AI chipsets-level cooling schemes generally use the combination of heat sinks and fans, using this forced convection mode. But AI chipsets liquid cooling into mainstream cooling scheme. The power of the cabinet exceeds 15kW, and the air cooling reaches the ceiling.The PUE supervision of IDC center becomes stricter, requiring less than 1.25/1.2.Liquid cold plate has been used more, immersion will become the mainstream scheme.
At present, AI server cooling investment increases significantly, its advantages are very obvious, from the technical characteristics, should be the future development trend and direction of research and development.Heat dissipation input of single KW power consumption increased to 4 times:The cost of air cooling is about 3000 yuan /kw, the cost of cold plate is about 4000 yuan /kw, and the cost of immersion liquid cooling is about 12000 yuan /kw.
Server cost ratio increased to 2.5-3 times:The power consumption of a single ordinary server is about 400w, the price is about 60,000 yuan, and the value of air cooling is about 1200 yuan, accounting for about 2%.The Inspur NF5468M6 AI server is used as an example. It consists of two intel lce Lake processors and eight NVIDIA A800 Gpus, consumes about 4kw, and sells for about 880,000 RMB. Liquid cooling is supported. The value of immersion liquid cooling is 48,000 yuan, accounting for about 5.5%.
AI chipsets liquid cooling industry into a period of mass:In 2025, the global market size of AI server chip-level liquid cooling in China is 22.3 to 33.3 billion yuan and 7.2-10.8 billion yuan.In 2027, the global music system will reach 702 to 94.1 billion yuan, and the domestic market will reach 20.3 to 27.2 billion yuan.
Four cognitive differences of temperature-controlled liquid cooling:
Poor perception 1: No clear policy support?
Recently, the Ministry of Industry and Information Technology said that by the end of 2025, the total energy consumption per unit of telecom business should be reduced by 15%, and the PUE value of newly built large or ultra-large data centers should be lower than 1.3. In order to meet the high power consumption requirements, liquid cooling will become a necessary option for new data centers.
Cognitive poor two: Industrial demand is not just demand?
The increasing demand for AI computing power greatly increases server energy consumption, and the future AI cluster computing power density is expected to reach 40+kw/ cabinet. When the cabinet power is greater than 15kw, the cost performance ratio of the liquid cooling solution becomes prominent. When the cabinet power is greater than 20kw, the liquid cooling solution is the only solution. It is expected that the penetration rate of the submerged liquid cooled data center will reach more than 40% in 2025.
Cognitive difference three: industry threshold is low, competition pattern is poor?
When liquid cooling becomes the mainstream solution, its security will form a strong technical barrier. As the core supporting facilities of the underlying computing power, the safety and reliability of the liquid cooling solution not only determines the amount of energy consumption, but also determines the security and stability of the related computing power cluster. The first-mover advantage of industry leading manufacturers will form a great barrier to subsequent growth, and the trial and error cost of downstream manufacturers is extremely high.
Poor perception four: Is the performance of the relevant company inflexible?
It is estimated that in 2025, the global market space of AI server chip-level liquid cooling will reach 30 billion yuan, and the market space of general server chip-level liquid cooling will reach 80 billion yuan, representing a geometric growth compared with the market size in 2022. With the rapid expansion of AI servers, solutions will be upgraded from “air-cooled – water-cooled – liquid-cooled plate – immersed”, and the value of products will gradually climb. It is estimated that the cost of AI servers for submerged solutions will account for nearly 6%.
