The Hotter Computing Power Gets, the Colder Liquid Cooling Becomes: A Hundred-Billion-Level Heat Dissipation Revolution Is Reshaping Data Centers

In the third week of September 2026, when the spotlight of HUAWEI CONNECT fell on the Ascend 960 supernode, one long-overlooked detail was worth pondering: this flagship product, featuring NPO optical engines and scalable to 4,096 cards, simultaneously adopted a full liquid-cooling design. On the same day, Baidu AI Cloud announced that its Tianchi supernode had achieved a full-stack domestic solution, with specifications for six major modules fully opened up. Among them, the heat dissipation module was listed as a core open area alongside power supply, nodes, and interconnection.

This is no coincidence. As single-cabinet power rises from tens of kilowatts to hundreds of kilowatts or even megawatt levels, traditional air cooling has reached its physical limits. Liquid cooling, once regarded as an “optional” heat dissipation solution, is becoming a “must-have” and a “hard threshold” for computing power infrastructure. From chip vendors to server manufacturers, from data center operators to coolant suppliers, a hundred-billion-level industrial chain centered on “cooling down” is accelerating into shape.

Air Cooling Hits the Wall: When Heat Dissipation Becomes the Bottleneck for Computing Power Expansion

In the past, data centers mainly relied on fans and air conditioning to “blow cool.” But after the accelerated deployment of high-power computing clusters such as 10,000-GPU intelligent computing centers and supernodes, chip power consumption has continued to climb, and traditional air cooling has become increasingly strained. A palm-sized cold plate now has to handle 2,000 watts or even 3,000 watts of heat dissipation. It fits tightly against the surface of high-power chips and continuously carries heat away through liquid circulation. A few years ago, the same cold plate only needed to handle a few hundred watts of heat.

The energy account is clear: traditional air cooling requires an additional 0.3 to 0.5 kWh of electricity for heat dissipation to complete the computing task of 1 kWh; the additional heat dissipation power consumption of full immersion liquid cooling can be reduced to about 0.15 kWh. This means that power consumption in the heat dissipation link can be cut by more than half. Under the “dual carbon” goals and increasingly tightening data center PUE requirements, this is not a multiple-choice question of “whether to do it,” but a survival question of “must do it.” Some industry experts have calculated a more intuitive account: for a data center with 100,000 servers, if liquid cooling is fully used to replace air cooling, the annual electricity cost savings can reach the level of hundreds of millions of yuan.

Policy is also accelerating the push. The Implementation Plan for Promoting Coordinated Digital and Green Transformation (2026–2030), jointly issued by the National Development and Reform Commission and six other departments, explicitly proposes strengthening the green design of computing power facilities, supporting the innovation, promotion, and application of liquid cooling, waste heat recovery, and low-power chips, exploring an 800V high-voltage direct current power supply architecture for computing power facilities, and promoting the deployment of single cabinets over 100 kW. At the same time, many places have already set a PUE below 1.25 as the entry threshold for new intelligent computing centers, and some eastern provinces even require it to be below 1.2. When dealing with high-density cabinets, air cooling solutions generally find it difficult to stably press PUE below 1.3, while cold plate liquid cooling can achieve 1.15 to 1.25, and immersion liquid cooling is expected to approach 1.05 further.

International giants are also accelerating their shift. Nvidia’s new-generation AI computing platform Vera Rubin has been fully put into production, completely abandoning traditional air cooling and adopting a fanless full liquid-cooling solution. Its GB200 NVL72 cabinet power has already reached more than 120 kW, and the next-generation Rubin platform’s single-cabinet power is expected to exceed 200 kW. Air cooling has completely failed at this power density. It can be said that liquid cooling is not a “future trend,” but a “present reality.”

Dual Drivers of Incremental and Existing Demand: An Underestimated Market

The explosive logic of the liquid cooling market is far more than simply building new intelligent computing centers. Data from CCID Consulting shows that more than 80% of existing air-cooled computer rooms in China have the potential for liquid cooling retrofitting. On one hand, new intelligent computing centers are increasing rapidly; on the other hand, demand for energy-saving retrofitting of old computer rooms is being released. Liquid cooling is simultaneously capturing both the “incremental” and “existing” markets. Although retrofitting existing facilities is more difficult and takes longer, the scale is huge, and the energy-saving benefits after retrofitting are directly reflected in electricity bills, creating strong incentives for both operators and equipment vendors.

