Nomura revises AI accelerator packaging roadmap: maintains 2-die GPU assumption, brings EMIB-T shipments forward

Overview

This is a Nomura event commentary on the Japanese electronic components and semiconductor packaging industry. The core change is a reassessment of the AI accelerator packaging roadmap: on one hand, Rubin Ultra may shift from the four-die structure presented at GTC to a two-die structure, leading to a lower GPU packaging price assumption; on the other hand, EMIB-T package substrate shipments may begin in FY28/3, one year earlier than previously expected. The report believes the two effects largely offset each other in Ibiden's earnings forecasts, but timely EMIB-T capacity expansion could bring earnings upside after FY29/3.

Core views

The report's core views include: first, Nvidia Rubin GPU packaging shipments began in FY27/3 Q1 and are expected to enter full-scale mass production in Q2; second, the packaging designs of Rubin Ultra and Feynman may lean more toward 2-die or 2-module structures, reflecting technical constraints in combining larger interposers with HBM4; third, EMIB-T does not rely on traditional interposers and can achieve larger packages through silicon bridges, helping custom ASICs integrate more AI accelerator dies, HBM, and I/O into large packages; fourth, agent-based coding models are more sensitive to HBM capacity, inter-GPU bandwidth, and tokens per watt, so the value of the large-package roadmap is increasing.

Analysis framework

The report combines industry shipment observations, macro statistics on shipment value and unit-area pricing of rigid module boards, changes in AI accelerator packaging structures, and the demands of agent-model workloads for memory bandwidth and energy efficiency to derive packaging substrate technology path changes and revisions to Ibiden's earnings forecasts.

Methodology notes

• Industry technology roadmap analysisRevision of AI accelerator packaging roadmap assumptions

  Assess changes in package substrate demand through die count, interposer size, HBM4 routing difficulty, and EMIB-T shipment timing.

  A shift in Rubin Ultra from four dies to two dies lowers the high-end GPU packaging price assumption; earlier EMIB-T shipments increase demand for large package substrates, and together these determine the net impact on Ibiden's earnings forecasts.

• Demand-driven analysisCompute efficiency constraints of agent models

  The performance of agent-based coding models is constrained by HBM capacity, memory bandwidth, and inter-GPU bandwidth.

  Large packages can increase HBM capacity share and inter-GPU bandwidth, reduce data transfer from host memory or SSDs, lower signal energy consumption, and improve GPU utilization.

Asset mapping & comparison

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

• Ibiden [4062]

  A beneficiary in the advanced package substrate supply chain; the report explicitly maintains Buy.

  Strengths

  Has growth exposure to GPU package substrates and EMIB-T, benefiting from demand for large AI accelerator packages.

  Weaknesses

  The Rubin Ultra 2-die assumption leads to a downward revision in the FY28/3 GPU packaging price forecast.

  Comparison

  Compared with the traditional interposer route, EMIB-T is more conducive to producing large package substrates and may be adopted by custom ASIC vendors.

  Risks

  Insufficient EMIB-T capacity may constrain shipments; changes in the Rubin/Feynman roadmap may alter package substrate pricing and demand timing.

• AI accelerator packaging supply chain

  Driven by growth in agent-model workloads.

  Strengths

  Large packages increase HBM capacity, inter-GPU bandwidth, and rack compute density, while reducing per-token energy consumption and latency.

  Weaknesses

  There are technical difficulties in combining ultra-large interposers with HBM4's high routing density.

  Comparison

  Compared with PLP interposer-based 3.5D packaging, EMIB-T places greater emphasis on achieving large-format packaging through embedded silicon bridges.

  Risks

  Technology route changes, mass-production yields, customer ASIC design choices, and capacity ramp-up may all affect demand realization.

Key data

• Rubin GPU packaging shipment timingStarted in FY27/3 Q1, with full-scale mass production expected in Q2The report makes this judgment based on industry developments from April to June.
• Estimated monthly shipment valueJan-Mar: ¥24.7bn; Apr-May: ¥25.8bn; May: ¥28.1bnEstimated from macro statistics on shipment value related to rigid module boards.
• Unit area price changeJan-Mar: ¥1.18mn/m²; Apr-May: ¥12.9bn/m²; May: ¥13.6bn/m²Disclosed this way in the original text; the units and figures show significant order-of-magnitude differences, so the original wording is preserved to avoid distortion from rewriting.
• Rubin Ultra packaging assumptionIndustry view strengthening that it will change from four dies to two diesThe report believes this may be related to the difficulty of combining an interposer larger than 5.5x reticle with HBM4's high routing density.
• EMIB-T shipment assumptionShipments to begin in FY28/3, one year earlier than previously expectedThis is the key variable behind the report's higher medium-term package substrate demand outlook.
• Nomura Group stock rating distributionBuy 58%; Neutral 39%; Reduce 3%, as of 2026-06-30From the report's disclosure page; background information on institutional rating distribution.

Impact & implications

For the supply chain, the competitive focus in advanced packaging is shifting from simply expanding GPU package size to balancing technical manufacturability, HBM integration, inter-GPU bandwidth, and energy efficiency. For Ibiden, the Rubin Ultra 2-die assumption weakens FY28/3 GPU packaging pricing, but earlier EMIB-T shipments and expanding custom ASIC demand provide an offset; if capacity expansion is timely, earnings elasticity may become more evident after FY29/3.

Risks

• If the actual packaging structure of Rubin Ultra or Feynman differs from the 2-die / 2-module assumption, pricing and demand forecasts may prove inaccurate.
• Current EMIB-T shipment volumes may be constrained by insufficient capacity; if expansion is not timely, the earnings upside after FY29/3 may not materialize.
• The technical difficulty of combining an interposer over 5.5x reticle with HBM4's high routing density may continue to affect the high-end packaging roadmap.
• There is uncertainty around the pace and scale of AI agent-model workload growth and adoption of EMIB-T by custom ASICs.
• Shipment value and unit-area pricing derived from macro statistics may not map perfectly to the revenue or profit of a single company.

What to watch

• Whether Rubin GPU packaging enters full-scale mass production as expected in FY27/3 Q2.
• Whether Rubin Ultra ultimately adopts a 2-die structure, and whether Feynman adopts a two-module, two-dies-per-module 3D stacking structure.
• Whether EMIB-T package substrates begin shipping in FY28/3 and whether customers expand from GPUs to more custom ASICs.
• Ibiden's pace of EMIB-T capacity expansion, yields, and capital expenditure plans.
• Measured improvements from large packages in HBM capacity, inter-GPU bandwidth, per-token energy consumption, and agent-model latency.