Google TPU v8 Architecture Differentiation Benefits Optical Interconnect and PCB Supply Chain

Overview

This report provides an in-depth analysis of Google's eighth-generation Tensor Processing Unit (TPU v8) released at Cloud Next 2026, focusing on its differentiated network architecture design for training (v8t) and inference (v8i) scenarios. The report believes this architectural evolution not only enhances performance but also profoundly changes the optical interconnect demands of AI data centers, particularly increasing reliance on Optical Circuit Switches (OCS). Simultaneously, the report notes that as AI enters the Agent era, the role of CPUs in clusters is being reassessed, driving growth in related supply chains. Based on this, the report is optimistic about key component suppliers in Google's TPU supply chain, such as MediaTek and Victory Giant Technology.

Core views

Core View 1: TPU Architecture Differentiation and Network Topology Innovation. Google first divides TPU into v8t for large-scale pre-training and v8i for real-time inference. v8t introduces the Virgo network topology, achieving a 4x increase in data center network bandwidth; v8i adopts the new Boardfly topology, reducing latency by 50% through fewer hops. Both architectures increase demand for OCS at scale-up and scale-out levels to address communication bottlenecks in large-scale clusters. Core View 2: CPU Value Reassessment in the Agent Era. As AI workloads shift from simple interactions to complex Agent continuous tasks, data orchestration and memory management deepen reliance on CPUs. Intel's CEO notes the CPU:GPU ratio is evolving from 1:8 to 1:1. Google's TPU v8 first uses its self-developed ARM architecture Axion CPU as the main controller, with CPU hosts per server doubling, significantly boosting demand for CPU PCBs and related server manufacturing. Core View 3: Supply Chain Dual-Vendor Strategy and Volume Expectations. TPU v8t continues co-design with Broadcom, while v8i introduces MediaTek for the first time, leveraging its expertise in low-power SoCs to optimize inference chip efficiency. Counterpoint predicts TPU v8 series shipments will reach 5 million units by 2028, over 10x growth from 2026's ~400,000 units, directly benefiting upstream PCB, substrates, and optical communication solution providers.

Analysis framework

The report adopts a logical chain of 'technical architecture breakdown → hardware demand mapping → supply chain beneficiary analysis.' First, by comparing TPU v7 and v8 technical specifications, it identifies the evolution from 3D Torus to Virgo/Boardfly topologies and their pursuit of low latency and high bandwidth. Second, based on physical connection methods (copper vs. optical) and switching mechanisms, it deduces specific incremental demand for OCS, optical modules, and PCB layers. Finally, combining shipment forecasts and vendor collaboration patterns (Broadcom/MediaTek), it identifies specific beneficiaries and conducts valuation analysis.

Methodology notes

• Industry/Industrial Analysis FrameworkUpstream-Midstream-Downstream Transmission

  AI chip architecture changes to upstream hardware component demand transmission

  The report demonstrates how changes in downstream AI chip (TPU) network topologies (e.g., Boardfly) derive specific demand changes in midstream network equipment (OCS, optical modules) and upstream basic materials (PCB, CCL), reflecting the analytical logic of industrial chain demand transmission.

• Valuation methodsPE/PEG valuation

  Target price setting based on historical P/E ranges

  The report uses PE multiples for MediaTek and Victory Giant Technology's valuations, referencing historical P/E ranges or industry medians, combined with future profit forecasts to provide target prices, a common relative valuation method in the electronic hardware industry.

Asset mapping & comparison

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

• MediaTek (2454.TW)

  Beneficiary: First-time co-design with Google for TPU v8i inference chip

  Strengths

  Low-power design experience in mobile SoC market transferable to inference chips, strong cost optimization capabilities

  Comparison

  Compared to Broadcom's focus on high-performance training chips, MediaTek has advantages in inference efficiency

  Risks

  Price competition from Qualcomm and Unisoc; weak smartphone demand; ASIC execution risks

• Victory Giant Technology (300476.SS / 2476.HK)

  Beneficiary: PCB and substrate supplier in Google TPU supply chain

  Strengths

  Entered Google TPU v8 supply chain, benefiting from AI server PCB demand growth

  Risks

  Downstream server and automotive PCB demand below expectations; intensified high-end PCB market competition pressures margins; rising raw material costs; geopolitical tensions

Key data

• TPU v8t Pod Scale9,600 cardsIncreased from TPU v7's 9,216 cards
• TPU v8i Pod Scale1,152 cardsBoardfly topology, maximum latency of only 7 hops
• DCN Bandwidth Increase4xData center network bandwidth increase from TPU v8t's Virgo network
• 2028 TPU v8 Shipment Forecast5 million unitsOver 10x growth from ~400,000 units in 2026
• CPU:GPU Ratio Trend1:1Intel CEO's forecast for the Agent era, currently 1:4
• MediaTek Target PriceTWD 3,400Based on 25x PE of 2027-28 average EPS
• Victory Giant (H-shares) Target PriceHKD 479.00Based on 27x PE of 2027 EPS CNY 15.44

Impact & implications

The report believes TPU v8's release confirms the trend of AI network architecture optimization for specific workloads, making Optical Circuit Switches (OCS) from optional to essential for large-scale AI clusters, benefiting the optical communication supply chain. Simultaneously, MediaTek's entry into Google TPU inference chip design marks a significant breakthrough in the AI ASIC field, potentially opening new growth curves. For PCB manufacturers, increased demand for high-end HDI and multilayer boards will support earnings growth.

Risks

• MediaTek faces intense price competition from Qualcomm and Unisoc
• Intensified high-end PCB market competition may pressure margins
• Raw material cost pressures higher than expected
• Escalating geopolitical tensions in global AI value chains
• Key technology changes (e.g., COWOP) may alter competitive landscapes

What to watch

• Mass production progress of SiPh-based OCS and its penetration in TPU v8 clusters
• Whether TPU v8 series actual shipments align with Counterpoint's forecasts
• MediaTek's subsequent product rollout execution in the AI ASIC field