Google's TPU v8 First-Ever Split into Training and Inference Versions; Optical Circuit Switching and PCB Supply Chains to Benefit

Overview

This research report, issued by Nomura, focuses on Google’s eighth-generation Tensor Processing Unit (TPU), unveiled at the Cloud Next conference in April 2026, and its innovative network architecture. The central thesis is that Google has, for the first time, explicitly segmented its TPU offering into dedicated training (TPU v8t) and inference (TPU v8i) variants — each optimized with purpose-built network topologies. This trend reflects the broader industry shift toward decoupling AI training and inference networks and will significantly drive demand across optical circuit switching (OCS), high-speed optical modules, and PCB/substrate supply chains. The report particularly highlights Victory Giant Technology (VGT) — a key PCB supplier in Google’s TPU supply chain — and assigns both its A- and H-shares a 'Buy' rating.

Core views

First-ever separation of training and inference chips, with distinct network architectures. Google’s TPU v8t targets large-scale pre-training and retains the 3D Torus topology, scaling up pod size to 9,600 accelerators and doubling bidirectional bandwidth to 19.2 Tb/s. v8t introduces the Virgo Network as its scale-out expansion fabric — a flat two-layer non-blocking architecture capable of increasing datacenter network bandwidth by 4x and supporting single training clusters comprising over one million chips. In contrast, TPU v8i is purpose-built for inference workloads, featuring the novel Boardfly topology — a flat two-layer architecture that reduces network diameter from 16 to 7 hops, improving latency by up to 50% for communication-intensive workloads. v8i also integrates a Collective Acceleration Engine (CAE), reducing on-chip collective communication latency by 5x and significantly boosting throughput in scenarios involving millions of concurrent AI agents. Boardfly architecture drives OCS demand from scale-out into scale-up. Boardfly’s hierarchical structure comprises: each tray (BB) formed by four chips interconnected via PCIe/PCB in a ring; eight BBs fully interconnected via copper cables to form a Group; and up to 36 Groups (supporting up to 1,024 active chips) connected via OCS to constitute a full Pod. This implies OCS is now deployed not only at the scale-out layer (between Groups) but also — for the first time — at the scale-up layer (between chip groups), generating incremental demand. Moreover, the report notes that silicon photonics (SiPh)-based OCS — capable of nanosecond-level switching — is well-suited to meet scale-up’s ultra-low-latency requirements. Key vendor iPronics has partnered with Fabrinet to build the world’s first mass-production line for SiPh OCS, scheduled to commence operations in Q2 2026. Rising Google TPU volumes and upgraded interconnect solutions benefit PCB and optical communication supply chains. Citing Counterpoint Research, the report forecasts combined TPU v8t/v8i shipments will surge from ~400,000 units in 2026 to nearly 5 million units by 2028 — over a tenfold increase. This volume growth will accelerate Google’s adoption of mainstream optical interconnect solutions, including 1.6T pluggable optics, NPO (Near-Package Optics), and CPO (Co-Packaged Optics). On the PCB/substrate front, Victory Giant Technology (VGT) is explicitly identified as Google’s supplier of UBB and CPU boards for TPUs. Furthermore, TPU v8 upgrades the CPU-to-TPU ratio from “1 CPU : 4 TPUs” to “1 CPU : 2 TPUs”, effectively doubling the number of CPU hosts — further expanding demand for CPU-specific PCBs.

Analysis framework

The report follows a structured analytical framework: 'Technical Architecture Decomposition → Incremental Demand Quantification → Supply Chain Mapping → Targeted Stock Valuation'. It begins with granular dissection of chip-level features (e.g., SparseCore, CAE, HBM capacity) and network topologies (3D Torus, Virgo, Boardfly) for both TPU v8t and v8i. Then, leveraging physical characteristics of competing OCS technologies (MEMS, DLC, SiPh), it evaluates which solution best satisfies scale-up’s stringent low-latency requirements. On the demand side, the report quantifies required components — DAC copper cables, optical modules, and OCS units — by reverse-engineering the BB-Group-Pod architecture of Boardfly, and cross-validates these estimates with Counterpoint’s shipment forecasts to project total market opportunity. Finally, technical roadmaps and demand projections are mapped to specific investment targets: For PCB/substrates, Victory Giant Technology is identified as the primary beneficiary based on Google’s supply chain mapping (including CCL material grade, layer count, and supplier distribution); for optical communications, the report comprehensively maps the end-to-end OCS supply chain — covering core components, global and domestic suppliers/manufacturers across MEMS, DLC, DLBS, and SiPh technology pathways. From a valuation perspective, Victory Giant’s A-share target price is derived using a 27x forward P/E multiple applied to its FY2027 earnings estimate — aligned with its historical median A-share P/E. The same multiple is applied to its H-share, yielding target prices of CNY 417.00 and HKD 479.00, respectively.

