国泰海通证券研究·2026-01-07 15:07Core Insights - Nvidia showcases comprehensive AI infrastructure deployment and iteration, emphasizing the application prospects of physical AI [1][3] - AMD and Intel have made significant progress in previously weaker areas, such as rack capabilities and the 18A process node [1][4] Group 1: Nvidia Developments - Nvidia's new AI platform Rubin has entered full-scale production, featuring six new chips: Vera CPU, Rubin GPU, NVLink 6 switch, ConnectX-9 super network card, BlueField-4 DPU, and Spectrum-6 Ethernet switch [3] - Rubin GPU achieves 3.5 times the training speed and 5 times the inference speed compared to the Blackwell architecture, with a peak computing power of 50 Petaflops [3] - The token cost during inference is reduced by up to 10 times compared to the Blackwell platform, and the number of GPUs required for training mixture of experts (MoE) models is reduced to one-fourth [3] - Nvidia introduces a memory storage platform driven by BlueField-4 to address KV Cache issues, enhancing token processing speed by up to 5 times in specific scenarios [3] - Microsoft and Coreweave are set to be the first customers to deploy Rubin in the second half of 2026, with Microsoft's next-generation Fairwater AI super factory scaling to hundreds of thousands of Vera Rubin chips [3] - Nvidia's Alpamayo series VLA open-source AI models and tools are aimed at autonomous vehicle development, with the DRIVE system entering mass production for the Mercedes-Benz CLA, expected to hit the US roads in 2026 [3] Group 2: AMD and Intel Innovations - AMD launches the Helios rack based on the MI 455X, featuring a fully liquid-cooled design with four Instinct MI455X GPUs and one EPYC Venice Zen6 CPU [4] - The MI500 series chips are in development, expected to enhance AI performance by 1000 times by 2027 [4] - Intel introduces its first computing platform based on the 18A process node, the Intel Core Ultra 3 series (codenamed Panther Lake), achieving a total AI computing power of 180 TOPS, with the GPU contributing 120 TOPS [4] - The Core Ultra 3 platform supports running a 70 billion parameter large language model locally with 32k context, with consumer laptops featuring this processor set to pre-sell on January 6, 2026 [4]