GF Securities: GTC Nvidia upgrades Agent computing power products, bringing new opportunities to the domestic AI industry.

GF SEC released a research report stating that at the GTC conference, NVIDIA Corporation (NVDA.US) showcased multiple new AI computing products, focusing on enhancing competitiveness in cluster computing and inferential computing products for Agent applications. The demand for inferential computing driven by Agent is rapidly increasing, and the process of domestic substitution of AI chips is expected to accelerate, with opportunities for further expansion in the long term. In addition, AI foundational software is also benefiting from the implementation and expansion of Agent-related applications. Key points from GF SEC include: At the GTC conference, NVIDIA Corporation showcased multiple new AI computing products for Agent applications. On March 16, 2026, NVIDIA Corporation presented multiple AI computing products at the GTC conference, including the Vera Rubin NVL72 super node product, Groq 3 LPU and LPX, and NemoClaw. From the direction of their product launches, NVIDIA Corporation is focusing on strengthening competitiveness in cluster computing and inferential computing product lines for Agent applications. Specifically, 1. Compared to the super node products under the Blackwell architecture, Vera Rubin NVL72 achieved a 5x improvement in inferential performance and a 3.5x improvement in training performance. The enhancement of the cluster capabilities of the Vera Rubin architecture is expected to better meet the computing demand for accelerating trillion-parameter AI models, multi-modal large models, and Agent inferential tasks. 2. For the common needs of long context and low latency in Agent inferential scenarios, NVIDIA Corporation introduced the dedicated chip Groq 3LPU. The LPU dedicated chip product, which integrates model and Agent algorithm principles, has shown significant improvement in computing performance, reflecting the increasingly clear trend of chip and algorithm fusion development. 3. For multi-agent collaboration scenarios, the Dynamo software stack achieves good performance improvement through KV-Cache storage optimization, large language model dynamic routing, and stepwise inferential technology. 4. The cuVS vector acceleration software stack mainly empowers data mining and semantic search scenarios by accelerating and optimizing the process of vector retrieval and search. 5. NemoClaw utilizes the NVIDIA Corporation Agent toolkit to optimize typical applications of OpenClaw. T...