证券研究报告 2025年09月17日 计算机 从Blackwell到Rubin：计算、网络、存储持续升级 ——AI算力"卖水人"专题系列（7） 评级：推荐(维持) 刘熹(证券分析师) S0350523040001 liux10@ghzq.com.cn 最近一年走势 相关报告 -12% 13% 37% 62% 86% 111% 2024/09/12 2025/09/12 计算机 沪深300 《从Tokens角度跟踪AI应用落地进展——计算机行业大模型及AI应用专题 （推荐）*计算机*刘熹》——2025-09-15 核心提要 B300基于Blackwell Ultra架构，采用TSMC 4NP工艺与CoWoS-L封装，FP4浮点算力可达15PFLOPS，是B200的1.5倍。GB300搭载288GB的 HBM3E显存。Vera Rubin NVL144性能是GB300 NVL72的3.3倍。Rubin Ultra NVL576性能较GB300 NVL72提升14倍，内存是GB300的8倍。 FY2026Q2，英伟达营收467亿美元，同比+56%。 《计算机事件点评：甲骨文RPO增至4550亿美元，AI算力强力增长 ...

March 19, 2025 12:41 PM GMT We provide read-across for our Asia supply chain coverage following Nvidia's recent GTC event. Key Takeaways: Nvidia CEO, Jenson Huang, discussed GB300 NVL72 coming in 2H25, Vera Rubin NVL144 in 2H26, Vera Rubin Ultra NVL576 in 2H27, and Vera Feynman in 2028. Nvidia Photonics was also announced, with Spectrum-X Photonics in 2H25 and Quantum-X Photonics in 2H26. Please see Joe Moore's note here for more details. Greater China Technology Hardware | Asia Pacific GTC 2025 – Implicati ...

点击招商研究小程序查看PDF报告原文 事件： 英伟达GTC 2025大会于3月17-21日举办，英伟达CEO黄仁勋于3月19日发表主题演讲，介绍AI加速计算 新时代、AI加速计算的新品、CPO芯片及集成光子引擎、以及AI技术在自动驾驶、机器人等前沿领域 的应用。综合演讲及材料信息，总结要点如下： 评论： 1、数据中心28年Capex望超万亿美元，训练推理带来100倍Tokens增长。 1）资本开支： 数据中心、CSP和相关基础设施的资本开支将大幅增加，预计2028年数据中心建设将达 到1万亿美元，绝大部分投资或将用于加速计算芯片。2024年公司向四大云服务商共交付130万颗Hopper GPU，预计2025年将出货360万颗Blackwell。英伟达认为，世界正在经历一场革命性转型，全球数据中 心正在向人工智能工厂转型，计算机将为软件生成Tokens，处理分析并生成AI驱动的研究和应用。 2） 训练及推理： 英伟达概述了三层次的AI，即Generative AI、Agentic AI以及Physical Al，从生成式到AI 推理再到理解物理层面；由于训练和推理模型的计算量剧增，AI基础设施在短时间内增长惊人 ...

Investment Rating - The report maintains a "Buy" rating for the company, indicating an expected price increase of over 15% in the next 6-12 months [4]. Core Insights - The company is expected to benefit as a leading AI chip manufacturer due to rapid hardware iteration and a rich software ecosystem, which enhances its competitive edge against rivals [4]. - The demand for AI computing power is anticipated to remain strong, driven by the complexity of models in the inference stage, which require significantly more computing resources compared to earlier generative AI models [2][4]. Summary by Sections Performance Review - The company held the GTC 2025 event on March 18, 2025, showcasing future product launches including GB300 (GB Ultra), Vera Rubin, Rubin Ultra GPUs, and CPO switches for Infiniband and Ethernet [1]. Operational Analysis - The transition from simple generative AI to assistant AI is expected to sustain demand for computing power, with inference requiring 100 times more tokens than before. The company anticipates strong customer demand, with major cloud providers expected to purchase 3.6 million Blackwell GPU dies in 2024 [2]. - Upcoming product releases include GB300 NVL72 in H2 2025, which will feature 288GB HBM3e memory and 1.5 times the computing power of GB200 NVL72. Vera Rubin NVL144 is expected in H2 2026, offering 3.3 times the computing power of GB300 NVL72, and Rubin Ultra NVL576 is projected for H2 2027, with 14 times the computing power of GB300 NVL72 [2]. Software Ecosystem - The company continues to enhance its software ecosystem, launching libraries tailored for various industries, such as cuLitho for lithography and CUDA-Q for quantum computing. Additionally, the introduction of the Dynamo system aims to improve GPU efficiency by assisting with prefill and decode tasks [3]. Profit Forecast and Valuation - The company forecasts net profits of $122.2 billion, $156.9 billion, and $177.9 billion for FY26, FY27, and FY28, respectively, with corresponding P/E ratios of 23, 18, and 16 [4][6].