, and tester demand benefits from greater CPU and ASIC volumes plus more intensive testing. Total WFE spending is forecast at US$158bn in 2026E, US$226bn in 2027E and US$274.5bn in 2028E. Memory WFE rises from US$74.5bn in 2026E to US$113.8bn in 2027E, with DRAM the largest contributor. The report sees beneficiaries across front-end equipment, logic and memory testers, packaging equipment, materials and outsourced assembly and test. Advanced packaging remains necessary to scale AI performance and power efficiency. Industry CoWoS capacity is forecast to rise from an average 133kwpm in 2026E to 233kwpm in 2027E, while utilization rises from 83% to 94%. TSMC's CoPoS panel-level packaging targets mass production in FY28, initially on 310mm by 310mm panels, with Nvidia Feynman expected as the first major customer and AMD and ASIC products following in 2029-30. Intel's EMIB-T is expected to ramp commercially in 2H27, with Google TPU v9 as the first major external product. These transitions expand opportunities for deposition, bonding, inspection, test, OSAT and substrate suppliers. Co-packaged optics is presented as a later-stage but substantial opportunity. CPO integrates optical engines with switch ASICs or accelerators to shorten copper connections, improve power efficiency and increase bandwidth density. Nvidia may begin using CPO Spectrum-X switches for scale-out connections from 2H26E, while scale-up adoption should initially be hybrid: NVLink switches move to CPO but GPU compute trays retain copper. Direct GPU scale-up adoption may mature around 2029E. UBS's base case puts the optical-transceiver market at US$32.3bn in 2030E, including US$6.5bn of CPO, while its upside case reaches US$55.2bn, including US$13.8bn of CPO. The AI strength contrasts with weak consumer devices. UBS forecasts PC units to decline 11% in 2026E and 4% in 2027E, while smartphones decline 10% in 2026E and 3% in 2027E before growing 2% in 2028E. Higher memory and component costs are pressuring demand and margins across PCs, Android smartphones, displays and consumer electronics. The report therefore favors companies with direct AI-server, memory, advanced-packaging, power, networking or test exposure and is more cautious on consumer-heavy hardware and display names. At the regional and subsector level, UBS is overweight Korea and Taiwan, neutral Japan and underweight China. It prefers IC substrates, leading-edge foundry, memory semiconductors, MLCCs and wafer-fabrication equipment, while taking a less favorable view of displays, silicon wafers, PCs, trailing-edge foundries, smartphones, OSATs and server ODMs. Regional top picks are Delta Electronics, MediaTek, NAURA Technology, Samsung Electronics and Tokyo Electron. The broader preferred list also includes TSMC, SK Hynix, Accton, Advantest, AEM Holdings, All Ring, ASE, Hon Hai, ISU Petasys, JCET, KYEC and SEMCO, while the least-preferred list highlights consumer-heavy, display or valuation-sensitive names such as Asustek, Compal, Desay Battery, Hanmi Semiconductor, Innolux and Largan.

Analysis framework

UBS starts with end-demand indicators such as AI-token activity, hyperscaler cash flow, capital raising and capex. It then builds bottom-up forecasts for servers, accelerators, HBM, DDR, NAND, foundry capacity, CoWoS and semiconductor equipment. Those demand estimates are compared with wafer capacity, utilization, technology transitions and supplier allocation to judge shortages, pricing and earnings. The report then maps the industry conclusions into company revenue, margins, cash flow and valuation using P/B, P/E, EV/EBITDA and, for selected companies, DCF.

Methodology notes

• Industry AnalysisSupply-demand framework

  Bottom-up semiconductor supply-demand modeling

  UBS compares bit demand, wafer output, capacity, utilization and inventory adjustments for DRAM, HBM, NAND, foundry and advanced packaging to estimate shortages, pricing and cycle duration.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  AI-infrastructure supply-chain transmission

  The report traces hyperscaler spending through accelerators, CPUs and memory into foundry, packaging, substrates, power, networking, testing, materials and equipment suppliers.

• Cycle and Business ConditionsBusiness-Cycle Inflection Analysis

  Semiconductor-cycle lead indicators

  UBS uses semiconductor revenue growth, foundry utilization, assembly-equipment revenue and memory operating profit to estimate industry and stock-cycle peaks and troughs.

• Valuation methodsPB valuation

  Target P/B based on long-term ROE and cost of equity

  For companies such as Samsung Electronics, SK Hynix and LG Display, UBS derives target book-value multiples from long-term return on equity and cost-of-equity assumptions.

• Valuation methodsP/E and PEG Valuation

  Forward P/E tied to earnings growth

  Many equipment, component and materials targets apply forward P/E multiples selected with reference to historical bands, peers and long-term EPS growth.

• Valuation methodsDCF (Discounted Cash Flow)

  Discounted cash-flow valuation

  SemiFive's price target is explicitly based on a DCF using a 9.9% WACC and 3.0% terminal growth rate.

Asset mapping & comparison

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

• Samsung Electronics (005930.KS)

  Regional top pick and key memory beneficiary; UBS assigns Buy with a Won535,000 target.