Methodology notes

• Industry/Industrial Analysis FrameworkUpstream-Midstream-Downstream Industry Chain Transmission

  End-to-end demand transmission: from chip design (TPU v8t/v8i) → network topology (Virgo/Boardfly) → interconnect components (DAC cables/optical modules/OCS) → PCB/substrates

  By deconstructing the technology stack layer-by-layer, the report establishes a causal logic chain: increased chip shipments → higher demand for networking infrastructure → greater usage of key components (PCBs, optical modules, OCS), enabling quantitative assessment of relative benefits across segments.

• Industry/Industrial Analysis FrameworkVolume-price decomposition

  Multiplying projected TPU shipment volume (quantity) by per-unit interconnect component requirements (price/unit value) to derive total market demand

  The report first cites TPU shipment forecasts (~5 million units by 2028), then calculates aggregate demand for DAC cables, optical modules, and OCS units by analyzing Boardfly’s architectural specifications — a classic volume-price decomposition methodology.

• Industry/Industrial Analysis FrameworkCost curve analysis

  Trade-offs among switching speed, cost, and performance across alternative OCS technology paths (MEMS, DLC, SiPh)

  The report compares millisecond-class switching speeds of MEMS/DLC versus nanosecond-class switching of SiPh, concluding SiPh is physically better suited for scale-up — though still in pilot production. Meanwhile, Google currently mitigates MEMS latency via software optimizations, indicating both approaches remain viable near-term options.

• Valuation MethodP/E or PEG Valuation

  Assigning target price based on historical median P/E multiple

  Victory Giant’s valuation applies a 27x FY2027 forward P/E — consistent with its historical A-share median P/E — representing a standard historical comparables approach, implicitly assuming future profitability remains broadly aligned with past performance.

• Company Fundamentals & Financial Framework

  Dual-vendor strategy analysis

  The report notes Google’s TPU v8 adopts a dual-vendor strategy: v8t co-designed with Broadcom, while v8i marks MediaTek’s debut collaboration with Google. Broadcom excels in peak compute and interconnect bandwidth, whereas MediaTek brings strengths in power efficiency and cost optimization — illustrating how Google matches design partners to chip use cases (training vs. inference).

Asset mapping & comparison

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

• Victory Giant Technology (300476 CH / 2476 HK)

  Core PCB/substrate supplier in Google’s TPU supply chain; supplies UBB and CPU boards across multiple TPU generations (v6p/v7p/v8t), directly benefiting from rising TPU volumes and increased per-unit PCB content

  Strengths

  Stable share within Google’s PCB supply chain; broad coverage of critical board types including TPU UBB and CPU boards; utilizes high-grade CCL materials (e.g., M8/M8.5), matching technical requirements for high-layer-count PCBs

  Comparison

  Competes with WUS, ISU, and TTM for Google TPU PCB business, yet the report exclusively assigns Victory Giant a 'Buy' rating and explicit target price

  Risks

  Lower-than-expected downstream demand for server/automotive PCBs; intensifying competition in high-end PCB segment pressuring margins; higher-than-anticipated raw material costs; worsening geopolitical friction across the global AI value chain

• MediaTek (2454 TT)

  First-time collaborator in Google TPU chip design, responsible for TPU v8i inference chip; leverages its low-power design expertise, well-aligned with inference chips’ power-efficiency requirements

  Strengths

  Low-power design experience accumulated in mobile SoC markets is transferable to inference chips; competitive advantages in power efficiency and cost optimization

