“性能暴增5倍”，英伟达称已全面投产，不忘提中国

Core Insights - NVIDIA's CEO Jensen Huang announced the new Rubin platform at the Consumer Electronics Show, highlighting significant improvements in inference cost and training efficiency, with full production set to begin and shipments planned for the second half of 2026 [1][3] - Huang emphasized the strong demand for the H200 AI chips in the Chinese market, which have been approved for export, while the company awaits further government approvals for shipments [1][6] Group 1: New Product Launch - The new Vera Rubin platform integrates six new chips, achieving a training performance that is 3.5 times better than the previous Blackwell platform and improving AI software performance by 5 times [3][4] - The Rubin platform can reduce inference token generation costs by up to 10 times compared to Blackwell [4] Group 2: Market Dynamics - Huang addressed concerns about increasing competition in the AI sector and the sustainability of AI spending growth, aiming to alleviate fears of a potential "AI bubble" [4] - Despite NVIDIA's dominance in training AI models, it faces growing competition from traditional rivals like AMD and major clients such as Google [4] Group 3: Export and Regulatory Challenges - The U.S. government, under former President Trump, allowed NVIDIA to export H200 chips to China, with conditions related to national security and a 25% revenue share for the U.S. [6] - The H200 chip reportedly has nearly six times the performance of its predecessor, the H100, enabling Chinese AI labs to build supercomputers comparable to top U.S. models, albeit at a higher cost [6] Group 4: Chinese Market Response - There are uncertainties regarding whether Chinese buyers will purchase the H200 chips, as some are reportedly reluctant to buy the "special version" designed for China [7][8] - Chinese companies are actively developing domestic AI chips to replace NVIDIA's market share, with significant investments from major tech firms like Huawei, Alibaba, Tencent, and ByteDance [8]