Guangfa Securities: GTC NVIDIA upgrades Agent computing power products, a new opportunity for the domestic AI industry

According to the Zhitong Finance APP, GF Securities released a research report stating that at the GTC conference, NVIDIA (NVDA.US) showcased several new AI computing products, focusing on enhancing competitiveness in the cluster computing and inference computing product lines for Agent applications. The demand for inference computing driven by Agents is rapidly increasing, and the process of domestic substitution for AI chips is expected to accelerate, with further long-term potential being opened up. Additionally, AI foundational software will also benefit from the implementation and expansion of Agent-related applications.

The main points from GF Securities are as follows:

At the GTC conference, NVIDIA showcased several new AI computing products for Agent applications

On March 16, 2026, NVIDIA presented multiple AI computing products at the GTC conference, including the Vera Rubin NVL72 supernode product, Groq 3 LPU and LPX, as well as NemoClaw, among others. From the direction of the products launched, NVIDIA is focusing on strengthening competitiveness in the cluster computing and inference computing product lines for Agent applications.

Specifically:

1. Compared to the supernode products under the Blackwell architecture, the Vera Rubin NVL72 achieves a 5-fold improvement in inference performance and a 3.5-fold improvement in training performance. The enhanced cluster capability of the Vera Rubin architecture is expected to better meet the computing demands of technology companies for accelerating trillion-parameter AI models, multimodal large models, and Agent inference tasks.
2. To address the common demands of long context and low latency in Agent inference scenarios, NVIDIA introduced the dedicated chip Groq 3LPU. The LPU dedicated chip product, which integrates model and Agent algorithm principles, shows significant improvements in computing performance, reflecting the increasingly evident trend of the integration of chips and algorithms.
3. For multi-agent collaborative scenarios, the Dynamo software stack achieves better performance improvements through KV-Cache storage optimization, dynamic routing of large language models, and step-by-step inference technology.
4. The cuVS vector acceleration software stack primarily empowers data mining and semantic search scenarios by accelerating and optimizing the process of vector retrieval and search.
5. NemoClaw optimizes the typical application of OpenClaw using NVIDIA's Agent toolkit; the launch of NemoClaw validates the previously reported viewpoint that "crawfish may change the future software application architecture, channels, and operational systems, becoming a battleground for entry."

Agent drives rapid increase in demand for inference computing, opening up space for domestic substitution of AI chips

At this GTC conference, NVIDIA not only strengthened the performance of Agent-related computing capabilities at the hardware level, such as chips and supernodes, but also further adapted Agent applications through software stacks like Dynamo and NemoClaw. This reflects the trend of rapidly increasing demand for inference computing driven by Agents in the future. On one hand, due to policy impacts, the sales of NVIDIA AI chips, including Vera Rubin, in the domestic market still face significant uncertainties; on the other hand, since the performance requirements for inference AI chips are relatively low, the technical difficulty for domestic AI chips to catch up with overseas AI chips represented by NVIDIA is lower. Under this trend, the process of domestic substitution for AI chips is expected to accelerate, with further long-term potential being opened up In addition, AI infrastructure software also benefits from the implementation and expansion of Agent-related applications.

Recommendations

① AI Hardware: Cambricon, Inspur Information, Unisplendour. ② Models: Zhiyuan, MiniMax, Alibaba, Tencent, with a focus on SenseTime and iFlytek. ③ AI Infrastructure Software: StarRing Technology, ZTE Information, Paradigm Intelligence. ④ Data Center Operations and Scheduling Services: Wangsu Science & Technology, Baoxin Software, Yunsai Zhilian, with a focus on Capital Online.

Risk Warning

The risk of limited production capacity for AI chips; the widening gap between China and the U.S. in AI computing power, posing challenges for the domestic AI industry chain to catch up; policy uncertainties affecting the supply of AI chips