Shanxi Securities: New Nvidia (NVDA.US) products accelerate DCI construction, DeepSeek is expected to increase the adaptation of domestic chips

The Zhitong Finance App learned that Shanxi Securities released a research report saying that Nvidia will launch Spectrum-XGS Ethernet DCI products, and Nvidia's entry is expected to accelerate the construction of computing power network interconnection (DCI). Among them, long-range optical modules, L3 layer switches, and DCI converters will be the main beneficiaries. DeepSeekV3.1 was released, using UE8M0FP8scale parameter accuracy, which is expected to increase the adaptation of domestic chips. Recently, the market continues to rotate at an accelerated pace among the main sectors of AI computing power, and is expected to continue to reach new highs. Short-term suggestions focus on logic, focus on marginal changes in individual stocks, and use performance to set safety margins to seize the timing of pullback and increase positions. It is recommended to pay attention to Dekley (688205.SH), Optical Bank Technology (300620.SZ), Shengke Communics-U (688702.SH), and Xinke Mobile-U (688387.SH).

The main views of Shanxi Securities are as follows:

Nvidia (NVDA.US) will launch Spectrum-XGS Ethernet DCI products and redefine “scaleacross.”

Nvidia will officially launch Spectrum-XGS, an Ethernet-based DCI product in HotChip 2025. This product is used to connect data centers in different geographical locations to build a billion-watt AI gigafactory. Nvidia believes that due to the maximum power limit of a single cabinet and physical space limitations in data centers, SpectrumXGS solves communication delays, congestion and synchronization of cross-region GPU clusters through core algorithms (dynamic adaptation to long-distance network characteristics), hardware collaboration (based on ConnectX8 network cards, Spectrum-X switches, Blackwell architecture chips), and software full-stack optimization (Dynamo, SpeculativeDecoding) conundrum. The first users of this product will include CoreWeave, which is also expected to support the Stargate project promoted by Oracle, SoftBank, etc. Nvidia's entry into the market is expected to accelerate the construction of computational power network interconnection (DCI). Among them, long-range optical modules (including coherent, light coherent, and ZR), L3 layer switches, and DCI converters (transponders) will be the main beneficiaries. New optical fibers such as hollow fiber and G654E are also expected to emerge in new DCI projects. In addition, the domestic Ethernet AI network standard GSE led by China Mobile also has a “logical long container” technology design that can effectively solve the problem of bandwidth fluctuations between data centers across regions. It is recommended to focus on GSE participating companies such as network cards, switching chips/switches, and optical modules.

DeepSeekV3.1 was released, and the agent's capabilities were upgraded and natively optimized for the next generation domestic chip FP8.

On August 21, DeepSeekV3.1 was officially released. This upgrade mainly brings a hybrid inference architecture, higher thinking efficiency, and better performance in tool use and intelligent tasks. According to the evaluation, DeepSeekV3.1 significantly improved the weight of the programming agent and search smart body compared to the previous DeepSeekR1-0528. One of the highlights of DeepSeekV3.1 is the marked improvement in performance achieved after training based on the v3-based base model, proving that post-RL training is still an effective method that can continue to be expanded; second, it uses UE8m0fp8scale parameter accuracy. DeepSeek emphasizes that this is a specially optimized design for the next generation of domestic chips that will be released soon. The market has high expectations for the release of DeepSeek's next generation model. The mainstream trend should be a significant increase in chip performance under the chiplet architecture, a significant increase in MoE performance under the supernode architecture, and further enhancements of software and hardware collaboration (such as native support for FP8). Domestic chips may gradually change from “usable” to “easy to use”, and the performance of the next generation of Shengteng 910 chips or closer to mainstream international standards. The market space for second-tier GPU manufacturers is significantly opened up, and the flexibility associated with supernodes (such as copper cabletray connections, domestic switching chips, and all-optical modules under all-optical solutions) , OIO, etc.)

Recommended to follow

DCI: Decolly, Optical Bank Technology, Zhongji Xuchuang, Xinyisheng, Changfei Optical Fiber, Hengtong Optoelectronics;

OCS: Optical Bank Technology, Ling Yunguang, Tengjing Technology, Saiwei Electronics, Tianfu Communications, Changxin Bochuang;

Domestic supernodes: Senke Communications, Lixun Precision, Huiju Technology, Huafeng Technology, Wall Nuclear Materials, Ricoda, ZTE;

Commercial aerospace: ICT Mobile, Shanghai Hanxun, Tongyu Communications, Zhenlei Technology, Tianyin Electromechanical, Shanghai Port.

Risk Alerts

Demand for overseas computing power fell short of expectations, domestic operators and Internet investments fell short of expectations, fierce market competition led to price drops exceeding expectations, and external sanctions escalated.