Core Insights - Nvidia is transitioning from a chip company to an AI infrastructure company, focusing on building intelligent infrastructure based on power and internet, evolving AI from perception and reasoning to autonomous decision-making and physical AI [1][8][11] - The introduction of the GB300 chip, which offers a 1.5x increase in inference performance compared to the GB200, along with enhancements in memory and network performance, signifies a major advancement in AI hardware [2][33] - Nvidia is collaborating with partners to establish a large AI supercomputer in Taiwan, aimed at enhancing the AI infrastructure and ecosystem in the region [3][38] Group 1 - Nvidia's vision includes the development of physical AI, where AI-driven robots will perform real-world tasks, supported by advancements in 5G/6G and quantum computing [1][11] - The GB300 chip features a full liquid cooling design, compatibility with existing interfaces, and achieves 40 petaflops per node, showcasing significant performance improvements [2][33] - The NV Link Fusion technology allows for the creation of semi-custom AI infrastructure, enabling users to mix and match Nvidia CPUs, GPUs, and third-party hardware [2][46] Group 2 - The open-source Isaac Groot N1.5 platform is expected to drive the development of humanoid robots, with Nvidia's collaboration with DeepMind on the Newton physics engine [3][22] - Nvidia's new DGX Spark and DGX Station products are designed for AI developers and researchers, providing powerful computing capabilities for training large AI models [2][54][58] - The RTX Pro Enterprise server supports both traditional industrial software and AI models, facilitating a comprehensive design-simulation-optimization process [2][64] Group 3 - Nvidia's focus on AI factories represents a shift from traditional data centers to facilities that produce valuable outputs, termed "tokens," which are integral to AI operations [11][34] - The company emphasizes the importance of a robust software ecosystem and collaboration with over 150 partners to enhance the capabilities of its AI infrastructure [44][46] - The introduction of the AIQ query system aims to revolutionize data storage and retrieval, enabling efficient semantic searches and data management [73][76] Group 4 - Nvidia's advancements in quantum computing integration predict a future where all supercomputers will incorporate quantum processors, enhancing computational capabilities [22][31] - The development of agentic AI, which can perform tasks autonomously and collaborate with other AI systems, is seen as a critical evolution in AI technology [23][63] - The company is committed to enhancing enterprise IT by integrating AI capabilities without overhauling existing infrastructures, focusing on incremental upgrades [87]

证券研究报告|行业简评报告 2025 年 05 月 19 日 NVLink Fusion 助力多体系融合，持续布局机器人等领域 英伟达 COMPUTEX 跟踪报告 TMT 及中小盘/电子 事件： 英伟达 CEO 黄仁勋 5 月 19 日在 COMPUTEX 2025 大会发表主题演讲，介绍 公司在 AI 基础设施和"物理 AI"领域最新进展，以及 Blackwell 等硬件平台与 机器人等技术前沿应用。综合演讲及材料信息，总结要点如下： 评论： 1、构建智能基础设施迈向物理 AI，5G/6G 以及量子计算持续推进。 英伟达表示未来智能基础设施建立在电力和互联网基础上，AI 从感知和推理向 自主决策演进，未来将实现物理 AI，AI 驱动的机器人执行现实世界任务，英伟 达发展历程从芯片公司到 AI 基础设施公司，ShonaAriel、Kulitho、DynamoQDF 等函数库，科学领域、物理领域加速应用，持续推进 AI 向 5G/6G、量子计算 领域迈进。 2、GB300 推理性能提升 1.5 倍，NV Link Fusion 提供半定制芯片平台。 1）GB300：英伟达 25Q3 将推出 GB300，在保持相 ...

整理 | 褚杏娟 "十年后我们会说，我们需要无处不在的人工智能，就像互联网和电力一样。"英伟达 CEO 黄仁勋 5 月 19 日在 Computex2025 的开幕演讲的开始再次强调了 AI 及英伟达作为 AI 基础设施的重要性。 这次演讲中，黄仁勋分享了关于人工智能和机器人未来的观点以及一些新的计算产品。但与之前的传 言不同，黄仁勋并未在演讲中提及 CPU。 此外，演讲最后，黄仁勋还称将建立中国台湾省办事处。从公布的渲染图可以看到，英伟达在中国台 湾地区的办事处大楼采用了类似宇宙飞船的设计风格，命名为 NVDIA Constellation。黄仁勋表示， 该园区将尽快开始建造，为员工提供办公空间。 黄仁勋表示， Physical AI 是机器人革命的基石，能对物理环境进行推理等。从 "单次响应" 转向 "思 考型"、"推理型"、"推理时间扩展型" AI，需要更多计算。因此，英伟达创建了新系统 Grace Blackwell。 "我很高兴地宣布，我们的 Grace Blackwell 系统在经历了艰难的开发过程后，现在已全面 投入生产。基于 HGX 主板的 Blackwell 系统自去年以来已全面投入生产，并 ...