GF Securities: GTC Nvidia Upgrades Agent Computing Power Products, Domestic AI Industry Welcomes New Opportunities

When it comes to stock trading, rely on Golden Kylin Analyst Reports—authoritative, professional, timely, comprehensive—helping you uncover potential thematic opportunities!

(Source: Zhihui Finance)

Zhihui Finance APP has learned that GF Securities released a research report stating that at the GTC conference, NVIDIA (NVDA.US) showcased several new AI computing products, focusing on strengthening competitiveness in Agent applications within clustered computing and inference product lines. The demand for inference computing driven by Agents is increasing rapidly, and the process of domestic AI chip replacement is expected to accelerate, with further long-term potential opening up. Additionally, AI foundational software also benefits from the implementation and expansion of Agent-related applications.

GF Securities’ main points are as follows:

At the GTC conference, NVIDIA presented multiple new AI computing products targeting Agent applications

On March 16, 2026, NVIDIA showcased several AI computing products at the GTC conference, including the Vera Rubin NVL72 supernode, Groq 3 LPU and LPX, and NemoClaw. Based on their product directions, NVIDIA is focusing on enhancing competitiveness in Agent applications within clustered computing and inference product lines.

Specifically:

① Compared to supernode products based on the Blackwell architecture, Vera Rubin NVL72 achieves a 5-fold increase in inference performance and a 3.5-fold increase in training performance. The enhanced clustering capability of the Vera Rubin architecture is expected to better meet the computing needs of tech companies for accelerating trillion-parameter AI models, multimodal large models, and Agent inference tasks. ② To address common requirements in Agent inference scenarios such as long context and low latency, NVIDIA launched the dedicated chip Groq 3LPU. This LPU chip, integrating model and Agent algorithm principles, shows significant improvements in computing performance, reflecting a clear trend of integrating chip and algorithm development. ③ For multi-agent collaboration scenarios, the Dynamo software stack improves performance through KV-Cache storage optimization, dynamic routing of large language models, and step-by-step inference techniques. ④ The cuVS vector acceleration software stack mainly accelerates and optimizes vector retrieval and search processes, empowering data mining and semantic search scenarios. ⑤ NemoClaw uses NVIDIA’s Agent toolkit to optimize typical applications like OpenClaw; its launch confirms the earlier report that “Lobster could change future software application architectures, channels, and operational systems, becoming a key entry point.”

Agents are driving rapid growth in inference computing demand, opening opportunities for domestic AI chip replacement

At this GTC conference, NVIDIA not only strengthened Agent-related computing performance at the hardware level with chips and supernodes but also further adapted to Agent applications through software stacks like Dynamo and NemoClaw. This reflects a trend of rapidly increasing inference computing demand driven by Agents in the future. On one hand, due to policy influences, sales of NVIDIA AI chips including Vera Rubin in China still face significant uncertainty; on the other hand, because inference AI chips have lower performance requirements, domestic AI chips are technically closer to catching up with overseas AI chips led by NVIDIA. Under this trend, the process of domestic AI chip replacement is expected to accelerate, with further long-term potential opening up. Additionally, AI foundational software also benefits from the implementation and expansion of Agent-related applications.

Recommendations to watch:

① AI hardware: Cambrian, Inspur, Unisplendour. ② Models: Zhipu, MiniMax, Alibaba, Tencent; also focus on SenseTime and iFlytek. ③ AI foundational software: Xinghuan Technology, Zhuoyi Information, Paradigm Intelligence. ④ Data center operation and scheduling services: Wangsu Technology, Baoxin Software, YunSai ZhiLian; also consider Capital Online.

Risk warnings:

Limited AI chip production capacity; widening gap in AI computing power between China and the US, challenges in domestic AI industry chain catch-up; policy uncertainties affecting AI chip supply.