Google Cloud Next'26 highlights expansion across the AI infrastructure supply chain, with TPU, ARM CPU, GPU, and liquid cooling chains all benefiting

Overview

This report is JPMorgan’s summary of conference takeaways on the Asian technology supply chain from Google Cloud Next'26. The core message is that Google is simultaneously advancing its in-house TPU, Axion ARM CPU, and NVIDIA Rubin GPU clusters in AI infrastructure, creating multi-layer supply chain opportunities spanning chip design, foundry, packaging, server ODM, liquid cooling, ABF substrates, PCB, and CCL. The report is overall positive, but more cautious on the view that MediaTek is involved in TPU v8t rather than the larger inference chip opportunity.

Core views

First, the specification differences between TPU 8t and 8i indicate rising Google demand for high-power, high-speed inference, with 8i potentially involving Broadcom in design and 8t potentially involving MediaTek. Second, TPU v8 clusters will use more Axion ARM CPU, significantly increasing Axion shipments after end-2026 and expanding the TAM for TSMC, GUC, Hon Hai, and related packaging and substrate supply chains. Third, Google launched NVIDIA Vera Rubin NVL72-based A5x GPU instances, showing that ASIC and GPU demand can grow in parallel amid surging token demand. Fourth, Taiwan’s server ODM, liquid cooling, ABF substrate, PCB, and CCL segments will benefit from TPU v8 volume ramp-up, specification upgrades, and tight supply.

Analysis framework

The report combines interpretation of conference product launches, supply chain checks, chip specification and supply chain mapping, and a summary of company ratings and price information to assess the incremental impact of Google TPU v8, Axion CPU, and Rubin GPU deployments on the Asian technology supply chain.

Methodology notes

• Supply chain analysisAI data center supply chain mapping

  Breaks down the Google TPU ecosystem into design, foundry packaging, testing, substrates, PCB, CCL, server boards, cold plates, CDU, and rack assembly, among other segments.

  This framework is used to identify the potential impact of Google Cloud Next'26 product launches on orders, share, and capacity for different suppliers.

• TAM analysisIncremental market opportunity assessment

  Compares the different market size opportunities created by TPU training chips, inference chips, Axion ARM CPU, and GPU clusters.

  The report believes the TAM for the v8t training chip is relatively smaller than that of the high-end v8i inference chip, so if MediaTek mainly participates in v8t, it is slightly negative for its eventual TAM.

• Supply-demand analysisSpecification upgrades and capacity tightness assessment

  Observes ABF substrate size, layer count, CCL grade, and the degree of PCB supply tightness.

  The TPU v8 substrate size is nearly twice that of v7, with layer count up 30% to 50%, and the CCL material is upgraded from M7 to M8, driving tightness in high-end materials and the PCB supply chain.

Asset mapping & comparison

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

• MediaTek Inc. (2454.TW)

  May be the design partner for Google TPU v8t

  Strengths

  Entry into the TPU design supply chain reflects recognition of its design capabilities.

  Weaknesses

  v8t is a training chip, and the report believes its TAM is smaller than the previously expected inference chip opportunity.

  Comparison

  Compared with v8i, which Broadcom may participate in, MediaTek’s corresponding opportunity may be smaller in scale and strategic significance.

  Risks

  The market may previously have believed both 8t and 8i were designed by Broadcom; the actual design split still requires confirmation from earnings calls or follow-up checks.

• Broadcom

  May be the design partner for Google TPU v8i

  Strengths

  If involved in the high-end v8i inference chip, it would correspond to stronger specifications and larger inference demand.

  Weaknesses

  The report does not provide a specific rating or price.

  Comparison

  Relative to MediaTek’s v8t, v8i is more aligned with the demand curve for high-power, high-speed inference.

  Risks

  Supply chain checks still involve uncertainty.

• TSMC (2330.TW)

  Core beneficiary of TPU and Axion CPU foundry and related supply chain

  Strengths

  Volume ramp-up in Axion ARM CPU and TPU expands demand for advanced process and packaging-related services.

  Weaknesses

  The report does not quantify TSMC’s standalone revenue elasticity in the main text.

  Comparison

  Positioned in the foundry/packaging segment of the Google TPU food chain.

  Risks

  Demand cadence depends on Google TPU v8 cluster deployment and AI data center capex.

• GUC (3443.TW)

  Potential beneficiary of the Axion CPU and related ASIC supply chain

  Strengths

  Growth in Axion shipments brings incremental opportunities.

  Weaknesses

  The report notes margins may be much lower.

  Comparison

  Compared with TSMC, GUC’s earnings elasticity may be constrained by margins.

  Risks

  Listed with a UW rating, so focus should be on earnings quality rather than only shipment growth.

• Hon Hai Precision (2317.TW)

  Potential share gains at the TPU system level and in rack assembly

  Strengths

  The report expects it to gain some share in the next generation at the CPU tray and TPU rack levels.

  Weaknesses

  U.S. server ODMs such as Celestica may still be the main beneficiaries of TPU servers.

  Comparison

  Compared with Celestica’s existing advantage, Hon Hai appears more like a share gainer as scale expands.

  Risks

  Share realization depends on Google’s scaled procurement and system project allocation.

• NVIDIA Corporation (NVDA)

  Google A5x GPU instances use NVIDIA Vera Rubin NVL72

  Strengths

  Up to 80k Rubin GPUs per site and up to 960k GPUs across multiple sites support NVIDIA demand growth.

