The Internet of AI Era: Demand, Innovation, and Investment Opportunities

With the emergence of large models, the network has become a key component in the AI era. This article explores why the network has become the "C position" in the AI era from a theoretical perspective, and discusses future innovations and investment opportunities on the network side.

1. Sources of Network Demand

In the era of large models, network demands mainly come from three aspects:

1. The increasingly large model size: single-card computing power can no longer meet the demand, and multi-server clusters have become a necessary choice.

2. Complex communication of multi-card synchronization: After model partitioning, alignment is required for each computation, which raises higher demands on network transmission.

3. Increasingly Expensive Failure Costs: Training large models takes months, and interruptions can result in significant losses; network stability is crucial.

2. Direction of Network Innovation

In response to AI demand, network innovation is mainly focused on the following directions:

1. Communication medium iteration: Optical modules pursue higher speeds while exploring cost reduction paths such as LPO and silicon photonics. Copper cables dominate in-cabinet connections due to their cost-effectiveness.

2. Network Protocol Competition: The inter-chip communication protocol is strongly bound to the graphics card, such as NV-LINK. The competition for inter-node communication is mainly between IB and Ethernet.

3. Changes in Network Architecture: The leaf-spine architecture is facing challenges, and new architectures such as Dragonfly and rail-only are expected to become the direction for ultra-large clusters.

3. Investment Recommendations

Core links of communication systems: Zhongji Xuchuang, Xinyi Sheng, Tianfu Communication, Industrial Fulian, Yingweik, Hu Dian Co., Ltd.

Innovation in Communication Systems: Changfei Fiber Optics, Taicheng Optoelectronics, Yuanjie Technology, Shengkai Communication-U, Cambrian, Dekoli.

4. Risk Warning

• AI demand is below expectations
• Scaling law failure
• Industry competition intensifies