Nvidia(US:NVDA) 硬AI·2026-01-06 01:40Core Insights - The new Rubin AI platform from NVIDIA significantly enhances performance, achieving 3.5 times the training performance and 5 times the performance for running AI software compared to the previous Blackwell platform [2][7] - The platform is set to be delivered to initial customers in the second half of 2026, marking NVIDIA's commitment to annual updates in the AI chip sector [3][5] Performance Enhancements - Rubin platform reduces the cost of inference token generation by up to 10 times and decreases the number of GPUs required for training mixed expert models by 4 times compared to Blackwell [7] - The Vera CPU integrated into the platform features 88 cores, providing double the performance of its predecessor, and is designed for efficient inference in large-scale AI factories [8] Chip Testing Progress - All six Rubin chips have returned from manufacturing partners and have passed critical tests, indicating that NVIDIA is on track to maintain its leadership in AI accelerator manufacturing [10] - The platform incorporates five innovative technologies, including the sixth-generation NVLink interconnect technology and a second-generation RAS engine for real-time health checks and fault tolerance [10] Ecosystem Support - Major cloud service providers such as Amazon AWS, Google Cloud, Microsoft, and Oracle Cloud are set to be the first to deploy instances based on the Vera Rubin platform in 2026 [12] - Prominent figures in the AI industry, including OpenAI's CEO and Meta's CEO, have expressed optimism about the Rubin platform's potential to enhance model capabilities and efficiency [12][13] Early Product Disclosure - NVIDIA has disclosed product details earlier than in previous years, aiming to maintain its position as a critical hardware provider in the industry [15] - The new hardware will also include networking and connectivity components, which will be part of the DGX SuperPod supercomputer and available as standalone products for modular use [15]