The AI semiconductor and server cycle has not yet peaked, with advanced packaging and component bottlenecks becoming the main theme

Overview

This is an Anchor Report by Nomura on the Asian AI semiconductor and server supply chain. The report's core judgment is that although conventional cycle signals such as price hikes, long-term agreements, and potential over-ordering indicate a very hot cycle, and stock prices have already risen sharply before pulling back in the short term, AI infrastructure demand remains real and strong. Tracking of global new data center construction, rising hyperscaler capex, and a two-year greenfield capacity cycle on the supply side all point to supply remaining tight around 2027F. The report therefore maintains a structurally bullish view and broadens its focus from pure CoWoS expansion to WoS, substrates, CCL, capacitors, PMICs, optical components, CPUs, and execution of advanced packaging technologies.

Core views

The report believes the AI hardware cycle has not yet peaked for several reasons: first, hyperscalers may continue to raise 2027F capex in the 'go big or go home' competition; second, global data center project tracking shows further upside room for hardware demand in 2027F and 2028F; third, large-scale greenfield capacity expansion on the supply side only starts from late 2025, implying that supply may still be insufficient in 2027F; fourth, supply bottlenecks are shifting from TSMC-dominated CoW to WoS and other small components, which may increase stock-price volatility but help extend the long-term cycle; fifth, price increases and upward earnings forecast revisions remain important catalysts for Asian semiconductor hardware stocks.

Analysis framework

The report combines its own tracking of global new data center construction projects, scenario analysis of CoWoS/CoPoS/SoIC capacity, assumptions on GPU/ASIC supply allocation, server shipment and revenue forecasts, and analysis of AI supply-chain component bottlenecks to derive revenue, capacity, and beneficiary names for the Asian semiconductor hardware supply chain from 2026F to 2029F. The analysis covers not only upstream foundries, advanced packaging, OSAT, memory, and silicon materials, but also downstream opportunities in PCB/CCL, power, thermal management, ODM, and server CPU-related areas.

Methodology notes

• Cycle assessmentAI infrastructure cycle tracking

  Use new data center projects and hyperscaler capex as leading indicators in place of traditional semiconductor cycle signals.

  The report believes traditional semiconductor cycle signals already indicate elevated heat, but AI demand has structural characteristics; therefore it uses global new data center projects, GW-scale capacity, and deployment schedules to judge whether hardware demand still has upside.

• Supply bottlenecksBreakdown of CoW, WoS, and small components

  Analyze TSMC-controlled CoW separately from WoS, substrates, CCL, capacitors, PMICs, and optical components handled by the external supply chain.

  The report believes that even if TSMC increases its CoWoS target capacity, WoS and multiple categories of small components may still become more severe bottlenecks from 2H26 to 2027F, thereby affecting GPU/ASIC supply allocation.

• Technology roadmapAdvanced packaging competition framework

  Compare the competitive relationship among TSMC CoWoS, CoPoS, SoIC, and Intel EMIB-T.

  The report views EMIB-T as one of the biggest potential threats to TSMC's advanced packaging, and believes SoIC and CoPoS are key technologies for TSMC to maintain its lead, especially in relation to nVidia Feynman, GPU-on-GPU stacking, and SiC carrier demand.

• Demand estimationGW-to-chip demand conversion

  Convert incremental GW capacity in data centers into rack counts, AI chip counts, and CoWoS demand.

  Using assumptions for GB300 and VR200 rack power consumption and chip configurations, the report converts deployed data center capacity into AI chip and CoWoS demand to assess future supply-demand gaps.

Asset mapping & comparison

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

• TSMC 2330 TT

  Core enabler of AI chips and advanced packaging; the report reiterates Buy and raises the target price.

  Strengths

  A more aggressive CoWoS capacity target, with SoIC and CoPoS helping maintain leadership in advanced packaging.

  Weaknesses

  WoS and other non-TSMC-controlled links may limit actual shipments, and after 2028F the company still depends on execution across multiple frontier technologies.

  Comparison

  Compared with Intel EMIB-T, TSMC needs to maintain its lead through the 14x reticle CoWoS roadmap, SoIC, and CoPoS.

  Risks

  EMIB-T competition, CoPoS mass-production timing, WoS bottlenecks, and advanced-packaging execution risk.

• ASE 3711 TT

  Beneficiary of OSAT and WoS outsourcing; the report reiterates Buy and raises the target price.

  Strengths

  TSMC's expansion of CoW capacity and WoS outsourcing directly benefit OSATs, while CPUs also create opportunities for CoWoS-like processes.

  Weaknesses

  WoS supply constraints may limit shipment elasticity, and insufficient mature yields could bring risk of large losses.

  Comparison

  Compared with foundries, OSATs benefit more directly from WoS outsourcing and growth in CPU packaging volume.

  Risks

  Substrate price increases, yield ramp-up, and the pace of customer CPU project volume ramp.

