Liquid cooling: from "optional" to "mandatory" — a hundred-billion-yuan track under the AI computing power surge
CICC predicts that 2026 will mark the first year of large-scale liquid cooling deployment, with the global liquid cooling market for intelligent computing centers reaching 114.7 billion yuan, up 273% year on year. The penetration rate of liquid-cooled servers in China is expected to jump from 20% in 2025 to 37% in 2026, with cold-plate liquid cooling accounting for more than 85% of the share; in the future, 80% of solutions on the market will be cold-plate liquid cooling. Standardization of liquid cooling fluids still needs to be unified; PG 25 has already gained preference from most customers, and Envicool focuses on 5-year maintenance-free operation.
Body · AI translation
① A liquid cooling industry expert said that in the future, 80% of the market will be cold plate liquid cooling solutions. Given the high requirements for energy efficiency and heat recovery in Europe and North America, the share of immersion liquid cooling will gradually climb. ② Although the domestic market still has disagreements among deionized water, PG 25 and EG 25 and has yet to form a final consensus, PG 25 has already gained the preference of most customers. Science and Technology Innovation Board Daily, July 28 (Reporter Huang Xinyi) Liquid-cooled data centers are becoming one of the focal points on the AI industry stage. CICC predicts that 2026 will usher in the first year of large-scale liquid cooling deployment, with the global intelligent computing center liquid cooling market reaching 114.7 billion yuan, up 273% year on year, and is still expected to maintain high growth in 2027.
"Liquid cooling has gone from being optional to a necessary choice," said Zhou Shaorong, a technical expert at Intel.
From air cooling to cold plate liquid cooling and then to fully liquid-cooled architectures, the evolution path of data center heat dissipation is becoming increasingly clear. This is an infrastructure reconstruction driven by the surge in computing power density under the wave of artificial intelligence.
▍Cold plate liquid cooling is timely
According to forecasts by the China Commercial Industry Research Institute, the penetration rate of liquid-cooled servers in China has risen from less than 3% in 2021 to 20% in 2025, and is expected to jump to 37% in 2026 and break through the 50% critical point in 2027. The liquid cooling penetration rate of global AI data centers has likewise risen from 14% in 2024 to 33% in 2025, and is expected to reach 40% in 2026.
In terms of technology routes, cold plate liquid cooling holds absolute dominance. CCID Consulting predicts that in the next three years, cold plate data centers will grow from 21.62 billion yuan to 43.75 billion yuan, consistently accounting for more than 85% of the share.
A liquid cooling industry expert said that in the future, 80% of the market will be cold plate liquid cooling solutions. Qiu Yuanyuan, Greater China business director at Castrol, said in an interview: "From the server perspective, most of the servers shipped on the market today are liquid-cooled servers. Liquid cooling has now become a necessary option for high AI computing power. Cold plate liquid cooling occupies the mainstream of market choice and is a timely technology. For a future of higher density, we have no doubt about the promising prospects of immersion liquid cooling."
Although immersion liquid cooling currently has a smaller share, it is growing faster. CCID Consulting predicts that immersion liquid-cooled data centers will increase from 1.63 billion yuan in 2026 to 3.29 billion yuan in 2028, with the scale nearly doubling. Its relatively low share is mainly due to its large initial investment and complex operations and maintenance, and it is currently only suitable for specific scenarios such as supercomputing and high-density AI training.
At the 2026 World Artificial Intelligence Conference, which concluded not long ago, the supernode systems of leading server manufacturers such as Huawei, ZTE, H3C and Inspur have all been upgraded to liquid cooling across the board. The Ascend 950 supernode, designed specifically for trillion-level large model training, emphasizes liquid cooling configuration, and ZTE's OEX supernode achieves full coverage of cold plate liquid cooling.
Zhou Shaorong, a technical expert at Intel, introduced that with the vigorous development of high-computing-power AI servers, especially given the high requirements for energy efficiency and heat recovery in Europe and North America, the share of immersion liquid cooling will gradually climb. However, he also admitted that immersion liquid cooling is difficult to retrofit into existing data centers, and issues such as load-bearing still need to be resolved. Newly built data centers are relatively easier to deploy.
Zhang Xiangbin, a liquid cooling product expert at Envicool, also said that data centers' acceptance of liquid cooling has greatly increased, directly thanks to the explosive growth in AI computing power demand. AI applications impose very high requirements on computing density, forcing the industry to stack as much computing power as possible within limited space, and liquid cooling breaks through the limits of air cooling heat dissipation, becoming an unavoidable approach. Traditional data centers are difficult to retrofit, but new projects have generally embraced liquid cooling solutions.
