The Zhitong Finance App learned that Guoxin Securities released a research report saying that AI is changing the computing power cluster network architecture. On the one hand, the scale of the single training cluster continues to expand, and the proportional relationship between connectivity requirements and acceleration units is expected to further increase; on the other hand, distributed computing power clusters are expected to boost the 10-20km DCI interconnection demand and accelerate the “related decline.” Taken together, Marvell believes that the growth elasticity of demand for optical connectivity is expected to surpass acceleration units, and that the optical connectivity market is expected to maintain a high boom. We continue to recommend optical devices and optical modules, focusing on Zhongji Xuchuang (300308.SZ), Tianfu Communications (300394.SZ), Xinyisheng (300502.SZ), Guangxun Technology (002281.SZ), etc.

The main views of Guoxin Securities are as follows:

On April 11, EST, Marvell hosted the Accelerated Infrastructure for the AI Era event to introduce the company's product layout and strategic direction for AI infrastructure. In 2023, the MarvelLAi-related business achieved revenue of 550 million US dollars, accounting for 10% of the company's overall business revenue; revenue is expected to exceed 1.5 billion US dollars in 2024, of which the connectivity business is expected to achieve revenue of 1 billion US dollars, and the custom computing business is expected to achieve revenue of 500 million US dollars, accounting for 30% of the company's overall revenue. During the event, Marvell CTO Noam Mizrahi shared his insights on how AI is reshaping network architectures. He believes that in large-scale AI clusters, the growth rate of optical connectivity is expected to surpass the growth rate of acceleration units (XPUs).

AI changes the computing power cluster network architecture, and DCI interconnection creates a new application market. Marvell divides the internal network connections of AI clusters into three types — namely computing unit internetworks for chip interconnection, front-end networks for AI server interconnection, and back-end networks for data input and output; AI clusters need to be connected to external networks through DCI networks. Considering that the scale of AI clusters continues to expand and distributed computing power clusters generate interconnection requirements with a distance of 10-20 km, DCI interconnection is expected to generate new application fields.

The growth elasticity of optical connectivity in large-scale AI clusters is expected to surpass acceleration units (XPUs). Guoxin Securities pointed out that in terms of demand elasticity, when the scale of the acceleration unit of the computing power cluster breaks through to the 100,000 or 1 million levels, the corresponding number of layers of the network may increase to 5-10 layers, then the ratio of acceleration units to optical modules is expected to rise from the current 1:2-1:3 to 1:5-1:10, and the usage elasticity of the cluster optical module will surpass that of the acceleration card. At the same time, in terms of price flexibility, AI shortens the speed upgrade cycle of optical connectivity from 4 to 2 years, and high-speed product iteration drives value increase.

The AI cluster interconnection market is expected to remain booming. Guoxin Securities said that from Marvell's connectivity business as an example, the CAGR is expected to reach 27% in 2023-2028. Among them, the corresponding CAGR for optical interconnection and copper interconnection within data centers is 25% and 59%, respectively, the data center interconnect (DCI) CAGR is 25%, and the AI cluster interconnection market continues to boom.