• MediaTek 2454 TT

  Beneficiary of Google TPU growth and the ASIC supply chain; the report reiterates Buy and raises the target price.

  Strengths

  The report expects TPU to be one of the fastest-growing AI logic semiconductors in 2027F, and MediaTek's share in TPU may rise from about 15% in 2026F to above 30% in 2027F.

  Weaknesses

  Highly dependent on Google's TPU roadmap and advanced-packaging capacity allocation.

  Comparison

  Compared with other GPU/ASIC suppliers, MediaTek benefits from Google's rising TPU share.

  Risks

  Google project timing, EMIB-T complexity, and supply-chain constraints such as CoWoS/substrates.

• ASPEED 5274 TT

  Beneficiary of rising demand for server CPUs and BMCs; the report reiterates Buy and raises the target price.

  Strengths

  Agentic AI is driving stronger-than-expected demand for CPU servers, and rising CPU counts and rack density expand BMC TAM.

  Weaknesses

  Demand is tied to server CPU architecture and shipment timing.

  Comparison

  Compared with pure GPU-chain names, ASPEED benefits from growth in general-purpose/CPU servers and peripheral demand in AI data centers.

  Risks

  Weaker-than-expected CPU server demand and delays in customer platform transitions.

• GWC 6488 TT

  Beneficiary of SiC carriers and Feynman thermal-design upgrades; the report reiterates Buy and raises the target price.

  Strengths

  Feynman's larger area and higher TDP may drive adoption of SiC carriers.

  Weaknesses

  The opportunity depends on the Feynman roadmap, SoIC stacking, and implementation of high thermal design power.

  Comparison

  Compared with traditional silicon carriers, SiC is viewed as an upgrade direction for thermal management in high-power advanced packaging.

  Risks

  Changes to the Feynman timetable, and uncertainty in material validation and mass-production progress.

• KYEC 2449 TT

  Beneficiary of AI chip testing; the report reiterates Buy and raises the target price.

  Strengths

  Greater AI chip complexity and shipment growth drive testing demand.

  Weaknesses

  Growth depends on AI chip customer volume ramp and test-capacity utilization.

  Comparison

  Compared with packaging and substrate segments, testing has more elasticity to back-end manufacturing demand.

  Risks

  Volatility in AI chip shipments, customer concentration, and pricing pressure.

• EMC 2383 TT / TUC 6274 TT

  Beneficiaries of AI PCB upgrades and CCL bottlenecks; the report reiterates Buy and raises the target price.

  Strengths

  Upgrades to low-loss materials, increased board counts around AI GPU/ASIC/CPU/switches, and tight supply support price increases.

  Weaknesses

  Highly affected by material costs and capacity bottlenecks.

  Comparison

  The report views CCL as one of the main supply bottlenecks, giving it stronger pricing elasticity than general components.

  Risks

  Demand slowdown, overly rapid capacity expansion, and fluctuations in material costs.

• ZDT 4958 TT

  Emerging AI PCB/HDI vendor; the report reiterates Buy and raises the target price.

  Strengths

  Expected to penetrate more deeply into the PCB/HDI supply chains of nVidia, Google, and AWS from 2H26F.

  Weaknesses

  Still at the stage of increasing penetration, making customer qualification and execution of mass production important.

  Comparison

  Compared with mature PCB makers, ZDT is positioned by the report as an emerging incremental beneficiary in AI PCB/HDI.

  Risks

  Customer onboarding below expectations, HDI technology upgrades, and yield risks.

• Delta 2308 TT / AVC 3017 TT

  Respectively aligned with the main themes of AI data center power supply and thermal management, with the report remaining positive.

  Strengths

  Delta benefits from +/-400VDC projects ramping from 2H26F; AVC benefits from penetration of VR200, Trainium 3, and Google TPU.

  Weaknesses

  The contribution from power and thermal solutions depends on platform transitions, customer validation, and shipment timing.

  Comparison

  Compared with the semiconductor manufacturing chain, these two asset types are more aligned with supporting upgrades in AI data center infrastructure.

  Risks

  Delays in 800VDC or HVDC progress, unsuccessful platform ramps, and changes in customer procurement timing.

• nVidia NVDA US / Google GOOGL US

  Key demand-side and competitive players in AI semiconductor supply allocation; not rated by the report.

  Strengths

  nVidia may still hold the largest CoWoS allocation; Google TPU continues to grow on the back of Gemini performance, ecosystem strength, and share gains.

  Weaknesses

  The two compete for CoWoS, substrates, and small components, potentially squeezing other GPU/ASIC vendors.

  Comparison

  The report uses the phrase 'elephants fight, grass gets trampled' to describe how competition between nVidia and Google squeezes other xPU/ASIC suppliers.

  Risks

  Insufficient supply-chain allocation, technology roadmap changes, and hyperscaler FCF pressure.

