华为云昇腾AI云服务行业：6A云化算力底座

Investment Rating - The report does not explicitly state an investment rating for the industry or company Core Insights - The emergence of large models has triggered an exponential growth in global computing power demand, with AI computing requirements increasing by 300,000 times from 2012 to 2023, and expected to grow another 500 times in the next decade [11][12] - Huawei Cloud's Ascend AI Cloud Service is positioned as the best cloud-based full-stack computing service for the era of large models, providing comprehensive support for model training, inference, and application development [17][18] - The AI industry is experiencing a paradigm shift, moving from specialized models to general-purpose models capable of handling diverse applications through extensive pre-training on massive datasets [8][12] Summary by Sections Section: Global Computing Demand - The demand for computing power has surged due to the rise of large models, with video generation models like Sora requiring significantly more computing resources compared to traditional models [11][12] - The report highlights that the scale of datasets will increase from a few terabytes to 100 terabytes, and the token length for models will expand from thousands to hundreds of thousands [11] Section: Full-Stack Computing Services - Huawei Cloud's Ascend AI Cloud Service integrates cluster computing, computing engines, and AI development frameworks to provide stable and reliable full-stack computing support for large model training and inference [12][18] - The service offers various computing usage, management, and deployment modes to cater to diverse business needs [19][21] Section: Business Innovation Focus - The report emphasizes that enterprises require full-stack computing services to focus on business innovation, leveraging accumulated training and operational experience to avoid redundant problem-solving [14][19] - The Ascend AI Cloud Service supports a wide range of AI frameworks and models, facilitating rapid application development and deployment [68][70] Section: Performance and Recovery - The service boasts rapid fault recovery capabilities, with industry practices showing that cluster faults can be detected in one minute, diagnosed in five minutes, and recovered in ten minutes [36][40] - The report indicates that the average interruption for large model training in the industry occurs every 2.8 days, with recovery times traditionally taking much longer [36][41] Section: Ecosystem and Community - The AI Gallery serves as a one-stop community service platform, fostering an open community for AI development and collaboration [69] - The D-Plan ecosystem partnership program aims to build a collaborative AI ecosystem, providing partners with comprehensive support for training, technology, marketing, and sales [70]