More GW-scale data center projects support stronger AI hardware demand in 2027 than in 2026

Overview

This report is Nomura's updated research on the Asian AI semiconductor and server hardware supply chain. Starting from global announcements of new data center construction, the report refreshes and expands its proprietary project tracking, increasing the sample from 130+ to 240+ and GW-scale projects from 30+ to 40+. Even after more conservatively addressing double counting and projects that only have data center shells, the report still concludes that AI hardware demand will continue to rise from 2026-2030F, with 2027F demand stronger than 2026F.

Core views

The core view is that announcements of AI infrastructure construction are shifting from a single theme to broader coordinated efforts among cloud service providers, neoclouds, chip companies, governments, and infrastructure partners. The report expects new data center construction demand of about 27/28GW in 2026F/2027F, implying CoWoS demand of about 467k/599k wafers in 2026F/2027F. More aggressive self-build projects and compute leasing partnerships by cloud service providers improve demand visibility, so the report reiterates its positive view on the AI semiconductor and hardware supply chain.

Analysis framework

The report uses a bottom-up announcement tracking approach, collecting announcements on data center construction, chip procurement, and compute partnerships from major cloud service providers, neoclouds, LLM companies, AI startups, and infrastructure partners; it then screens for GW-scale projects, excludes potential duplicates and shell projects without clear IT deployment demand, and converts GW capacity into AI chip volumes and CoWoS wafer demand.

Methodology notes

• Leading indicator trackingNomura proprietary global new data center construction tracker

  Use data center construction announcements as a leading indicator of Asian AI hardware demand.

  The report believes that only observing orders in the Asian technology supply chain may be lagging, so it tracks global data center projects, cloud service provider partnerships, and neocloud expansion plans to anticipate demand for AI chips, servers, and components in 2026-2030F.

• Demand conversionEstimating AI chip and CoWoS demand from GW capacity

  Convert newly deployed data center capacity into AI chip volumes and CoWoS wafer demand.

  The report assumes that incremental capacity before 2026F is mainly based on GB300 NVL72, while after 2027F it is mainly based on the Vera Rubin platform, and combines rack power, chip count per GW, and CoWoS ratios to estimate chip and CoWoS demand.

• Risk controlExclusion of double counting and shell projects

  Avoid repeatedly counting chip partnerships, Stargate-related projects, and projects that only have infrastructure shells.

  This time, the report retains only OpenAI's 26GW chip agreement and excludes data center shell projects without clear tenants or IT deployment; chip partnership announcements are handled separately from data center projects.

Asset mapping & comparison

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

• TSMC (2330 TT)

  A core beneficiary of AI chip manufacturing.

  Strengths

  The report lists it as the top pick in logic semiconductors, emphasizing its leading position in AI chip manufacturing.

  Weaknesses

  The report does not disclose target price or specific valuation sensitivity within the scope of this summary.

  Comparison

  It is listed as a top pick within the logic semiconductor supply chain.

  Risks

  If AI data center construction or advanced process demand falls short of expectations, upside realization will be affected.

• MediaTek (2454 TT)

  Related to the long-term TPU trend and non-NVIDIA AI chip demand.

  Strengths

  The report upgrades it from Neutral to Buy due to improved risk-reward and the long-term TPU uptrend.

  Weaknesses

  Earnings realization depends on continued expansion in TPU and custom chip demand.

  Comparison

  It was upgraded from a Neutral rating of about 2,000 days to Buy, the explicit rating change in this report.

  Risks

  TPU demand timing, customer project execution, and changes in the competitive landscape may affect performance.

• NVIDIA CORP (NVDA US)

  GB300, Vera Rubin, and neocloud partnerships are important anchors for the data center demand estimates.

  Strengths

  It collaborates frequently with ecosystem partners such as CoreWeave, Nebius, Meta, and AWS, and is driving new platform deployments.

  Weaknesses

  It is marked as Not Rated in the report, and some partnerships may overlap in scope with data center projects.

  Comparison

  It is a key platform supplier in the AI hardware cycle, but the report's investment recommendations are mainly focused on the Asian supply chain.

  Risks

  Product transitions, double counting of partnership announcements, and deployment delays may affect estimation accuracy.

• CoreWeave (CRWV US) / Nebius (NBIS US)

  Neocloud expansion and partnership with NVIDIA improve visibility into AI compute deployments.

  Strengths

  Both companies are involved in capacity deployment plans of more than 5GW by 2030 and have deep partnerships with NVIDIA.

  Weaknesses

  Both are marked as Not Rated in the report, and project capex and delivery capabilities still need to be tracked.

  Comparison

  Compared with traditional cloud service providers, neoclouds are more aggressive and are an important source of new AI factory capacity.

  Risks

  Financing, power supply, rack delivery, and customer contract execution risks may affect construction pace.

• Samsung Electronics (005930 KS) / SK Hynix (000660 KS)

  Beneficiaries of the memory upcycle and AI server memory demand.

  Strengths

  The report maintains a positive view on the memory cycle and keeps both companies rated Buy.

  Weaknesses

  The report does not provide company-specific valuation details within the scope of this summary.

  Comparison

  Both are core beneficiary assets within the memory segment.

  Risks

  Memory pricing cycles, HBM supply-demand dynamics, and fluctuations in AI server demand may affect earnings.

• EMC (2383 TT) / Unimicron (3037 TT)

  The PCB, CCL, and substrate supply chain benefits from Vera Rubin, AI ASICs, and high-end server upgrades.

  Strengths

  EMC is expected to benefit from mid-board PCBs in Vera Rubin, new AI ASIC boards, and more favorable pricing; Unimicron benefits from substrate and PCB demand related to NVIDIA Blackwell, Google TPU/CPU, AWS Trainium 3, and Meta ASICs.

