How much incremental liquid cooling market has the AI big model brought?

The scale of large and data center racks in China has grown rapidly. In recent years, the scale of data center racks in China has steadily increased. According to the standard rack size of 2.5kW, as of the end of 2021, the scale of data center racks in use in China has reached 5.2 million, with a CAGR of over 30% in the past five years. Among them, the scale of large and above racks is 4.2 million, accounting for 80%. China's data centers are mainly based on general computing power. 

With the increasing demand for AI and edge computing, the number of supercomputing and intelligent computing data centers needs to grow. Currently, the scale of universal computing power racks accounts for over 90%, and supercomputing centers are mainly used in national scientific research, with few commercial scenarios; The intelligent computing center is gradually transitioning from early experimental exploration to commercial pilot projects. With the enrichment of artificial intelligence application scenarios, the expected growth rate of scale will reach 70%.

According to calculations by Inspur and IDC, the scale of China's intelligent computing power will reach 15.52 billion floating-point calculations per second (EFLOPS) in 2021, and 268.0 EFLOPS in 2022. It is expected that by 2026, the scale of intelligent computing power will enter the level of one billion floating-point calculations per second (ZFLOPS), reaching 1271.4 EFLOPS. In 2021, the scale of general computing power in China will reach 47.7 EFLOPS, and it is expected that by 2026, the scale of general computing power will reach 111.3 EFLOPS. During the period from 2021 to 2026, it is expected that the compound annual growth rate of China's intelligent computing power scale will be 52.3%, while the compound annual growth rate of general computing power scale during the same period will be 18.5%.

At present, the cooling technology of the liquid cooling plates or heatsink plates has a high maturity and is in the mainstream in the liquid cooling Technology roadmap, assuming that the current proportion is 80%. In the future, with the maturity of immersion liquid cooling technology, the overall proportion is expected to gradually increase. Based on comprehensive calculations, AI large model training and inference will bring a liquid cooling market space of 4 billion yuan. With the increase of model parameters and promotion of use, the liquid cooling heatsinks market will experience a compound annual growth rate of 60% over the next four years.
According to our calculations, it is expected that the scale of server cooling modules will maintain a compound growth of nearly 30% by 2026. Based on the following assumptions: 1) IDC predicts a global server shipment volume of 18.77 million units in 2026; 2) The proportion of AI servers is increasing year by year; 3) As the power increases, the water cooling penetration rate increases, and the number of water cooling servers increases by 25% annually; 4) As the power increases and the water cooling heatsinks requirements increase, the prices of air-cooled and water-cooled ASP have been increasing year by year, reaching 15% and 10% respectively.