中金 | AI寻机系列04：AIDC冷水机组升级，磁悬浮压缩机逐步国产替代

Group 1 - The rapid growth of the cooling business in AI data centers is driven by the increasing demand for cooling solutions due to the surge in global computing power investments, with major cloud providers' capital expenditures reaching nearly $100 billion in Q3 2025, a year-on-year increase of over 70% [2][5] - The cooling system's core component is the chiller, which accounts for approximately 40% of the total energy consumption of data centers, with chillers consuming about 60-85% of the cooling equipment's energy [12][17] - Liquid cooling solutions are emerging as a key technology to meet the stringent PUE (Power Usage Effectiveness) requirements, with the potential to reduce PUE to around 1.2, significantly improving energy efficiency [9][14] Group 2 - Compressors are the core components of chillers, with magnetic levitation centrifugal compressors expected to become mainstream due to their higher cooling efficiency and lower energy consumption compared to screw compressors, potentially saving over 40% in energy [3][18] - The market for magnetic levitation chillers is projected to reach approximately 28.83 billion yuan by 2026, with a year-on-year growth of 70.4% if all new AI data centers adopt this technology [3][18] - Domestic manufacturers are making technological breakthroughs in key components of magnetic levitation compressors, which may lead to a gradual replacement of foreign brands in the market [3][49] Group 3 - The increasing power density of AI servers necessitates higher cooling capacities, with single cabinet power requirements exceeding 50 kW, driving the demand for more efficient cooling solutions [8][39] - The shift from traditional air cooling to liquid cooling is essential for meeting the cooling demands of high-density AI data centers, with centrifugal compressors being more suitable for large-scale cooling needs [39][41] - The adoption of magnetic levitation technology in centrifugal compressors enhances their efficiency and durability, making them well-suited for the high power demands of AI data centers [43][46]