Evercore的分析师Amit Daryanani在周四的报告中指出，如果假设GPU占AI集群总支出的80%，那么总成本可能达到25亿美元，其中10%可能用于基础设施。他进一步分析称：“如果Arista能够赢得这项业务，这可能意味着2.5亿美元的收入机会。”According to Evercore ISI's analysis, Arista Networks is likely to become a network partner for the large-scale training cluster that Meta is about to establish.
According to the analysis of Evercore ISI, a Wall Street investment institution, Arista Networks (ANET.US) is likely to become a network partner for Meta Platforms (META.US)'s upcoming large-scale model training cluster. This training cluster is expected to be supported by over 0.1 million GPUs, with most likely being Nvidia's NVIDIA H100 model. This news drove Arista's stock price up nearly 4% on Thursday. If Arista successfully secures this contract, it will help the company achieve its $750 million AI revenue target by 2025. Evercore maintains an 'outperforming the market' rating for Arista and continues to adhere to a target stock price as high as $400.
It has been revealed that Meta is making final adjustments to its training cluster, which will be used to train its Llama 4 large language models. This massive project is expected to be equipped with $2 billion worth of chips. Since InfiniBand technology cannot meet the needs of this scale cluster, the cluster will be interconnected using Ethernet, leading Evercore to speculate that Arista Networks may provide some of the infrastructure.
Evercore analyst Amit Daryanani pointed out in Thursday's report that if GPUs account for 80% of total AI cluster spending, total costs could reach $2.5 billion, with 10% possibly allocated to infrastructure. He further analyzed, "If Arista can secure this business, it could mean a revenue opportunity of $0.25 billion."
Evercore also noted that Meta had previously used Arista's switches to interconnect a cluster containing 24,000 GPUs. Daryanani added, "Given Arista's close collaboration with Meta, we believe they are likely to provide switch infrastructure for this new cluster."