▍Standards for liquid cooling working fluids urgently need to be unified
The performance and indicators of liquid cooling working fluids are among the most important factors for the normal and reliable operation of the liquid cooling secondary loop system. As the key medium responsible for transferring heat in the entire liquid cooling circulation system, the liquid cooling working fluid can quickly absorb and carry away the large amount of heat generated by various components in the data center, just like the blood of the liquid cooling circulation system.
Although industry standards such as the Technical Requirements for Cold Plate Liquid Cooling Data Centers have been released, there is still a great deal of work to be done on standardization in the field of liquid cooling working fluids.
For example, liquid cooling working fluids such as PG 25, EG 25 and deionized water have not yet been unified. Du Liguang, director of the platform technology R&D department for China at Intel's Data Center Group, said that PG 25 is more recommended because of its convenient maintenance and extremely low toxicity, while EG 25 has a certain degree of toxicity. Both PG 25 and EG 25 are freeze-resistant; pure water freezes at 0°C, and outdoor pipes face the risk of freezing in winter, while PG 25 can effectively avoid this problem.
In addition to lowering the freezing point, liquid cooling working fluids must also be supplemented with various functional additives. These formulations are the core technical secrets of each manufacturer and are not disclosed publicly. Du Liguang introduced that the standard composition of PG 25 is 25% propylene glycol and 75% deionized water; the former must ensure purity and stability, while the latter requires the highest possible cleanliness. Beyond these two basic components, manufacturers add small amounts of special additives to specifically inhibit problems such as scaling, bacterial growth, corrosion and erosion. Therefore, the basic formulation standard of PG 25 is clear, but the differentiated competitiveness of each manufacturer is reflected precisely in those small amounts of undisclosed additive formulations.
Zhang Xiangbin, a liquid cooling product expert at Envicool, said that at present, the choice of liquid cooling working fluids at home and abroad is generally trending toward PG 25. Although the domestic market still has disagreements among deionized water, PG 25 and EG 25, and no final consensus has yet been formed, PG 25 has already gained a majority usage preference. Although deionized water has advantages in heat dissipation performance, its application scenarios are limited: first, the system scale should not be too large, otherwise maintenance complexity rises significantly; second, the system must be as closed as possible and have short links. By contrast, PG 25 is more suitable for deployment in large data centers with long liquid cooling links, while the advantages of deionized water are more reflected in short-link, small-scale systems.
Zhang Xiangbin further stated that liquid cooling working fluids usually contain various functional additives such as antioxidants, corrosion inhibitors and antibacterial agents, and each manufacturer's formulation is different, which directly determines the system's maintenance cycle. Envicool focuses on 5-year maintenance-free operation, matching the typical 5-year life cycle of servers; while the maintenance cycles of other manufacturers' products may range from 1 to 3 years, and the root of the difference lies precisely in each manufacturer's complex additive formulations.
There is also no unified industry standard for the replacement cycle of liquid cooling working fluids. Du Liguang admitted frankly: "Cold plate liquid cooling is still in the early stage of development in China, and it is too early to discuss the issue of working fluid service life; it may take some time to form a clear answer. As the technology is promoted on a large scale and applications become increasingly widespread, the industry will naturally conduct in-depth discussions around the standards that liquid manufacturers should meet, the requirements of server and pipeline manufacturers, and reasonable addition or replacement cycles, and gradually reach a consensus. At present, the liquid cooling industry has not yet developed to that stage."
▍Scale effects reduce liquid cooling costs
In addition, the operational complexity and construction cost of liquid cooling remain unavoidable practical challenges.
Zhang Xiangbin analyzed that the economics of liquid cooling depend on scale effects; only when a sufficient deployment scale is reached can costs be effectively shared, otherwise customers still tend to adopt air cooling. At present, newly built data centers generally adopt hybrid air-liquid architectures, but the proportion of air cooling continues to decline, and in the coming years it is expected to transition to a fully liquid-cooled model. However, at the level of the entire data center, liquid cooling is not the only solution, because natural cooling sources can still play a role, and air-liquid same-source solutions are also becoming increasingly mature; but from the IT equipment side, the continued rise in liquid cooling penetration is already a clear trend.
The market already has clear expectations for this. A report by CCID Consulting shows that in the next three years, the scale of China's liquid cooling data center market will grow from 23.25 billion yuan in 2026 to 47.04 billion yuan in 2028, achieving more than double growth.
It can be said that from air cooling to liquid cooling, this cooling revolution concerning the future of data centers has already begun. And whoever can seize the initiative in the standardization process and build barriers in large-scale delivery is expected to win the final position in this track.