  Strengths

  Improving HBM4 execution, capacity flexibility, expected 41% HBM bit share in 2027 and potential for materially higher shareholder returns.

  Weaknesses

  Non-memory divisions face weaker consumer demand, and HBM execution has not fully matched the strongest peer historically.

  Comparison

  UBS expects Samsung to overtake SK Hynix in 2027 HBM bit share, at 41% versus 37%.

  Risks

  Memory affordability and weaker-than-expected HBM execution or shareholder distributions.

• SK Hynix (000660.KS)

  Preferred memory exposure; UBS assigns Buy with a Won3,000,000 target.

  Strengths

  Strong HBM position, substantial free-cash-flow generation and a commitment to return at least 50% of cumulative 2025-27 free cash flow.

  Weaknesses

  UBS expects 2027 HBM shipment growth of 15%, below the market, as wafers are shifted toward DDR.

  Comparison

  Forecast 2027 HBM bit share is 37%, below Samsung's 41% but above Micron's 22%.

  Risks

  HBM pricing, market-share normalization and the affordability of rapidly increasing memory content.

• Delta Electronics (2308.TW)

  Regional Key Call Buy and beneficiary of rising AI-server power requirements; target NT$2,750.

  Strengths

  Broad grid-to-chip product portfolio and an estimated roughly 40% share of a growing AI-server power market.

  Weaknesses

  Near-term growth is constrained by power-IC shortages.

  Comparison

  UBS estimates the AI-server power TAM will compound 69% from 2025 to US$91.1bn in 2030.

  Risks

  Hyperscaler cash-flow limits, grid constraints and persistent component shortages.

• MediaTek (2454.TW)

  Regional top pick with Buy and a NT$7,300 target, driven by TPU and other custom-ASIC opportunities.

  Strengths

  Strong positioning in Google TPU v9 and potential additional ASIC programs.

  Weaknesses

  EMIB-T supply could constrain TPU v9 volumes above 4m units.

  Comparison

  UBS forecasts TPU sales of US$18.0bn, US$43.5bn and US$52.5bn in 2027E, 2028E and 2029E.

  Risks

  Packaging-supply constraints, project execution and weak smartphone demand.

• NAURA Technology (002371.SZ)

  Regional top pick and major beneficiary of Chinese memory investment and equipment localization; Buy with a Rmb905 target.

  Strengths

  Leading domestic WFE position, strong Chinese memory-capex exposure and localization-driven share potential.

  Weaknesses

  The investment case depends heavily on sustained domestic fab spending.

  Comparison

  UBS expects China's memory capex and local-equipment penetration to support above-market growth.

  Risks

  Slower Chinese memory investment, pricing pressure or delayed overseas-customer penetration.

• Tokyo Electron (8035.T)

  Regional top pick with Buy and a ¥93,600 target, geared to DRAM and broader WFE expansion.

  Strengths

  High DRAM exposure, pricing initiatives and potential China share recovery.

  Weaknesses

  Earnings remain sensitive to customer investment cycles.

  Comparison

  Management is targeting price increases consistent with a 50% gross-margin objective.

  Risks

  Delayed memory investment or weaker China equipment demand.

• TSMC (2330.TW)

  Preferred leading-edge foundry and advanced-packaging exposure; Buy with a NT$3,650 target.

  Strengths

  AI/HPC visibility into 2029-30, expanding CoWoS capacity and a roadmap into CoPoS.

  Weaknesses

  Very high capital requirements accompany capacity expansion.

  Comparison

  TSMC remains the dominant supplier in leading-edge foundry and CoWoS capacity.

  Risks

  Execution on large capacity additions and a slowdown in AI-platform demand.

• Samsung Electro-Mechanics (009150.KS)

  Preferred MLCC and substrate exposure; Buy with a Won2.7m target.

  Strengths

  AI-server MLCC demand, FC-BGA project ramps and improving product mix.

  Weaknesses

  High-end MLCC production has longer cycle times and lower yields.

  Comparison

  UBS forecasts operating-profit growth of 131% in 2026E and 165% in 2027E.

  Risks

  Slower AI-server demand, delayed substrate ramps or weaker pricing.

• ISU Petasys (007660.KS)

  Preferred AI-switch and accelerator PCB exposure; Buy with a Won190,000 target.

  Strengths

  Richer high-layer MLB mix, ASIC project wins and potential additional capacity.

  Weaknesses

  Its capacity ramp is slower than some aggressive competitors.

  Comparison

  Blended ASP is forecast to compound 24% from Won3.8bn/km² in 2024 to Won8.1bn/km² in 2028E.

  Risks

  Relative wallet-share loss or insufficient capacity expansion.

• Advantest (6857.T)

  Preferred logic-testing exposure; Buy with a ¥43,600 target.

  Strengths

  More than 70% logic exposure and leverage to chiplets, CoWoS, CoPoS, EMIB and custom ASIC testing.

  Weaknesses

  The market remains concerned about possible share loss to Teradyne.

  Comparison

  Chiplet architectures are estimated to raise testing costs by around 15-30% versus a single die.

  Risks

  Tester competition and lower-than-expected test intensity.

