黄仁勋罕见提前宣布：新一代GPU全面投产

Core Viewpoint - NVIDIA has accelerated its AI chip platform release with the introduction of the "Rubin" platform at CES 2026, marking a shift in its product announcement strategy and emphasizing the growing demand for AI computing in both training and inference [2][4]. Group 1: Rubin Platform Overview - The Rubin platform, which integrates six chips including the NVIDIA Vera CPU and Rubin GPU, is designed for extreme collaborative performance, enhancing AI training performance by 3.5 times and operational performance by 5 times compared to the previous Blackwell architecture [4]. - The Rubin platform's inference token cost can be reduced by up to 10 times, and the number of GPUs required for training MoE models is decreased by four times [5]. - The NVL72 system, which includes 72 GPU packaging units, was also announced, with each unit containing two Rubin dies, totaling 144 Rubin dies in the system [5]. Group 2: Ecosystem and Market Response - Major cloud providers and model companies such as AWS, Microsoft, Google, OpenAI, and Meta have shown strong interest in the Rubin platform, indicating a positive market response [6]. - NVIDIA's early announcement of Rubin aims to provide engineering samples to ecosystem partners for preparation of subsequent deployment and scaling applications [6]. Group 3: Full-Stack AI Strategy - NVIDIA's presentation at CES included a range of AI products, signaling a shift from training scale to inference systems, with the introduction of the Inference Context Memory Storage Platform designed for efficient management of KV Cache [8]. - The company is expanding its focus on physical AI, releasing open-source models and frameworks that extend AI capabilities into robotics, autonomous driving, and industrial edge scenarios [8][9]. - The Cosmos and GR00T series models were introduced for robotics, enabling machines to understand and act in the physical world, while the Alpamayo model family was launched for autonomous driving, supported by a high-fidelity simulation framework [11].