广发证券:AI数据中心网络多层互连趋势确定 持续关注产业链投资机会
Zhi Tong Cai Jing·2025-08-26 07:09

Core Insights - The rapid increase in AI demand has led to power and capacity bottlenecks within individual data center facilities, making interconnection between data centers an industry trend [1] - NVIDIA's launch of Spectrum-XGS Ethernet technology enables the combination of multiple distributed data centers into a billion-watt AI super factory, significantly enhancing multi-node communication performance [1][2] Group 1: Data Center Trends - Individual data center facilities are experiencing power and capacity bottlenecks due to surging AI demand, necessitating the need for interconnection between data centers [1] - NVIDIA's CEO Jensen Huang stated that the AI industrial revolution is underway, with giant AI factories being the core infrastructure of this transformation [1] Group 2: Spectrum-XGS Ethernet Technology - Spectrum-XGS Ethernet is fully integrated into the Spectrum-X platform, dynamically adapting the distance between network and data center facilities through algorithms [2] - The Spectrum-XGS Ethernet has nearly doubled the performance of NVIDIA's collective communication library, accelerating multi-GPU and multi-node communication capabilities [2] - North American cloud service providers (CSPs) have begun accelerating the deployment of the Spectrum-XGS platform, with CoreWeave being one of the first companies to utilize this technology [2] Group 3: Light Module Companies and Investment Opportunities - To meet the rapidly growing computing power demands, NVIDIA plans to introduce optical interconnects in the Scale-Up sector, prompting leading optical module companies to make strategic moves [3] - The report suggests that leading optical module companies will leverage their strong R&D capabilities, flexible solution designs, and agile global delivery to help CSP clients reduce costs and drive new technology iterations in the Scale-Up market [3] - Continued recommendations are made for leading optical module companies such as Zhongji Xuchuang (300308) and Xinyi Sheng (300502) [3]