2025年全球人工智能网络超级周期_由技术升级供应短缺驱动报告_76页_16mb

报告摘要

Summary of ANCHOR REPORT: Global AI Networking Supercycle

Core Content

The Anchor Report from Nomura highlights the ongoing global AI networking supercycle that is expected to continue in 2026 and extend into 2027, driven by technological upgrades and supply shortages. Key technologies include Silicon Photonics (SiPh), Co-packaged Optics (CPO), Optical Circuit Switches (OCS), Active Electronic Cable (AEC), and Hollow Core Fiber (HCF). These innovations are critical for meeting the growing demand for high-speed, low-latency interconnects in AI data centers, particularly for large language model (LLM) training and inference.

The report emphasizes the importance of optical components in the AI infrastructure upcycle, particularly in the context of scale-up, scale-out, and scale-across network architectures. It also analyzes the supply chain dynamics and market strategies of key players, including Nvidia, Broadcom, Google, and Chinese firms such as InnoLight, YOFC, TFC, and T&S Communications.

Main Points and Key Insights

AI Networking Supercycle: The global AI networking supercycle is expected to continue in 2026F and 2027F, fueled by technology upgrades and supply constraints.
800G/1.6T Transceivers: Shipments of 800G and 1.6T transceivers are projected to grow significantly, with SiPh transceivers capturing 50-70% of the market.
Market Leaders: Chinese firms InnoLight, YOFC, TFC, and T&S Communications are identified as key beneficiaries of the AI networking upcycle.
CPO Migration: CPO is expected to gain momentum in 2026 due to aggressive bundling strategies by global AI leaders. TFC and T&S Communications are seen as potential beneficiaries.
YOFC Upgrade: YOFC is upgraded to Buy from Neutral due to buoyant AIDC demand and stabilized telecom market.
AI Networking Roadmaps: The report outlines the networking roadmaps of major AI players, including Nvidia, Google, and Broadcom, and their interconnect technologies such as NVLink, PCIe, Ethernet, and InfiniBand.
Scale-up, Scale-out, Scale-across: These are the three primary levels of AI data center expansion. Scale-up involves connecting multiple GPUs within a rack, Scale-out expands the number of server nodes, and Scale-across connects multiple data centers into a large AI factory.
Interconnect Standards: NVLink, UALink, SUE, and PCIe are key standards for scale-up networks, while InfiniBand, Ethernet, and OCS are critical for scale-out and scale-across.
CSPs' Strategies: Global CSPs are increasingly adopting open protocols and custom solutions to reduce costs and improve flexibility, which may impact the margins of branded switch manufacturers.

Stocks for Action

Company	Ticker	Rating	Mkt Cap (USD mn)	Avg. TO (USD mn)	Target Price (LC)	Price (As of 8 Jan)	Upside (%)	Company Description	NV's Revenue Exposure % in FY25F	Other Global AI Tech Players' Revenue Exposure % in FY25F
InnoLight	300308 CH	Buy	95,604	2,547	799.00 ↑	595.45	34.2%	Global No.1 DC optical transceiver supplier	25%~30%	50%~60%
YOFC	6869 HK	Buy	†	2,727	142	64.50 ↑	28.0%	One of China's leading optical fiber makers	N.A	10~20%
TFC	300394 CH	Buy		21,742	1,026	243.00 ↑	24.5%	High-end MPO product maker, key partner to Corning	50%~60%	10~20%
T&S Communications	300570 CH	Buy		3,774	260	148.00 ↓	27.6%	Core supplier to Corning	0~10%	60%~70%

Key Technology Trends

Scale-up Networks: Dominated by copper interconnects (e.g., NVLink), but optical solutions are becoming necessary due to the increasing size of GPU clusters.
NVLink Fusion: Nvidia is opening its ecosystem to third-party vendors, allowing integration with custom chips and heterogeneous architectures.
PCIe 7.0: Offers doubled bandwidth compared to PCIe 6.0 and supports optical connectivity for long-distance transmission.
Ethernet for AI: UEC (Ultra Ethernet Consortium) is developing open protocols for AI training with microsecond latency and multipath transmission.
Optical Circuit Switches (OCS): Used by Google in TPU systems, providing high performance in large-scale AI clusters.
CPO Adoption: Accelerating due to increased demand from global AI leaders, with InnoLight, TFC, and T&S Communications as key beneficiaries.
Hollow Core Fiber (HCF): A promising technology for scale-across networks, offering low latency and high bandwidth over long distances.

Conclusion

The global AI networking supercycle is driven by technological innovation and supply constraints, with optical interconnects playing a central role. InnoLight, YOFC, TFC, and T&S Communications are positioned to benefit from the upgrading of transceivers and adoption of CPO. Nvidia and Google are leading the charge with their interconnect technologies, such as NVLink, UALink, and OCS, which are shaping the future of AI data center infrastructure. The report underscores the importance of optical components in the AI networking ecosystem and highlights the strategic positioning of key firms in this rapidly evolving market.