AI的进击时刻系列13：英伟达GTC大会开幕，走向代理式AI&物理AI时代

Investment Rating - The report recommends a "Buy" rating for multiple companies within the AI hardware industry, including Lixun Precision, Northern Huachuang, and others, with specific target prices set for each [5]. Core Insights - The report highlights the transition of AI from perception-based to generative AI, moving towards Agentic AI and Physical AI, as stated by NVIDIA's CEO Jensen Huang during the GTC conference [15]. - NVIDIA's new distributed inference service library, NVIDIA Dynamo, significantly enhances performance, achieving a 25-fold increase compared to the previous Hopper architecture [27]. - The anticipated demand for NVIDIA's Blackwell architecture is projected to reach 3.6 million units in 2025, indicating a robust growth trajectory for AI hardware [19]. Summary by Sections Section 1: NVIDIA's Transition to Agentic AI & Physical AI - Jensen Huang announced the evolution of AI technologies at the GTC conference, emphasizing the shift towards more advanced AI models [15]. - The scaling law for AI development is described as hyper-accelerated, focusing on pre-training, post-training, and testing time expansions [15]. Section 2: AI Models and Applications - The report discusses the increasing computational demands of AI models, with NVIDIA's Dynamo library optimizing inference processes to maximize token revenue [23][25]. - The performance of the Grace Blackwell NVLink72 chip is highlighted, showing a 25-fold improvement in efficiency when using the Dynamo service [27]. Section 3: Investment Recommendations - The report suggests that the rapid development of AI models and applications will benefit the AI hardware supply chain, with specific companies identified as key players in this growth [5]. - The introduction of new NVIDIA chips, such as Blackwell Ultra and Vera Rubin, is expected to further enhance AI processing capabilities, with significant performance metrics outlined [28][33].