
CITIC Securities: NVIDIA announces it will resume sales of H20 chips in China, focusing on server and cloud development

CITIC Securities released a research report stating that NVIDIA will resume the sales of the H20 chip in China. The H20 chip has competitive advantages in the software ecosystem and interconnect fields, and it is expected that the subsequent B30 products will also be launched. CITIC Securities recommends paying attention to server and cloud computing manufacturers, believing that the resumption of H20 sales will boost customers' capital expenditures. The H20 chip complies with the export restrictions set by the U.S. Department of Commerce, featuring a computing performance of 148Tflops@FP16 and an interconnect bandwidth of 900GB/s
According to the Zhitong Finance APP, CITIC Securities has released a research report stating that, according to NVIDIA's (NVDA.US) official website, the H20 chip will resume sales in China. The H20 chip still has competitive advantages in software ecology, interconnect, and other areas. CITIC Securities expects that subsequent B30 products are also likely to be launched, and downstream AI customers are expected to increase their Capex investment, thereby bringing opportunities for the development of the industrial chain. It is recommended to pay attention to: 1) server manufacturers; 2) cloud computing manufacturers.
The main points of CITIC Securities are as follows:
According to NVIDIA's official website, on July 15, 2025, NVIDIA announced that it would resume sales of the H20 chip in China. In this regard, we comment as follows:
The H20 chip is a product designed to meet the restrictions imposed by the U.S. Department of Commerce's Bureau of Industry and Security, and it still has advantages in software ecology and cluster interconnect.
In 2022, U.S. BIS export control measures targeted the total processing power (TPP) of chips and interconnect communication speed: 1) chips with I/O bandwidth transmission rates greater than or equal to 600Gbyte/s; 2) TPP greater than or equal to 4800TOPS are prohibited from export. In 2023, the U.S. BIS further tightened restrictions on TPP and PD (performance density): (1) TPP exceeding 4800 or TPP exceeding 1600 and PD exceeding 5.92; (2) TPP in [2400, 4800), and PD in [1.6, 5.92), or TPP in [1600, +∞), and PD in [3.2, 5.92). As long as either condition is met, it is restricted. The H20 chip emerged in this context, and according to NVIDIA's official website, its peak computing performance is 148Tflops@FP16, with NVLink interconnect achieving 900GB/s bandwidth. The H20 weakens peak computing performance but retains interconnect bandwidth capability to some extent, thus still having advantages in cluster interconnect and other aspects, while NVIDIA's CUDA ecosystem also makes it an important option for downstream customers such as the internet.
The resumption of H20 sales may further boost customer Capex, and it is recommended to pay attention to the development of server and cloud manufacturers.
According to the annual performance communication meetings of leading internet manufacturers, each manufacturer is actively investing in the development of the artificial intelligence industry, maintaining high levels of investment in Capex. The H20 chip has advantages in cluster interconnect and software ecology, and its resumption of sales may further enhance customer investment in this field, promoting the development of the artificial intelligence industry. We expect downstream AI manufacturers to increase their investment in AI infrastructure in the second half of 2025 to meet the future demands of artificial intelligence development, and related segments of the industrial chain are expected to welcome development opportunities. Server manufacturers are likely to directly benefit from the shipment of related products and achieve performance growth, while cloud computing companies are expected to provide services to downstream customers with more sufficient chip supply.
Risk factors:
The risk of artificial intelligence industry development falling short of expectations; the risk of capital expenditure by major internet companies falling short of expectations; the risk of overseas advanced chips being subject to regulatory control; the risk of limited production capacity supply, etc