Data confirm this judgment. In 2025, China’s liquid cooling data center market reached 15.98 billion yuan, a year-on-year increase of 45.2%. It is expected to maintain a growth rate of more than 35% over the next three years, and the overall market size will reach 47.04 billion yuan by 2028. From a global perspective, the global intelligent computing liquid cooling market may exceed 100 billion yuan in 2026. In terms of penetration, the share of liquid cooling in AI servers is rising rapidly. TrendForce estimates that the penetration rate of liquid cooling in AI chips will rise from 33% in 2025 to 53% in 2026, and is expected to approach 60% in 2027.

The capital market’s reaction has also been intense. On September 7, liquid cooling concept leader Zhongshi Technology hit the “20cm” daily limit during trading, with its stock price breaking through 100 yuan per share for the first time and hitting a historical high. Zhongyuantong also hit the “20cm” daily limit; Dunan Environment, Haoersai, and Jitai Shares hit the daily limit; Gaolan Shares rose more than 8%. More noteworthy is the density of institutional research: since July, 19 liquid cooling concept stocks have received institutional research, among which Feilong Shares and Dayuan Pump both received research from more than 100 institutions. The focus of institutional attention has shifted from “whether liquid cooling will become mainstream” to “which company has true mass-production capability and customer barriers in cold plates, CDUs, quick connectors, coolants, and other links.”

The intensification efforts of industry leaders are even more direct. On the evening of September 11, Inspur Information announced a plan to raise no more than 9 billion yuan, of which 1.4 billion yuan will be specifically invested in the “R&D and Industrialization Project of Liquid-Cooling-Native Computing Power Infrastructure for Green Intelligent Computing Clusters,” focusing on three major directions: ultra-high-density liquid-cooling-native whole-cabinet servers, prefabricated liquid cooling data center rapid delivery systems, and liquid-cooling-native computing architecture. Inspur Information has been laying out liquid cooling since 2015, has accumulated more than 1,000 core patents, and has led or participated in the formulation of more than 20 liquid cooling-related design and technical standards. At the same time, vendors such as Sugon, Lenovo, H3C, and xFusion are also accelerating the launch of liquid cooling whole-cabinet solutions. Liquid cooling is moving from “optional” to “standard.”

The Battle of Technology Routes: The Boundaries of Cold Plate, Immersion, and Phase Change

Liquid cooling is not a single technology, but a spectrum that includes cold plate, immersion, and phase-change liquid cooling, among other technical routes. Different routes have their own applicable boundaries and are forming a clear division of labor.

At present, cold plate liquid cooling has the highest maturity and is also the most widely used solution. Its principle is not complicated: a metal cold plate is installed above the chip that generates the most heat, and coolant circulates in the flow channels inside the cold plate to carry heat away. Shenling Environment clearly stated in an investor relations activity that liquid cooling has gradually become the standard configuration for heat dissipation in high-power computing servers, and its penetration rate will inevitably continue to rise in the future. But cold plate cooling has also encountered new challenges. Industry insiders point out that after single-unit power exceeds 100 kW, the pressure on cold plate liquid cooling will increase significantly, requiring higher-flow coolant, more complex pipeline design, and more powerful CDUs (coolant distribution units). As single-cabinet power moves toward 200 kW or even higher, the balance between heat dissipation efficiency and pump power consumption in cold plate solutions will become increasingly difficult.

Greater market imagination comes from immersion liquid cooling. In China’s first immersion liquid cooling computing cabin in Guiyang, Guizhou, servers are directly immersed in insulating coolant. At present, the project is testing four types of coolant: silicone oil, fluorinated liquid, synthetic oil, and transformer oil. The technology is already running; the next step is to compete on scale and cost. The advantages of immersion liquid cooling lie in extremely high heat dissipation efficiency, extremely low noise, no need for fans, and the ability to support higher power density. But the challenges are equally obvious: high coolant cost, inconvenient server maintenance, the need for customized cabinets and motherboards, and significant differences in material compatibility among different coolants. When AI server cabinet power reaches more than 200 kW, immersion solutions will be more advantageous, but in low-to-medium power density scenarios, cold plate cooling is still the more economical choice.

