Shanxi Securities: NVIDIA's new products accelerate DCI construction, DeepSeek is expected to increase the adaptation of domestic chips

According to the Zhitong Finance APP, Shanxi Securities released a research report stating that NVIDIA will launch the Spectrum-XGS Ethernet DCI product. NVIDIA's entry is expected to accelerate the construction of the Data Center Interconnect (DCI) network, with long-distance optical modules, L3 switches, and DCI converters being the main beneficiaries. The DeepSeek V3.1 has been released, utilizing the UE8M0FP8 Scale parameter precision, which is expected to enhance the adaptation of domestic chips. Recently, the market continues to accelerate rotation along the main line of AI computing power, with the potential to reach new highs. In the short term, it is recommended to focus on the logical space and pay attention to the marginal changes of individual stocks, while setting a safety margin based on performance to seize the opportunity for increasing positions during pullbacks. Recommended stocks include Dekeli (688205.SH), Guangku Technology (300620.SZ), Shengke Communication-U (688702.SH), and Xinke Mobile-U (688387.SH).

The main points of Shanxi Securities are as follows:

NVIDIA (NVDA.US) will launch the Spectrum-XGS Ethernet DCI product and redefine "scale across."

NVIDIA will officially launch the Ethernet-based DCI product Spectrum-XGS at Hot Chips 2025, which is designed to connect data centers in different geographical locations to create a billion-watt AI super factory. NVIDIA believes that Scale-up and Scale-out can no longer meet the gigabit-level AI demands due to the power limits of single cabinets and the physical space constraints of data centers. Spectrum-XGS addresses the communication latency, congestion, and synchronization challenges of geographically distributed GPU clusters through core algorithms (dynamically adapting to long-distance network characteristics), hardware collaboration (relying on ConnectX8 network cards, Spectrum-X switches, and Blackwell architecture chips), and full-stack software optimization (Dynamo, Speculative Decoding). The first batch of users for this product will include CoreWeave, and it is also expected to support Oracle, SoftBank, and other companies in advancing the Stargate project. NVIDIA's entry is expected to accelerate the construction of the DCI network, with long-distance optical modules (including coherent, light-coherent, and ZR), L3 switches, and DCI converters (Transponders) being the main beneficiaries. New types of optical fibers such as hollow-core fibers and G654E are also expected to shine in new DCI projects. In addition, the domestically produced Ethernet AI network standard GSE led by China Mobile has a "logical long container" technology design that can effectively address bandwidth fluctuations between regional data centers, and it is recommended to pay attention to companies participating in GSE, such as network cards, switching chips/switches, and optical modules.

DeepSeek V3.1 released, upgrading agent capabilities and natively designed for next-generation domestic chip FP8 optimization.

On August 21, DeepSeek V3.1 was officially released, bringing a hybrid inference architecture, higher thinking efficiency, and better performance in tool usage and agent tasks. According to evaluations, DeepSeek V3.1 shows significant improvements in programming agents and search agents compared to the previous DeepSeek R1-0528 One of the highlights of DeepSeek V3.1 is the significant improvement in performance achieved through post-training based on the V3 base model, proving that RL post-training is still an effective way to continue expanding; secondly, it uses the UE8M0FP8Scale parameter precision, and DeepSeek emphasizes that this is specifically optimized for the upcoming next-generation domestic chips. The market has high expectations for the release of DeepSeek's next-generation model, with the mainstream trend expected to be a substantial improvement in chip performance under chiplet architecture, a significant enhancement in MoE performance under supernode architecture, and further strengthening of software-hardware collaboration (such as native support for FP8). Domestic chips may gradually transition from "usable" to "user-friendly," with the performance of the Ascend 910 next-generation chip potentially approaching mainstream international levels, significant market space opening up for second-tier GPU manufacturers, and considerable elasticity related to supernodes (such as cable tray copper connections, domestic switching chips, optical modules under all-optical solutions, OIO, etc.).

Recommended Focus

DCI: Dekoli, Guangku Technology, Zhongji Xuchuang, NewEase, Changfei Fiber Optics, Hengtong Optoelectronics;

OCS: Guangku Technology, Lingyun Optics, Tengjing Technology, Saiwei Electronics, Tianfu Communication, Changxin Bochuang;

Domestic Supernodes: Shengke Communication, Luxshare Precision, Huiju Technology, Huafeng Technology, Wolong Nuclear Materials, Ruikeda, ZTE;

Commercial Aerospace: Xinke Mobile, Shanghai Hanxun, Tongyu Communication, Zhenlei Technology, Tianyin Electromechanical, Shanghai Port Bay.

Risk Warning

Overseas computing power demand may fall short of expectations, domestic operators and internet investments may not meet expectations, intense market competition may lead to price declines exceeding expectations, and external sanctions may escalate