報告_Aletheia_AI網通_20261006

PDF 原檔：報告_Aletheia_AI網通_20261006_original.pdf

圖片清單（已驗證 2026-10-09）

本份 PDF 文字層因字型編碼損毀無法使用（抽出全為數字亂碼），正文改為逐頁 Read proceed/3.unprocessed_data/parsed/報告_Aletheia_AI網通_20261006_pages/p1.png～p8.png 人工轉寫（見下方「原始內容」）。本節為 docling 抽出 attachment 圖的獨立驗證。

原始內容（人工逐頁轉寫，原文為英文，保留原文不譯）

本報告 PDF 文字層損毀，以下為逐頁 Read 頁面圖片後的人工轉寫全文；圖表數字、公司名、評等均核對頁面圖片。第 7 頁（研究團隊聯絡資訊）精簡列出姓名與地區，不逐字列出電話/email；第 8 頁為 Firm Disclosures／Analyst Certification／Disclaimer，依規範不轉寫。

p1（封面）

AI Networking ／ #AI

Companies: Lumentum, Coherent, EZconn, Corning, Furukawa, Broadcom, Celestica, Arista, Google, Amphenol, TEC, Sumitomo, InnoLight

分析師：Skye Chen（+852 3470 0011，skye.chen@aletheia-capital.com）、Warren Lau（+852 9181 4766，warren.lau@aletheia-capital.com）

Upside Surprise from TPU8 Networking Upgrade

Google's networking supply chain — ethernet switches, optical transceivers, optical cables, laser diodes, and OCS systems — is positioned for strong earnings upside in 2027–28, supported by a 2–5x increase in networking-hardware content value per TPUv8 versus TPUv7. These elevated hardware requirements stem from Google's efforts to fully realise the Virgo (East–West) and Jupiter (North–South) architectures, which enable TPUv8 clusters to scale to 134,400 units, reduce latency and hop count, and double ICI bandwidth to 19.2 Tb/s per TPU. Lumentum, EZconn, and Celestica are likely to be key beneficiaries, given their significant sales exposure to Google and dominant market positions.

TPUv8 will see a big network upgrade on hardware and capex

At the Hot Chips Conference in August, Google further disclosed the networking architecture for TPU8t and TPU8i. To support its bandwidth of ICI doubling to 19.2Tb/s, expanding scalability, connecting with more different type of racks and with multi-data center sites, we see Google meaningfully upgrading its network architecture hardware and capex budget, with the aim to realise the Virgo (East-West) and Jupiter (North-South) network system.

• Ethernet switch: We estimate that the TPUv8 architecture will require more ethernet-switching capacity across the ToR layer, the Virgo East–West fabric, and the Jupiter North–South network. We estimate bandwidth per TPUv8 rising to roughly 2Tbps, up from about 0.4Tbps in TPUv7. In addition to this, we also see adoption of 200Gbps SerDes on TPUv8 network.
• Transceiver module and laser diode: TPUv8 node will require roughly three transceiver modules, up from fewer than two in TPUv7. This step-up in optical I/O also implies a substantial increase in laser-diode demand across 100Gbps, 200Gbps, and CW (continuous-wave) laser categories.
• Optical cable: Given the larger pod and cluster scale in TPUv8 — and the need to interconnect different rack types — the optical-cable count per TPU is expected to more than double, rising from 8 units in TPUv7 to roughly 18 units in TPUv8.
• OCS: We are seeing more adoption of OCS in TPU8 for the Jupiter network and TPU8i rack connections. Hence, we estimate total OCS demand for the TPU8 life cycle could be as much as 258,000 units compared to 20,800 units in TPU7 cycle.
• Copper cable (DAC/ACC cable): Although ICI SerDes will upgrade to 200Gbps from 100Gbps in TPU7, the adoption of Boardfly for TPU8i could reduce the number of copper cables per rack. We estimate copper cables per TPU8 will increase slightly to 1.9 from 1.5 for TPU7.

Beneficiaries of TPU v8 network upgrade

Meaningful TPU demand growth — from 3.6m units in 2026 to 8.0–8.5m units in 2027 — combined with strong rising networking-hardware content value per TPU, will materially benefit the hardware supply chain. Among them, Lumentum, EZconn, and Celestica warrant particular attention given their high sales exposure to Google and their dominant market position. We rate all three stocks as BUYs and have them in our Google G7 portfolio.

p2（Figure 1）

TPU8 will see a big network upgrade

At the Hot Chips Conference, a premier annual semiconductor industry event at Stanford University at the end of August, Google further disclosed the networking architecture for its upcoming TPU8t (Zebrafish – MediaTek) and TPU8i (Sunfish – Broadcom). To support its bandwidth of ICI (Inter-Chip Interconnect) doubling to 19.2Tb/s, expanding scalability, connecting with more different racks (such as general server and storage racks) and with multi-datacenter sites, we expect Google will meaningfully upgrade its network architecture and capex budget.

Google is employing two major network architectures — Jupiter (North–South traffic) and the newly announced Virgo (East–West traffic) — to expand TPUv8 scalability and integrate TPU clusters with storage systems, general-purpose servers, and Google's global WAN.