  Comparison

  Complements Broadcom (responsible for v8t training chip): Broadcom leads in peak compute and interconnect bandwidth; MediaTek excels in power efficiency and cost optimization

  Risks

  Price competition from Qualcomm and UNISOC; execution risk on product specifications, pricing, and cost delivery; volatility in smartphone demand in China and emerging markets; ASIC execution and competitive risks

Key data

• TPU v8t/v8i Estimated Shipments in 2028Nearly 5 million unitsOver 10x growth from ~400,000 units in 2026 (Counterpoint forecast)
• TPU v8t Pod Scale9,600 acceleratorsAn increase from Ironwood’s 9,216 accelerators
• TPU v8t Per-Accelerator Scale-Up Bidirectional Bandwidth19.2 Tb/sDouble that of the prior-generation Ironwood
• TPU v8i Pod Scale1,152 chipsConnected via Boardfly topology; up to 1,024 active chips
• Boardfly Network DiameterReduced from 16 to 7 hopsLatency improvement of up to 50% for communication-intensive workloads
• TPU v8i CAE Acceleration Effect5x reduction in on-chip collective communication latencyDirectly enhances throughput for scenarios with millions of concurrent AI agents
• CPU Configuration UpgradeFrom 1 CPU : 4 TPUs to 1 CPU : 2 TPUsDoubling CPU host count; uses Google’s in-house ARM-based Axion CPU
• Victory Giant A-Share Target PriceCNY 417.00Based on 27x FY2027 forward P/E; ~9.9% upside from current price of CNY 379.30
• Victory Giant H-Share Target PriceHKD 479.00Based on 27x FY2027 forward P/E; ~9.7% upside from current price of HKD 436.80

Impact & implications

The report concludes that Google’s TPU v8 series — with its explicit separation of training and inference chips and dedicated network architectures — represents a pivotal trend in AI infrastructure. This trend implies: For the optical communications industry: OCS demand expands from scale-out to scale-up, creating pure incremental market opportunity; concurrently, surging TPU volumes accelerate adoption of advanced optical interconnects (1.6T, NPO/CPO), benefiting the entire optical module and optical switching supply chain. For the PCB/substrate industry: With TPU shipments growing more than tenfold and CPU count doubling, demand for high-layer-count, high-grade CCL-based PCBs will rise substantially. As a core Google TPU supplier — already providing UBB and CPU boards across v6p, v7p, and v8t generations — Victory Giant Technology stands to benefit continuously. For the chip design industry: The inference chip’s first-time collaboration with MediaTek signals a strategic shift in AI inference chips — from singular focus on raw compute power toward balanced optimization of compute, power efficiency, and cost — aligning closely with mobile SoC design philosophy and opening new growth avenues for MediaTek. Longer term, the CPU’s role in AI inference evolves from a preprocessing bottleneck to a continuously saturated main controller. Arm’s CEO projects the datacenter CPU market to exceed USD 100 billion by 2030 — benefiting both CPU PCB suppliers and server contract manufacturers.

Risks

• PCB Demand Risk: Downstream PCB demand from servers and automotive electronics may fall short of expectations.
• Intensifying Competition Risk: Growing competition in the high-end PCB market may pressure Victory Giant’s margins.
• Raw Material Cost Risk: Unexpected cost pressures on CCL and other raw materials may erode profitability.
• Geopolitical Risk: Escalating geopolitical friction across the global AI value chain could disrupt supply chain stability.
• Technology Roadmap Risk: The OCS technology path (MEMS vs. SiPh) remains unsettled; SiPh solutions are still in pilot production, introducing uncertainty.

What to watch

• Commercialization Progress of SiPh-Based OCS: Monitor timing of iPronics-Fabrinet mass-production line ramp-up and its ability to satisfy Google’s scale-up network requirement for nanosecond-level switching.
• TPU v8 Shipment Execution: Track actual shipment cadence and supply chain order confirmations against Counterpoint’s 2028 forecast (~5 million units).
• Google’s Optical Interconnect Upgrade Pathway: Monitor real-world deployment pace and market share evolution of 1.6T optical modules, NPO, and CPO within Google’s TPU clusters.
• Victory Giant’s TPU v8 PCB Share: Monitor shifts in supplier landscape and progress in certification for high-layer-count HDI PCBs.