  Weaknesses

  Capacity expansion, cost inflation, and customer pricing mechanisms remain key variables.

  Comparison

  Both are key Buy-rated names in the downstream hardware materials and board-level supply chain.

  Risks

  If AI server platform upgrades slow or customer ROI models fall short of expectations, demand and earnings leverage may decline.

• Delta (2308 TT) / AVC (3017 TT) / Bizlink (3665 TT)

  Benefit from higher AI server rack power, HVDC upgrades, liquid cooling, and high-speed interconnect trends.

  Strengths

  Delta holds a leading position in AI power and HVDC upgrades; AVC benefits from full liquid cooling designs and contributions from new customers; Bizlink benefits from rising power consumption and transmission speed upgrades.

  Weaknesses

  The report does not disclose target prices for individual companies or short-term order details.

  Comparison

  Compared with chip manufacturing, these companies benefit more directly from higher value per watt and increasing rack system complexity.

  Risks

  Changes in rack design solutions, the pace of liquid cooling adoption, and slower customer capex may affect realization.

• Zhongji InnoLight (300308 CH), Suzhou TFC (300394 CH), Shenzhen T&S (300570 CH)

  Participants in the networking and potential CPO supply chain.

  Strengths

  The report believes Zhongji InnoLight has attractive risk-reward and notes that potential CPO supply chain participants have more time to prepare for technical challenges.

  Weaknesses

  The report mentions that the timeline for scale-up CPO adoption has been pushed back slightly.

  Comparison

  In networking and optical modules, compared with segments already seeing clear volume ramp-up, CPO remains more of a medium- to long-term trend.

  Risks

  There is uncertainty around the CPO adoption timetable, technical barriers, and customer validation progress.

Key data

• Project sample size240+ projectsPreviously 130+ projects; the sample scope has expanded significantly.
• Number of GW-scale projects40+ projectsPreviously 30+ projects; this time more projects with explicit GW-scale deployment plans are included.
• New data center construction demandAbout 7GW in 2025F; about 27GW in 2026F; about 28GW in 2027FEstimated based on 40+ GW-scale projects, excluding the remaining roughly 200 samples.
• Incremental capacity deployment2026F 26.67GW; 2027F 28.48GW; 2028F 21.35GW; 2029F 16.89GW; 2030F 7.09GWThe report's estimates show that new deployments in 2027F are higher than in 2026F.
• Implied AI chip demandAbout 7.479 million units in 2026F; about 5.394 million units in 2027FThe text summarizes this as annual AI chip demand of about 4-7 million units in 2026-2028F.
• Implied CoWoS demandAbout 467k wafers in 2026F; about 599k wafers in 2027FThe main text summarizes this as incremental CoWoS demand of about 450k/600k wafers in 2026F/2027F.
• Share of projects from the top four cloud service providersProject count share rose from about 24% to 30%+In the capacity estimates, the top four cloud service providers account for 31% in 2026F and 24% in 2027F.
• OpenAI capacity share2026F 13%; 2027F 26%; 2028F 40%The report emphasizes that OpenAI's share in the portfolio has relatively declined while the cloud service provider share has increased, making the demand mix more balanced.
• Greenfield data center construction costAbout USD20-40bn/GWA rough budget range provided by the report.

Impact & implications

If the report's estimates materialize, AI compute infrastructure construction will continue to drive demand for advanced process nodes, advanced packaging, memory, server motherboards, PCB/CCL, power, thermal, interconnect, and networking optical modules. The beneficiary chain is not limited to GPUs, but also includes the Vera Rubin platform, ASICs, TPUs, Trainium, HBM, CoWoS, and components related to high-power racks. From an investment perspective, the report prefers Asian hardware suppliers with manufacturing leadership, pricing power due to tight capacity, or direct benefits from higher rack power and transmission speed upgrades.

Risks

• Data center project announcements may overlap with chip partnerships or Stargate-related projects, leading to double counting.
• Some infrastructure projects may only be power shells or campus developments; without tenants or IT deployment, they cannot be directly converted into AI hardware demand.
• If market concerns about an AI bubble intensify, capex and valuations may come under pressure.
• There is uncertainty in the OpenAI-Microsoft relationship; the report mentions that Microsoft may consider legal action regarding the OpenAI-Amazon transaction.
• If assumptions around the timing, power consumption, and rack configuration of the transition from GB300 to the Vera Rubin platform deviate significantly, chip and CoWoS demand estimates will be affected.
• Power supply, grid connection, high-voltage transmission and distribution, land, and cooling facilities may become bottlenecks for GW-scale data center rollout.
• The CPO adoption timeline has already been pushed back slightly, and the realization pace of the networking supply chain may come in below expectations.
• The report does not provide valuations and target prices for all beneficiary companies, so investment judgments still need to be combined with company fundamentals and prices.

What to watch

• Follow-up announcements and actual groundbreaking and energization milestones for major projects such as AWS, Google, Meta, Microsoft, HUMAIN, OpenAI, Nscale, Nebius, and IREN.
• Whether new GW capacity in 2026F and 2027F materializes in line with the report's estimates, especially whether 2027F remains stronger than 2026F.
• Whether the share of self-built projects by cloud service providers continues to increase, and whether there are further scope adjustments for OpenAI-related projects.
• Order trends, delivery pace, and early neocloud deployments for the NVIDIA Vera Rubin platform.
• Changes in CoWoS capacity, advanced packaging supply-demand dynamics, and assumptions on the number of chips per CoWoS wafer.
• MediaTek TPU-related progress and demand trends for non-NVIDIA AI ASICs.
• Whether value content continues to increase in AI server power, HVDC, liquid cooling, PCB/CCL, substrates, and interconnects.
• Whether the timelines for CPO and high-speed network upgrades reaccelerate.