Jupiter handles North–South traffic, connecting TPU compute racks to storage, general servers, and Google's global WAN through a combination of ethernet switching and optical circuit switching (OCS). When paired with the new Boardfly ICI topology, TPU8i benefits from lower latency and fewer hops, improving inference responsiveness and model-weight access.

Virgo, designed for East–West traffic, is a co-design of hardware network architecture, optical switching, and software orchestration. This architecture dramatically increases horizontal scalability, enabling Google to expand the maximum TPU8t cluster size to 1 million units — a tenfold increase from the 100,000-unit limit of TPUv7 (Ironwood).

FIGURE 1: KEY NETWORK UPGRADE IN TPU V8 VS TPU V7

Source: Aletheia Capital/TAG

This major upgrade is highly reliant on networking hardware advancements, particularly in optical communications. Our deep-dive research into Google's new TPU network architecture leads us to conclude that we will see meaningful demand increase for ethernet switches, optical transceivers, laser diodes (both EML and CW lasers), optical cables, and optical circuit switches (OCS). For copper cabling within the TPU rack, we see a mixed trend: while TPU8 upgrading the SerDes to 200Gbps (up from 100Gbps in TPU7) would benefit content value, the adoption of the Boardfly topology could reduce the overall volume of copper cables required per rack in TPU8i (accounting for 50% of TPU8 volume).

• Ethernet switch: We estimate that the TPUv8 architecture will require more ethernet-switching capacity across the ToR layer, the Virgo East–West fabric, and the…（接續 p3）

p3（Figure 2、Figure 3）

…Jupiter North–South network. As a result, the aggregate network bandwidth per TPU is projected to increase sharply — rising to roughly 2Tbps per TPU, up from about 0.4Tbps in TPU7. In addition to this, we also see adoption of 200Gbps SerDes in TPU8 network.

• Transceiver module and laser diode: The increased reliance on ethernet switches in Virgo naturally drives higher demand for optical transceiver modules. The reduction in the number of TPUs per rack to 32 units from 64 will also further add to transceiver requirements. Based on our model, each TPU8 node will require roughly three transceiver modules, up from fewer than two in TPUv7. This step-up in optical I/O also implies a substantial increase in laser-diode demand across 100Gbps, 200Gbps, and CW (continuous-wave) laser categories.
• Optical cable: Given the larger pod and cluster scale in TPUv8 — and the need to interconnect different rack types — the optical-cable count per TPU is expected to more than double, from 8 units in TPUv7 to roughly 18 units in TPUv8.
• OCS: We are seeing more adoption of OCS in TPU8 for the Jupiter network and TPU8i rack connections. Hence, we estimate total OCS demand for the TPU8 life cycle could be as much as 258,000 units compared to 20,800 units in TPU7 cycle.
• Copper cable (DAC/ACC cable): Although ICI SerDes will upgrade to 200Gbps from 100Gbps in TPU7, the adoption of Boardfly for TPU8i could reduce the number of copper cables per rack. We estimate copper cables per TPU8 will increase slightly to 1.9 from 1.5 for TPU7.

FIGURE 2: ESTIMATE OF OPTICAL NETWORK HARDWARE DEMAND

Source: Aletheia Capital/TAG

Meaningful TPU demand growth—from 3.6m units in 2026 to 8.0–8.5m units in 2027—combined with strong rising networking-hardware content value per TPU, will materially benefit the hardware supply chain. Below we outline the major solution providers and manufacturers supporting Google's TPU networking ecosystem. Among them, Lumentum, EZconn, and Celestica warrant particular attention given their high sales exposure to Google and their dominant market position.

FIGURE 3: KEY SUPPLIER OF NETWORK HARDWARE FOR TPU8

Source: Aletheia Capital/TAG（原圖以粗體標示 Lumentum、Celestica、EZconn 為報告重點標的）

p4（Details of Virgo and Jupiter、Figure 4）

Details of Virgo and Jupiter

Google's upcoming TPUv8 generation is expected to introduce a differentiated networking architecture for 8i and 8t, reflecting their distinct roles within Google's AI stack. While the company has not fully disclosed the boundary between the two, industry signals suggest TPU8i is optimized for inference workloads, whereas TPU8t will lean toward training-optimized performance.

We observe meaningful architectural differences in how TPU chips are interconnected within each rack, as well as in the East–West rack-to-rack topology used to assemble a full TPU Pod. These variations reflect Google's workload-specific optimization for scaling compute domains and fabric bandwidth across the 8i and 8t configurations.

However, both TPU 8i and TPU 8t are expected to leverage Google's Jupiter networking fabric, using North–South traffic flows to interface with the broader data center stack. This architecture enables each TPU domain to connect efficiently to general-purpose servers, storage racks/servers, and ultimately to Google's global backbone network (Global WAN).