Key data

• SOX gain during the period+85% since the March update; +211% since the AI theme review in May 2025The report believes the recent pullback is a healthy consolidation after a rapid rise, but does not mean the cycle has already peaked.
• Global sample of new data center projectsAbout 280, versus about 240 previouslyThe number of projects continues to increase, supporting potential hardware demand over the next 2-3 years.
• Number of GW-scale projectsAbout 50, compared with more than 40 previouslyThe report mentions newly added or advancing GW-scale projects such as Nebius, Softbank, and SK Telecom.
• Incremental capacity deploymentAbout 26-27GW in 2026F; about 32GW in 2027F; about 23GW in 2028FThe 2027F estimate was raised from about 28GW to about 32GW, and 2028F from about 21GW to about 23GW.
• TSMC CoWoS target capacity2026F 1,100kpcs; 2027F target 2,000kpcsThe report believes TSMC is responding more aggressively to AI chip demand and defending against EMIB-T and FOPLP competition.
• TSMC long-term CoWoS demand2,500-3,500kpcs of CoWoS output required in 2029FThis is a scenario analysis based on achieving management's target of '50% CAGR in AI revenue from 2024-29E'.
• Potential CoPoS capacity demand700-800kpcs in 2029FAssumes nVidia Feynman production fully migrates to CoPoS in 2029F.
• Server revenue growth forecastGlobal servers +74%/+65% YoY in 2026F/2027FThe report raises its previous forecasts significantly, reflecting stronger AI and general-purpose/CPU servers.
• AI server revenue growth forecast+78%/+76% YoY in 2026F/2027FThe previous 2026F forecast was +58% YoY.
• General-purpose/CPU server revenue growth forecast+67%/+43% YoY in 2026F/2027FThe previous 2026F forecast was +16% YoY, with Agentic AI driving stronger-than-expected CPU demand.
• 2026F GB/VR rack assumption54.5k unitsRaised from 50k units; VR200 is expected to account for 15-20% of 2026F, mainly concentrated in 4Q26F.
• 2027F rack forecast62k unitsThe report introduces a 2027F forecast and assumes a possible transition from Rubin to Rubin Ultra in 2Q27F.
• CoWoS allocation assumptionnVidia about 55%; Google TPU about 26%The report believes that while nVidia is still seeking a 60% allocation, Google TPU will further squeeze other GPU/ASIC vendors.

Impact & implications

If the report's judgment is correct, the main investment theme in the Asian AI semiconductor hardware chain will expand from single AI GPU demand to broader advanced packaging, WoS outsourcing, OSAT-owned CoWoS-like processes, CPU/BMC, SiC carriers, PCB/CCL, power, thermal management, and memory. Supply-demand mismatches will create opportunities for price increases and upward earnings revisions, but may also cause short-term stock volatility and squeeze supply in non-AI consumer electronics, automotive, and other sectors. For portfolios, the report favors supply-chain companies that can benefit from real bottlenecks, technology upgrades, or market-share gains, while remaining cautious on other GPU/ASIC vendors squeezed by nVidia and Google.

Risks

• An unprecedented component supply-demand mismatch may emerge from 2H26 to 2027F, with WoS, PCB/CCL, IC substrates, high-end capacitors, PMICs, and optical components all potentially facing shortages.
• Rising memory costs may leave hyperscalers short of free cash flow in 2027F, triggering market concerns about the sustainability of AI capex.
• After 2028F, the hardware roadmap requires multiple technological breakthroughs, including EMIB-T, CoPoS, GPU-on-GPU SoIC, MCL, CPO, 336G/448G SerDes, M9Q/M10Q PCB, PTFE, and HDI materials/tools.
• Rising U.S. yields, inflation, and changes in the macro environment may pressure valuations and amplify stock-price volatility.
• If AI infrastructure projects are delayed, canceled, or lack sufficient tenants, visibility for chip and server hardware demand will weaken.
• While supply-chain price increases benefit earnings, they may also further affect supply and costs for downstream sectors such as non-AI consumer electronics and automotive.

What to watch

• Whether full-year visibility for global new data center construction in 2028F continues to improve over the next 3-6 months.
• Whether TSMC's 2027F CoWoS target capacity of 2,000kpcs can translate into actual shipments, and whether WoS becomes the larger bottleneck.
• Whether CoPoS can enter mass production before 2029F and align with nVidia Feynman's 2H28F timeline.
• Whether SoIC capacity doubles consecutively in 2027F and 2028F as expected in the report.
• Whether Google TPU's share of CoWoS allocation rises to about 26% in 2027F, and whether MediaTek's share in TPU exceeds 30%.
• Whether nVidia can maintain about 55%-60% of CoWoS resources and sell out its reserved capacity.
• Whether platform transitions from GB300 to VR200, Rubin to Rubin Ultra, and Trainium 3 proceed smoothly.
• Changes in pricing and lead times for small components such as PCB/CCL, IC substrates, capacitors, PMICs, and optical components.
• Trends in hyperscaler capex, free cash flow, and memory costs.
• Whether revenue growth forecasts for AI servers and general-purpose/CPU servers continue to be revised upward.