GF SEC: AI data center network multi-layer interconnection trend confirmed, continue to pay attention to investment opportunities in the industry chain.

GF Securities released a research report stating that with the surge in demand for AI, there are bottlenecks in power and capacity within individual data centers, and interconnection between data centers has become an industry trend. NVIDIA's Spectrum-XGS Ethernet technology can combine multiple distributed data centers into a billion-watt AI super factory, significantly improving multi-node communication performance. Leading optical module companies are actively expanding into the Scale-Up market, and they are continuing to monitor investment opportunities in these companies. On August 21st, NVIDIA announced the launch of NVIDIASpectrum-XGS Ethernet. This cross-regional scaling technology combines multiple distributed data centers into a billion-watt AI super factory. Key points from GF Securities include: - Bottlenecks in power and capacity within individual facilities, with interconnection between data centers becoming an industry trend - With the increasing demand for AI, there are bottlenecks in power and capacity within individual data centers. Breaking through the limitations of buildings and overcoming existing Ethernet network infrastructure bottlenecks such as high latency, high jitter, and performance is necessary to expand data centers. NVIDIA CEO Jensen Huang stated that the AI industrial revolution has arrived, and giant AI factories are the core infrastructure of this transformation. Following vertical and horizontal scaling technologies, NVIDIA Spectrum-XGS Ethernet provides cross-regional scaling technology to combine data centers from different cities, countries, and even continents into massive billion-watt AI super factories. - Improved multi-node communication performance, CSP accelerated deployment - Spectrum-XGS Ethernet is fully integrated into the Spectrum-X platform, dynamically adapting distances between networks and data center facilities through algorithms. Spectrum-XGS Ethernet has nearly doubled the performance of the NVIDIA collective communication library, accelerating the communication performance of multiple GPUs and nodes, achieving predictable performance for remote AI clusters. North American cloud factories have begun accelerated deployment of the Spectrum-XGS platform, with CoreWeave being one of the first companies to use Spectrum-XGS Ethernet. The Spectrum-X Ethernet network platform is designed for ultra-large-scale AI factories, offering 1.6 times higher bandwidth density than traditional Ethernet, enabling ultra-low latency. NVIDIA Spectrum-XGS Ethernet is now available as an extension of the NVIDIA Spectrum-X Ethernet platform. - Optical module companies entering the Scale-Up market, with continued focus on investment opportunities in leading optical module manufacturers - To ensure that subsequent network architectures meet the rapidly growing demand for computing power, NVIDIA plans to introduce optical interconnection in the Scale-Up field, with leading optical module companies making strategic moves. From AOCs to light sources to optical engines, the industry believes that leading optical module companies will help CSP customers reduce equipment costs, drive the iteration of new technologies in the Scale-Up market, and play a significant role, due to their strong R&D capabilities, flexible solution design, and agile global delivery. Continued recommendation for leading optical module companies Zhongji Innolight/Eoptolink Technology Inc. Risk factors include: - Risks of AI application development falling short of expectations - Risks of local shortages in AI hardware components - Risks of stricter overseas export controls - Risks of cloud providers' capital expenditures falling short of expectations.