Investment Rating - The industry investment rating is "Positive" [2] Core Insights - NVLink Fusion builds a semi-custom AI infrastructure, enhancing the ecosystem's moat. Nvidia's CEO Jensen Huang announced NVLink Fusion at COMPUTEX 2025, marking the opening of Nvidia's proprietary high-performance interconnect technology NVLink to partners for integrating third-party CPUs and AI accelerators, thus creating a semi-custom AI infrastructure. This aims to overcome traditional data center bottlenecks in scale and performance, providing more flexible and optimized system design solutions for cloud service providers and large enterprises [6][11] - The GB300 is expected to launch in Q3 2025, with multiple personal and enterprise products announced. The new AI computing platform Grace Blackwell and its upgraded version GB300 will offer 1.7 times the inference performance compared to the previous H100, equipped with 1.5 times the HBM memory and double the network bandwidth, achieving up to 40 petaflops per node. The Blackwell system will start mass production by the end of 2024, with GB300 expected to launch in the third quarter of 2025 [13][23] - The data center market is transitioning to a nearly trillion-dollar market driven by AI factories and infrastructure. Nvidia's CEO stated that the data center market is on the verge of a prosperous future, driven by AI factories and infrastructure. The expansion of AI infrastructure investment is expected to increase orders for quality companies in the domestic AI industry chain, with fundamentals likely to continue to deliver [23] Summary by Sections - **NVLink Fusion**: NVLink Fusion offers two main configurations: one connects third-party custom CPUs with NVIDIA GPUs via NVLink, and the other connects NVIDIA's Grace series CPUs with non-NVIDIA custom accelerators (GPU, ASIC, FPGA) to meet various computing needs. Initial adopters of NVLink Fusion include MediaTek, Marvell, Alchip, Astera Labs, Synopsys, and Cadence [7][11] - **GB300 Launch**: The GB300 is set to launch in Q3 2025, featuring significant performance improvements over its predecessor, the H100. The Blackwell system will be available on platforms like CoreWeave, and several new products targeting developers and enterprises were also introduced [13][14][16] - **Market Transformation**: The report emphasizes the transformation of the data center market into a trillion-dollar industry, highlighting the impact of AI infrastructure and factory investments on the domestic AI industry chain [23]

Investment Rating - The industry investment rating is "Positive" [2][29] Core Insights - NVLink Fusion builds a semi-custom AI infrastructure, enhancing the ecosystem's moat. Nvidia's CEO Jensen Huang announced NVLink Fusion at COMPUTEX 2025, marking the opening of Nvidia's proprietary high-performance interconnect technology NVLink to partners for integrating third-party CPUs and AI accelerators, thus creating a semi-custom AI infrastructure. This aims to overcome traditional data center bottlenecks in scale and performance, providing more flexible and optimized system design solutions for cloud service providers and large enterprises [6][11] - The GB300 is expected to launch in Q3 2025, with multiple personal and enterprise products announced. The new AI computing platform Grace Blackwell and its upgraded version GB300 were introduced, with GB300 offering 1.7 times the inference performance of the previous H100, equipped with 1.5 times HBM memory and 2 times network bandwidth, achieving up to 40 petaflops per node [13][23] - The data center market is transitioning to a nearly trillion-dollar market driven by AI factories and infrastructure. Nvidia's CEO stated that the data center is on the verge of becoming a trillion-dollar market, driven by AI factories and infrastructure. The expansion of AI infrastructure investment is expected to increase orders for quality companies in the domestic AI industry chain, with fundamentals likely to continue to deliver [23] Summary by Sections NVLink Fusion - NVLink Fusion provides two main configurations: one connects third-party custom CPUs with NVIDIA GPUs via NVLink, and the other connects NVIDIA's Grace series CPUs with non-NVIDIA custom accelerators (GPU, ASIC, FPGA) to meet various computational needs. Initial adopters of NVLink Fusion include MediaTek, Marvell, Alchip, Astera Labs, Synopsys, and Cadence [7][11] GB300 Launch - The GB300 is set to launch in Q3 2025, with significant performance improvements over its predecessor. The Blackwell system is expected to start mass production by the end of 2024 and has already been deployed on platforms like CoreWeave [13][23] Market Transformation - The report emphasizes the transformation of the data center market into a trillion-dollar industry, highlighting the impact of AI infrastructure and factory investments on the domestic AI industry chain [23]

梦晨 闻乐 发自 凹非寺 量子位 | 公众号 QbitAI 英伟达官宣新办公室落户中国台湾省台北市，但居然是从太空飞下来的吗？ 当然，这只是黄仁勋在Computex大会上展示的CG特效，但背后还有这样一层意思： 如此逼真的画面中， 只有1/10的像素是显卡渲染的 ，剩下全部内容为AI预测。 当初是GeForce显卡让AI成为现实，现在AI反过来给GeForce显卡带来颠覆性变革。 除了正式宣布新办公室地址，会上英伟达还宣布多款重磅新品： 面向个人有 DGX Station ，在桌面跑1T参数大模型，这回刀法也不用了，足足给到800GB内存。 面向企业有 RTX PRO Server ，可加速从AI智能体、物理AI、科学计算到渲染、3D图形和视频等各种用例。 同时老黄还表示英伟达正在经历一次转型， 从一家芯片公司扩展到一家AI基础设施公司 。 老黄表示，AI基础设施市场将以数万亿美元衡量。 此外，这场演讲中的其他亮点还有： 下面一起来看。 发布半定制化NVLink Fusion（混搭，有IP就可用） 推出英伟达AI数据平台 宣布与迪士尼合作机器人项目 …… "买的越多，创造的越多" 今年第三季度 ，英伟达主打产 ...