• Hanmi Semiconductor (042700.KS)

  Least-preferred HBM bonding-equipment name; UBS assigns Sell with a Won170,000 target.

  Strengths

  A leading role in HBM thermocompression bonding at SK Hynix and Micron.

  Weaknesses

  Current valuation already reflects an optimistic scenario, while supplier diversification may reduce share.

  Comparison

  UBS forecasts market share at SK Hynix and Micron falling to 46% and 84% in 2027E and normalizing to 45% and 75% by 2030E.

  Risks

  Loss of bonding-tool share, more upgrade spending instead of new tools and limited Samsung wins.

• Largan Precision (3008.TW)

  Least-preferred name; UBS downgraded it from Buy to Sell with a NT$5,000 target.

  Strengths

  Precision-manufacturing capabilities and longer-term CPO fiber-array-unit potential.

  Weaknesses

  Qualification is early, yields remain low and the competitive field is crowded.

  Comparison

  UBS believes the 242% share-price rise since April moved ahead of qualification progress.

  Risks

  Delayed Gen-3 optics adoption, weak yields, integrator acceptance and competition from incumbents or vertically integrated suppliers.

Key data

• Top-12 hyperscaler capexUS$1.0086tn in 2026E; US$1.4475tn in 2027E; US$1.6186tn in 2028EForecast growth of 98.5%, 43.5% and 11.8%, respectively.
• 2027 hyperscaler operating cash flow+30%UBS increased its 2027 OCF forecasts for hyperscalers and neoclouds by 30%.
• Memory-cycle durationDRAM toward 2Q28; NAND toward 4Q27UBS's revised expected endpoints for the respective upcycles.
• 3Q26/4Q26 blended contract ASP changeDRAM +25%/+9% QoQ; NAND +20%/+8% QoQPricing outlook amid tight fulfillment and limited supply.
• Current DRAM fulfillment ratio60%Evidence of continued industry undersupply.
• 2027 server-memory demand growthServer DDR +81% YoY; server SSD +50% YoYUBS says unconstrained server-SSD demand could be twice the supplied amount.
• 2027 HBM supply growth+47%Expected to remain below demand as capacity is reallocated toward DDR.
• 2027 HBM bit-share forecastSamsung 41%; SK Hynix 37%; Micron 22%Samsung is expected to move ahead of SK Hynix.
• Industry LTA coverageAbout 50% of total bits by year-endUBS expects 40% to be covered by fixed volume and pricing.
• CoWoS capacity133kwpm in 2026E; 233kwpm in 2027EIndustry utilization is forecast at 83% and 94%, respectively.
• AI GPU/accelerator units16.933m in 2026E; 28.493m in 2027EA 68% YoY increase in 2027E.
• Consumer-device outlookPC units -11%/-4% in 2026E/2027E; smartphones -10%/-3%/+2% in 2026E/2027E/2028EConsumer hardware remains materially weaker than AI infrastructure.
• CPO market opportunityUS$6.5bn base case or US$13.8bn upside case in 2030ECPO share of optical transceivers reaches 20% in the base case and 25% in the upside case.

Impact & implications

UBS expects AI spending to support a broad but uneven APAC technology upcycle. Memory producers benefit first from shortages and pricing, followed by equipment, testing, advanced-packaging, substrate, MLCC, power and networking suppliers as capacity and system complexity increase. Consumer-oriented PC, smartphone and display businesses face a less favorable demand and margin backdrop because higher memory and component costs reduce affordability.

Risks

• The gap between hyperscaler operating cash flow and capital expenditure is the report's leading financial risk, even though UBS considers it manageable.
• Affordability could constrain demand as memory-industry revenue approaches US$1.64tn in 2027E and memory absorbs a very large share of hyperscaler budgets.
• Weak PC, smartphone and other consumer demand may worsen as higher memory and component prices are passed through.
• Power IC, packaging and other component shortages may restrict near-term order fulfillment despite strong underlying demand.
• Rapid technology changes, intense competition and macroeconomic cyclicality can make technology-company earnings and valuations unusually volatile.
• CPO, CoPoS, EMIB-T, HBM4 and hybrid-bonding adoption may be delayed or produce different supplier shares than UBS assumes.

What to watch

• Track hyperscaler and neocloud operating cash flow, financing and capex to assess whether the funding gap remains manageable.
• Watch whether agentic-AI adoption continues to lift CPU, ASIC, HBM, DDR5 and server-SSD demand into 2027.
• Monitor DRAM and NAND contract pricing, fulfillment ratios and the expected extension of the cycles toward 2Q28 and 4Q27.
• Track LTA coverage, fixed-price terms and whether these agreements improve through-cycle margins and return on equity.
• Watch Samsung's HBM4 execution and whether its 2027 HBM bit share reaches 41%.
• Monitor shareholder-return announcements from Samsung Electronics and SK Hynix.
• Track CoWoS capacity and utilization, the 2H27 EMIB-T ramp, FY28 CoPoS production and CPO adoption from 2H26 through 2029.
• Watch whether AI-server strength can offset declining PC and smartphone volumes across APAC hardware companies.