大模型训练及推理算力需求持续增长 国内产业链加速成熟

Core Insights - The capabilities of large AI models are continuously upgrading, with an increase in application scenarios, making computing power, algorithms, and data critical [1] - There is a rapid growth in market demand, and the domestic industry chain is maturing quickly [1] Computing Power - Large-scale computing power is required for training smart glasses, with efficient internal connections and operations needed to overcome computing power bottlenecks [1] - Many regions across the country have been investing in the construction of large data centers to provide sufficient computing power for AI model training [1] Current Statistics - As of March 31, 2025, the number of operational computing power standard racks in China is expected to reach 10.43 million, with intelligent computing power reaching 748 exaFLOPS, providing a smart foundation for massive data calculations [1] - The volume of data being processed is significant, with a figure of 603,138 being mentioned [1] Industry Insights - The training phase of large models is likened to "learning knowledge," while the inference phase is described as "performing tasks," with the industry optimizing algorithms to enhance computing power efficiency [1]