Higher-power technology has already begun to be deployed. In Zhengzhou, Henan, domestic immersion phase-change liquid cooling heat exchange technology has been officially put into use, with single-cabinet power reaching megawatt level, 3 to 5 times that of traditional liquid cooling solutions. After being heated, the refrigerant undergoes a phase change, changing from liquid to gas to carry away chip heat, and then condenses back into liquid for cyclic use. Compared with traditional air cooling, it saves more than 30% of electricity and simultaneously helps improve chip performance by more than 10%. The uniqueness of phase-change liquid cooling lies in its use of latent heat of phase change, with heat exchange efficiency far higher than single-phase liquid cooling, but system complexity and sealing requirements are also higher.

There is also a more niche but imaginative route—spray liquid cooling. At the China Computing Power Conference, spray liquid cooling made a new appearance. Unlike cold plate and immersion liquid cooling, it simulates a “shower” through customized spraying, directly spraying coolant onto the surface of heat-generating components to precisely carry away heat. This solution has unique advantages in specific high-density scenarios, but places extremely high requirements on coolant insulation and spray uniformity.

Nvidia’s technology choices are accelerating the convergence of industry routes. Its new-generation AI computing platform Vera Rubin has been fully put into production, completely abandoning traditional air cooling and fully adopting a fanless full liquid-cooling solution. A CITIC Securities research report estimates that the value of liquid cooling components per cabinet is expected to rise from $41,500 for GB200 to $55,700 for Rubin NVL144, an increase of another 17% over GB300. This means that as power density increases, the per-cabinet value of liquid cooling components continues to grow, and the market space is far from reaching its ceiling.

From Components to Systems: Liquid Cooling Is Redefining Data Centers

The deeper change in the liquid cooling industry is that it is moving from “selling cold plates” to “selling the entire intelligent computing center.” What companies compete over is no longer just “whose cold plate is better.” Liquid cooling is beginning to be designed together with cabinets, power, and data center construction, turning from a device or a set of components into part of the overall intelligent computing center design. Through prefabricated production and supply chain collaboration, some companies have even proposed completing the delivery of an entire AIDC artificial intelligence data center within eight months from the customer’s order. Under the traditional construction model, a data center of equivalent scale often takes a year and a half or even longer from design to commissioning.

At the China Computing Power Conference, a containerized distributed computing power unit achieved factory prefabricated production, with a prefabrication rate of 90%. After being transported to the site and powered on, it can be put into operation in as little as two hours, greatly compressing the traditional data center construction cycle of about one year. This kind of “computing power container” integrates and tests liquid cooling, power supply, network, computing, and other modules in the factory. On site, only external water and electricity access and network connection need to be completed, greatly improving delivery efficiency.

Huawei, based on the Lingqu architecture, launched a new-generation Ascend 960 liquid-cooled intelligent computing supernode, providing computing power support for trillion-level large models. Compared with the previous generation, network bandwidth is increased 4 times, RTT latency is reduced by one third, and memory capacity is increased 11 times. This supernode adopts a full liquid-cooling design, deeply coupling liquid cooling with high-speed interconnection and unified memory addressing, reflecting the design concept of “liquid-cooling native”—not adding liquid cooling after the server is designed, but treating liquid cooling as a basic constraint from the very beginning of server architecture design.

A CICC research report points out that the parameter scale of large models, long context, and parallel strategies have pushed up inter-card communication demand. The bottleneck of AI infrastructure is shifting from single-chip computing power to interconnection efficiency, resource scheduling, and system collaboration. Supernodes are expected to become an important organizational form of AI computing power, driving demand expansion and technological upgrading in links such as switch chips and switch trays, optical interconnection, liquid cooling, and high-power power supplies. This means that liquid cooling is no longer an accessory to the data center, but a core subsystem on par with computing and networking. The design of liquid cooling pipelines, the flow distribution of coolant, and the redundant configuration of CDUs all directly affect the reliability and performance of the entire supernode.

