Gf Securities has released a research report stating that at the GTC conference, NVIDIA showcased multiple new AI computing products, with a focus on enhancing the competitiveness of its product lines for cluster computing and inference computing tailored for Agent applications. The rise of Agents is driving a rapid increase in demand for inference computing power, which is expected to accelerate the process of domestic substitution for AI chips and open up further long-term growth potential. Additionally, foundational AI software is also benefiting from the implementation and expansion of Agent-related applications. The main viewpoints of Gf Securities are as follows:

At the GTC conference, NVIDIA showcased multiple new AI computing products designed for Agent applications. On March 16, 2026, NVIDIA presented several AI computing products at the GTC conference, including the Vera Rubin NVL72 super-node product, Groq 3 LPU and LPX, and NemoClaw. Judging from the direction of the launched products, NVIDIA is focusing on strengthening the competitiveness of its product lines for cluster computing and inference computing specifically for Agent applications.

Specifically: 1. Compared to super-node products based on the Blackwell architecture, the Vera Rubin NVL72 achieves a 5x improvement in inference performance and a 3.5x improvement in training performance. The enhanced clustering capabilities of the Vera Rubin architecture are expected to better meet the computing power demands of technology companies for accelerating trillion-parameter AI models, multimodal large models, and Agent inference tasks. 2. To address the common requirements of long context and low latency in Agent inference scenarios, NVIDIA launched the dedicated Groq 3 LPU chip. This dedicated LPU chip product, which integrates model and Agent algorithm principles, shows significant improvements in computational performance, reflecting the increasingly evident trend of convergence between chip design and algorithms. 3. For multi-agent collaboration scenarios, the Dynamo software stack achieves notable performance improvements through KV-Cache storage optimization, dynamic routing for large language models, and step-by-step reasoning techniques. 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 utilizes the NVIDIA Agent toolkit to optimize typical applications of OpenClaw; its launch validates the viewpoint from a previous report that "small-scale agents may change the future architecture, channels, and operational systems of software applications, becoming a focal point of competition."

Agents are driving a rapid increase in demand for inference computing power, opening the door for domestic substitution of AI chips. At this GTC conference, NVIDIA not only enhanced the computing performance related to Agents at the hardware level, including chips and super-nodes, but also further adapted its software stacks, such as Dynamo and NemoClaw, for Agent applications. This reflects the trend that future Agents will lead to a rapid increase in demand for inference computing power. On one hand, due to policy influences, the sales of NVIDIA's AI chips, including the Vera Rubin, still face significant uncertainty in the domestic market. On the other hand, as inference AI chips have lower performance requirements, it is less difficult for domestic AI chips to catch up technically with overseas counterparts represented by NVIDIA. Under this trend, the process of domestic substitution for AI chips is expected to accelerate, potentially opening up further long-term space.

Furthermore, foundational AI software also benefits from the implementation and expansion of Agent-related applications. It is recommended to focus on: 1. AI hardware: Cambricon, Inspur Information, Unisplendour Corporation. 2. Models: Zhipu AI, MiniMax, Alibaba, Tencent. It is also suggested to monitor SenseTime and iFlytek. 3. Foundational AI software: Transwarp Technology,卓易信息,范式智能. 4. Data center operation and scheduling services: Wangsu Technology, Baosight Software, YunSai ZhiLian. It is also suggested to monitor Capital Online.

Risk warnings include limited production capacity for AI chips; the widening gap between China and the US in the field of AI computing power, presenting challenges for the domestic AI industry chain to catch up; and policy uncertainties affecting the supply of AI chips.