FIGURE 4: NETWORKING STRUCTURE FOR TPU8i AND TPU8t（見圖片清單 報告_Aletheia_AI網通_20261006_003.png；內容：左側 TPU8i (Sunfish) 經 OCS 與 Boardfly via DAC/ACC cable 串接 ToR 機架，SPOCS 負責 East-West 流量，單 pod 最高擴充至 1,152 顆 TPU；右側 TPU8t (Zebrafish) 經 OCS 與 3D Torus via DAC/ACC cable 串接，Virgo network（200G SerDes ethernet switch）負責 East-West 流量，單 pod 最高擴充至 9,600 顆並合併 14 pods（134,400 顆）；兩側 ToR 上方皆接 CPU-based general server、Storage rack/server，再經 Jupiter network（100G SerDes ethernet switch + OCS）接 Global WAN）

Source: Google at Hot Chips Conference and Aletheia Capital/TAG

TPU connection inside of rack

TPU v7 (Ironwood) uses a 3D-Torus interconnect implemented through ICI copper cabling. Google will continue to adopt 3D Torus for its TPU 8t (Zebrafish), aligning with its training-oriented design.

In contrast, TPU 8i will shift to Google's new Boardfly topology — a high-radix, multi-way hierarchical fabric that behaves more like a high-radix switch than a traditional torus. Boardfly is engineered for instant, any-to-any node communication, reducing hop count and latency, and is optimized for MoE (Mixture-of-Experts) and emerging agent-style inference architectures, where rapid expert routing and low-latency fan-out/fan-in patterns dominate.

p5（Figure 5、Figure 6）

FIGURE 5: GOOGLE'S NEW BOARDFLY STRUCTURE FOR ICI（見圖片清單 ..._004.png 及原圖右半 TPU 8t 3D Torus 部分；內容：左「TPU 8i (Sunfish) – ICI on Boardfly structure」三張示意圖——① Boardfly 4 TPUs on each tray fully connected、② Group of 8 fully connected trays（BB-BB#2 cu links）、③ Fully connected up to 36 groups（G-G-G-G）；文字框「Boardfly will reduce max number of hop to 7 from 16 that using 3D Torus: this could reducing demand of ICI copper cable per rack」。右「TPU 8t (Zebrafish) – ICI on 3D Torus」三張示意圖——3D Torus 座標網格圖、1 OCS／3 OCS 方塊圖、8x8x16 立方堆疊示意圖）

Source: Google at Hot Chips Conference

East-West (Virgo): building a Pod for TPUv8

TPU 8i will use the Boardfly topology inside each rack and extend it across racks through SPOCS (Optical Circuit Switch). This optical fabric allows Google to interconnect 36 racks, enabling 1,152 TPU8i chips to operate cohesively as a single Pod with low-latency, any-to-any communication that is optimized for popular MoE and AI agent.

For TPU8t, Google is introducing the new Virgo network architecture, which appears to rely primarily on ethernet-switch-based connectivity built on 200Gbps SerDes. Virgo provides the scale and bandwidth needed to aggregate roughly 300 TPU8t racks (9,600 TPUs) into a Pod and can further interlink 14 Pods into a single Virgo cluster comprising 134,400 TPUs.

FIGURE 6: VIRGO NETWORK ARCHITECTURE（原圖題文字為「VIFGO」，應為 Virgo 之誤植；見圖片清單 ..._006.png）

Source: Google at Hot Chips Conference

North-South network (Jupiter): storage, general server and global WAN

Jupiter network handles North–South traffic, including receiving external user inference requests (prompts), returning generated tokens to users, and loading the latest AI model weights from external high-bandwidth storage systems. Hence both TPU8i and TPU8t will connect to Jupiter. This network system is built by both OCS and ethernet switch.

p6（Figure 7）

FIGURE 7: JUPITER NETWORK ARCHITECTURE（與 Figure 6 同一張架構圖，紅色虛線圈改標示右側 Jupiter Network 區塊；見圖片清單 ..._007.png）

Source: Google at Hot Chips Conference

p7（研究團隊與聯絡資訊，精簡列出）

Alētheia Research Team

Macro/Strategy：Global Strategy – Jonathan Wilmot；Asianomics – Dr. Jim Walker；Tech Thematic Strategy – Keith Woolcock；Asia Equity Strategy – David Scott；China Strategy – Vincent Chan；Tactical Alpha Strategy – Justin Collazo；Global Commodities – Steven Schlegel。

Sector：Technology Hardware – Warren Lau（本報告作者之一）；Consumer & Internet – Nirgunan Tiruchelvam；China Technology – Eric Wen；Telecoms – Chris Hoare；CrossASEAN – Angus Mackintosh；India – Maulik Patel；MENA – Jaap Meijer。

Product Marketing Teams（Serving 300+ clients, 8,000+ touch points, in 20+ geographies with $15tr+ AUM）：Al Park（Americas-New York）、Linda Gustafsson（Europe-Nordics）、Michael Chambers（Asia-Singapore）、Graeme Bateman（Asia-Hong Kong）、Wayne Chang（Americas-New York）、Daren Riley（Europe-London）、Richard Wallace（Asia/Europe-Hong Kong）、Terrence Alford（Americas-Dallas）、Dhananjay Mahurkar（Europe-London）、Augustine Chen（Asia-Hong Kong）。

p8

Firm Disclosures／Analyst Certification and Disclosures／Disclaimer（免責聲明頁，依規範不轉寫）。