Orders Booked Through the End of the Year: The Capacity Race Has Begun

The expansion of demand quickly transmitted to the production side. On a high-end AI server production line in Jinan, Shandong, a batch of liquid-cooled servers is being urgently assembled and tested; at a CDU assembly line for core liquid cooling equipment components in Laixi, Qingdao, Shandong, orders have been booked through the end of December; a liquid cooling component manufacturer in Foshan, Guangdong, has seen its orders, output value, and delivery volume all grow by nearly 100% year-on-year this year, with 70% of its production lines at full capacity. Some company executives said that in the past, customers usually required delivery within three months after placing an order, but now some urgent orders even require delivery within one month, putting enormous pressure on capacity.

This tight capacity situation reflects a deeper industrial reality: the speed at which the liquid cooling industry chain matures is becoming a key variable constraining the pace of computing power infrastructure construction. When leading vendors such as Huawei, Baidu, and Inspur are all racing to deploy liquid-cooled supernodes, the supply capacity, delivery cycle, and cost control capability of liquid cooling components will directly determine who can seize the initiative in this round of computing power competition. Key links such as cold plates, CDUs, quick connectors, coolants, and liquid cooling pumps are all undergoing a leap from “small-batch customization” to “large-scale mass production.” Whoever can complete capacity ramp-up and yield improvement first will be able to take the initiative in the next round of order competition.

At the same time, the localization rate of the liquid cooling industry chain is also rising rapidly. In the past, high-end coolants, quick connectors, and other components relied heavily on imports. Now domestic companies have made breakthroughs in coolants such as fluorinated liquids, silicone oils, and synthetic oils, and domestic substitution of quick connectors and CDUs is also accelerating. This not only reduces costs but also improves supply chain security.

Cold Thinking: Liquid Cooling Is Not a Panacea

While the liquid cooling industry is advancing triumphantly, it is also necessary to calmly view the challenges involved. First, the initial investment in liquid cooling systems is still higher than air cooling. The construction cost of a cold plate liquid cooling data center is usually 10% to 20% higher than air cooling, and immersion cooling is even higher. Although energy-saving benefits can recover the incremental cost within a few years, initial investment pressure still exists for budget-sensitive projects.

Second, the degree of standardization is still insufficient. Links such as cold plate dimensions, quick connector specifications, coolant types, and CDU interfaces lack unified standards, making it difficult for solutions from different vendors to interoperate and complicating spare parts management for operations and maintenance. Some industry insiders have called for the industry to accelerate the formulation of standards for liquid cooling interfaces, coolants, and O&M specifications to reduce users’ selection costs and lock-in risks.

Third, the O&M system needs to be restructured. O&M experience from the air cooling era partially fails in liquid cooling scenarios. Leak detection, coolant replacement, pipeline cleaning, and pump and valve maintenance all require new skills and tools. O&M personnel in liquid cooling data centers need specialized training, and O&M processes also need to be redesigned. Although leakage risk can be reduced through multiple protective measures, it remains one of the reliability issues users care about most.

Fourth, the value of waste heat recovery has not yet been fully released. The heat discharged by liquid cooling systems has a relatively high temperature and has the potential for waste heat recovery and utilization, which can be used for district heating, greenhouse cultivation, industrial preheating, and other scenarios. However, most liquid cooling data centers have not yet been equipped with waste heat recovery systems, and heat is directly discharged into the atmosphere, causing energy waste. As the “dual carbon” policy advances further, waste heat recovery is expected to become standard in liquid cooling data centers.

From a broader perspective, the rise of the liquid cooling industry is a microcosm of the computing power infrastructure shifting from “scale-driven” to “efficiency-driven.” When single-chip power consumption breaks through physical limits, and when PUE changes from a “bonus point” to an “entry line,” heat dissipation is no longer a supporting issue that can be ignored, but a fundamental constraint determining whether computing power can continue to expand. Behind the hundred-billion-level liquid cooling market is a redefinition of the physical form of data centers—from “blowing air” to “taking a bath,” from “component procurement” to “system delivery,” and from “incremental market” to “existing retrofitting.” This “cold” revolution is providing an indispensable physical foundation for the increasingly “hot” computing power industry. And in this revolution, whoever can first solve the problems of cost, standardization, and O&M will be able to occupy an advantageous position in the computing power infrastructure competition of the next decade.