  Weaknesses

  The report mentions that overall sentiment toward the GPU supply chain remains relatively subdued.

  Comparison

  Not zero-sum with ASIC deployment, but rather jointly meeting growth in token demand.

  Risks

  Market sentiment toward the GPU supply chain and expectations for AI capex may affect valuation.

• Asia Vital Components (3017.TW)

  New supplier of TPU v8 liquid-cooling cold plate modules

  Strengths

  The report believes AVC will become a major beneficiary starting with v8 and the second-largest cold plate module supplier.

  Weaknesses

  Fujikura is still expected to remain the leader.

  Comparison

  In the cold plate competitive landscape, it ranks behind Fujikura and ahead of the potential positions of Boyd and Cooler Master.

  Risks

  Share ramp-up, yield, and Google procurement qualification still need to be tracked.

• Unimicron (3037.TW)

  Major supplier of TPU ABF substrates

  Strengths

  v8i share is expected to exceed 70%, with even higher share for v8t, benefiting from upgrades in size and layer count.

  Weaknesses

  Tight high-end ABF capacity may constrain deliveries.

  Comparison

  Google will become its second-largest substrate customer, significantly elevating its position.

  Risks

  Capacity expansion, yield, customer concentration, and specification migration risks.

• Elite Material Co (2383.TW)

  High-end CCL supplier

  Strengths

  With TPU v8 upgrading from M7 to M8, EMC’s share is expected to rise from 65% to above 75%, and it may continue to benefit if v9 partially adopts M9.

  Weaknesses

  High-grade material qualification and capacity need to match demand.

  Comparison

  Gaining share amid high-end CCL upgrades.

  Risks

  Changes in material specifications, new competitor entry, and customer volume ramp timing.

• ISU Petasys (007660.KS)

  Google TPU PCB supplier

  Strengths

  The report expects its share of Google’s PCB wallet to reach around 25% in FY26/27E, and JPMorgan rates it OW.

  Weaknesses

  Disclosures indicate that J.P. Morgan has market-making, shareholding, client, and compensation-related relationships with ISU Petasys.

  Comparison

  Together with WUS, TTM, and new PCB suppliers, it helps alleviate supply tightness and prepare for the v9 HDI technology transition.

  Risks

  PCB supply tightness, HDI technology migration, customer concentration, and conflict-of-interest disclosures warrant attention.

Key data

• TPU 8t specifications1 Compute die, 6 HBM stacksThe report judges that 8t is focused on training and may involve MediaTek in the design.
• TPU 8i specifications2 Compute dies, 8 HBM stacksThe report says 8i is focused on high-end inference, with higher HBM density and SRAM density, and may involve Broadcom in the design.
• Estimated server ARM CPU shipments5 million to 6 million units in 2027Driven by growth in Axion, NVIDIA Grace/Vera, and AWS Graviton, among others.
• NVIDIA Rubin cluster scaleUp to 80k Rubin GPUs per site and up to 960k GPUs across multiple sitesGoogle A5x GPU instances use NVIDIA Vera Rubin NVL72, ConnectX-9 SuperNICs, and Google Virgo networking.
• TPU v8 substrate specification changesSize nearly doubles versus v7, with layer count up 30% to 50%The report believes this will consume already tight high-end ABF capacity.
• Unimicron TPU substrate shareMore than 70% for v8i, and higher for v8tThe report says Google will become Unimicron’s second-largest substrate customer.
• EMC share in TPU CCLRises from 65% to above 75%TPU v8 uses M8 material, and v9 may partially adopt M9.
• ISU Petasys share of Google PCB walletReaches around 25% in FY26/27EThe report expects TPU v8 volume ramp-up to tighten PCB supply and drive share gains.

Impact & implications

From an investment perspective, Google’s AI infrastructure strategy is not simply a single ASIC replacing GPU, but rather concurrent expansion of its in-house TPU, ARM CPU, and NVIDIA GPU. Beneficiaries span TSMC, GUC, Hon Hai, Celestica, Inventec, AVC, Fujikura, Unimicron, EMC, ISU Petasys, WUS, TTM, and others; meanwhile, MediaTek’s opportunity needs further confirmation at next week’s earnings call.

Risks

• The division of labor among Google TPU v8 design partners still depends on industry checks, and the actual roles of MediaTek and Broadcom may change.
• If MediaTek mainly participates in the v8t training chip rather than the v8i inference chip, its TAM may be below prior market expectations.
• If AI data center capex, token demand, and Google’s external sales model come in below expectations, supply chain volume ramp-up will be affected.
• Capacity expansion, qualification, yield, and delivery of high-end ABF, CCL, PCB, and liquid cooling may become bottlenecks.
• The report includes existing ratings and compliance disclosures for multiple companies, and some companies have J.P. Morgan client, shareholding, or market-making relationships.

What to watch

• What MediaTek says at next week’s earnings call about its design role in TPU v8t or v8i.
• The adoption ratio of Axion CPU in Google TPU v8 clusters from end-2026 and the actual shipment cadence.
• Deployment scale of Google A5x Rubin clusters and order visibility for the NVIDIA Rubin supply chain.
• Changes in share for Taiwan server ODMs such as Hon Hai and Inventec in CPU tray, TPU rack, and v9 system projects.
• The competitive landscape among AVC, Fujikura, Boyd, and Cooler Master in TPU cold plate modules.
• Share and capacity utilization of Unimicron, EMC, ISU Petasys, WUS, TTM, and others amid ABF, CCL, and PCB specification upgrades.