GTC大会召开，催生供应链新增量

Investment Rating - The industry rating is optimistic, expecting overall returns to exceed the CSI 300 index by more than 5% in the next six months [64]. Core Insights - The release of the Blackwell Ultra platform marks the beginning of a new era in AI inference, with significant enhancements in training and testing capabilities [5][10]. - The Rubin Ultra NVL576 is anticipated to deliver a substantial performance boost, with a focus on vertical scaling [19][41]. - Changes in the GB300 and Rubin platform architectures are expected to stimulate new supply chain growth [3][48]. Summary by Sections I. Blackwell Ultra Platform Release - The Blackwell Ultra platform was unveiled at the GTC conference, aiming to enhance inference capabilities through increased computational power [5][10]. - The platform features a single card FP4 floating-point performance of 15 PetaFlops, a 50% increase compared to the B200 [10]. - Memory capacity has been upgraded to 288GB, also a 50% increase from the previous generation [14]. II. Rubin Ultra NVL576 Performance Enhancement - The upcoming Rubin Ultra NVL576 is expected to launch in the second half of 2027, featuring a performance increase of 14 times compared to the GB300 NVL72 [41]. - It will include HBM4 memory and achieve FP4 floating-point performance of 15 ExaFlops, significantly surpassing previous models [41][43]. III. GB300 & Rubin Platform Architecture Changes - The transition to a UBB+OAM architecture in the GB300 will allow for more flexible customer configurations compared to the previous Bianca board [48]. - The introduction of PTFE backplane architecture is anticipated to enhance transmission efficiency due to its low dielectric loss [49]. - Companies in the supply chain, such as Huadian Co., Shenghong Technology, and Jingwang Electronics, are expanding their capacities to meet the growing demand in AI-related sectors [52][55][58].