Recently, the market's focus on AI infrastructure has shifted to the network communication end. The impressive second quarter performance has driven the stock price of communication chip leader Broadcom to a historic high, with the latest market value reaching $778.1 billion. Arista in the same industry also achieved a new high on the same day, with a market value exceeding 100 billion US dollars.
Broadcom
Arista
Both companies are excellent network infrastructure suppliers dedicated to Ethernet solutions. The former's main products include Ethernet switching chips, packet processors, ASCI, etc., while the latter is the leader in high-end Ethernet switches, mainly provided for large-scale data centers such as Microsoft and Facebook.
Similar to the stock price trend of Nvidia, the stock prices of Broadcom and Arista began to rapidly rise around the turn of 2022 and 2023, which coincided with the release and popularity of ChatGPT.
After the stock price reaches a new high, will it be another new high? Wall Street analysts are very optimistic about this track.
Bank of America stated that Broadcom, like Nvidia, is the best choice in the field of AI. After releasing impressive performance reports, Broadcom still has greater potential and is likely to become the next member of the trillion dollar club in the US stock market. The Ethereum network is seen as an important driving force for its growth.
What is the relationship between Ethernet and AI?
Networking devices represented by switches are an important component of AI computing hardware. In terms of network route selection, currently, NVIDIA's IB (Infini Band) solution is at the forefront in the field of AI connectivity, with advantages over Ethernet solutions.
IB has better performance, with advantages such as high bandwidth, low latency, high reliability, and high scalability. However, the premium is high, and users often need to pay more invisible networking costs. Under the same budget, the number of graphics cards, optical modules, and switch hardware purchased will be smaller.
Ethernet, as the most widely used communication protocol in the past few decades, has a large audience, high user acceptance, and uses universal network devices, with lower costs and strong compatibility. However, due to its limited use in high-performance AI computing scenarios and dispersed suppliers in the past, the comprehensive effect of Ethernet in the early stages of AI development was somewhat lacking compared to the closed Nvidia solution.
The cost advantage of Ethernet is highlighted, and the switch/switching chip and optical module industries are expected to benefit
The high cost of computing infrastructure is a major pain point in the development of AI. With the acceleration of the AIGC development process, the demand for cluster capabilities and computing power is becoming larger, and the cost advantage of Ethernet is gradually becoming apparent. According to a research report by Guosheng Securities, in Meta's testing, the Ethernet networking price is less than half of IB, and the performance is 10% higher, confirming that Ethernet is buy more save more.
Nvidia has chosen to take a "two pronged approach". In his speech at the Computex conference on June 2nd, Huang Renxun officially showcased Nvidia's layout in the Ethernet switching system. In response, Guosheng Securities believes that Nvidia's entry into Ethernet is a reconfirmation of the demand for AI networks, and it will also drive Ethernet into a new rate update cycle.
In July 2023, the Ultra Ethernet Consortium (UEC) was established, with members including AMD, Arista, Broadcom, Cisco, Meta, and Microsoft, aiming to address many shortcomings in the practical application of Ethernet.
Zhongtai Securities stated that with the establishment of UEC, several issues with Ethernet running AI workloads may be effectively resolved, but they do not believe that IB and Ethernet are in a completely replacement role, or that Ethernet will replace IB networks. For a longer period of time, thanks to the high iteration efficiency of Nvidia GPU products, IB networks will occupy a higher market share. In the short term, Ethernet benefits more from factors such as market, cost, and commercial competition. And the ultimate results of the Super Ethernet Alliance and Ethernet in product/technical practice also need to be verified to some extent. But regardless of the protocol, exchanging underlying chips and exchanging whole machine products will benefit.
Guosheng Securities also stated that NVIDIA's bet on Ethernet is not about being pessimistic about its own IB, but about seizing a broader AI inference market in advance. Regardless of which solution (IB vs. Ethernet) dominates, it will promote the expansion of the switch market space and the acceleration of the overall development of the industry. It is recommended to pay attention to the switch industry, as chip, manufacturing, and brand manufacturers are expected to benefit from the gradual increase in penetration rate of Ethernet solutions in AI application scenarios.
In addition, the institution stated that one of the biggest advantages of NVIDIA/UEC's AI Ethernet solution over the IB solution is its low cost. In order to continue to implement the system's low-cost and low-power characteristics, it also tends to use low-cost LPO (Linear Drive Pluggable Optical Module) solution in the selection of optical modules. Therefore, the penetration rate of LPO is expected to accelerate in the future. However, due to the need to cooperate with switches, the requirement for upstream and downstream cooperation and collaboration among optical module manufacturers in the industry is higher. Leading companies such as Zhongji Xuchuang and Xin Yisheng will have more advantages.
Guotai Junan believes that we should attach importance to investment opportunities in NVIDIA, as well as optical module suppliers and switches in the industry under the trend of using Ethernet solutions for networking.
