In the AI infrastructure track, liquid cooling has taken the lead.
Entering 2025, tech giants have been extremely generous in AI infrastructure construction, often investing in the hundreds of billions of dollars.
On August 15th, Eastern Time, OpenAI CEO Altman stated that the company will invest trillions of dollars in the future to strengthen AI infrastructure to support various artificial intelligence services.
Last month, Meta CEO Zuckerberg also said that Meta will spend hundreds of billions of dollars to build several large - scale AI data centers.
Statistics show that since the beginning of 2025, countries and regions such as China, the United States, and Europe have announced investments of over tens of billions of dollars in the field of AI infrastructure.
In the AI data center industry chain, the cooling system is a key part of the upstream facility layer. According to SemiAnalysis, the main expenses for a new data center fall into four categories: land and building structure, cooling system, electrical system, and interior decoration.
Among them, the cooling system is the second - largest expense. The cold source production end accounts for about 15% of the AI data center construction investment (excluding IT equipment), the pump and pipeline system accounts for 7%, and the terminal equipment accounts for 10%.
However, in the face of such a promising market, people have found that the previously mainstream air - cooling technology can no longer meet the increasing heat dissipation needs of AI data centers.
Therefore, the next - generation cooling technology - liquid cooling - has started its "rapid development".
01
Liquid cooling has become a "must - have" for servers
Driven by the demand for AI, the power density of data center server cabinets has increased rapidly, expanding from 10kW per cabinet to over 120kW. Traditional air - cooling technology cannot cope with this trend due to its physical limitations.
The heat transfer capacity of air - cooling is limited. The upper limit of the power density of cabinets that can be effectively cooled is about 40 - 50kW. Once this range is exceeded, huge fans and air ducts are required, causing the fan power to increase exponentially with the air flow. The cabinet size becomes bulky, and the risk of component failure also rises.
Even enhanced air - cooling solutions such as the Rear Door Heat Exchanger (RDHx) introduce complexity, increase capital expenditure, and come with high fan energy consumption costs. At the same time, the heat generation of chips themselves is also surging. For example, the thermal design power of the NVIDIA H100 chip has reached 700W, and the power consumption of its subsequent products will exceed 1500W, far beyond the processing capacity of air - cooling.
Economic and environmental factors also force the industry to seek alternative solutions. As the power demand of US data centers is expected to triple from 25 gigawatts to 80 gigawatts in five years, the huge energy consumption pressure forces the industry to fundamentally reshape its architecture.
Therefore, liquid - cooling technology has become an inevitable choice due to its excellent heat dissipation efficiency. First of all, in terms of heat dissipation capacity, the efficiency of liquid cooling far exceeds that of air - cooling. Since the thermal conductivity of water is more than 23 times that of air, and the heat that can be absorbed per unit volume is about 3243 times that of air, it can achieve higher heat transfer and heat flux density.
Liquid cooling supports higher deployment density. Since there is no need to reserve a large amount of space for fans, computing units can be deployed more compactly, reducing the floor area of the data center. This advantage becomes particularly obvious when the single - rack density exceeds 20kW.
From the perspectives of energy saving and cost, liquid cooling directly absorbs the heat of core hardware, reducing the energy consumption of equipment such as fans and air conditioners, and can effectively reduce the PUE value and total cost of ownership (TCO) of the data center.
Therefore, an article in "EE Times" stated: "Telecom operators, industrial enterprises, large - scale enterprises, sovereign institutions, and research organizations have all reached the same conclusion, that is, liquid cooling is not an optimization option but a prerequisite."
02
Three technical paths for liquid cooling
According to whether the internal components of the server are in direct contact with the coolant, liquid cooling can be divided into indirect - contact and direct - contact types.
In the indirect - contact cold - plate liquid cooling, a cold plate filled with coolant is attached to the heat - generating components for heat exchange. Its heat dissipation effect is significantly improved compared to air - cooling, but there is a problem of uneven temperature distribution, and it usually needs to be used in combination with air - cooling.
Direct - contact liquid cooling mainly includes spray - type liquid cooling and immersion - type liquid cooling. In spray - type liquid cooling, the coolant is directly sprayed onto the components, and the heat dissipation is relatively uniform. The effect is better than air - cooling but weaker than immersion - type liquid cooling.
In immersion - type liquid cooling, IT equipment is completely immersed in a non - conductive coolant, and heat is taken away through liquid flow or phase change. It is currently the best heat - dissipation method. However, immersion - type technology is the most difficult, facing many challenges such as coolant selection, long - term equipment reliability, and post - operation and maintenance.
Currently, cold - plate liquid cooling is the most widely delivered and mature liquid - cooling method in China. It requires the least change when transitioning from an air - cooling solution, and is cheaper than other liquid - cooling solutions, making it the first choice for current data center liquid cooling.
03
Big companies are scrambling for suppliers
Currently, the liquid - cooling market is in a stage of rapid expansion.
According to the latest liquid - cooling industry research by TrendForce, with the large - scale shipment of NVIDIA's GB200 NVL72 cabinet - type servers in 2025, cloud service providers are accelerating the upgrade of AI data center architectures, promoting the transition of liquid - cooling technology from early pilot projects to large - scale implementation. It is estimated that its penetration rate in AI data centers will increase significantly from 14% in 2024 to 33% in 2025 and will continue to grow in the next few years.
This year, both SemiAnalysis and Dell'Oro Group predict that driven by the demand for AI, edge computing, and sustainable development, the industry will rapidly transform to the chassis - level and chip - level direct liquid - cooling (direct - to - chip) architecture. It is expected that liquid - cooling expenditure will grow at a compound annual growth rate (CAGR) of 14% and will exceed $61 billion by 2029, even though chip - level and full - immersion liquid - cooling systems face limitations in scalability and maintenance.
When building AI data centers, tech giants have already reserved space for liquid - cooling systems. At the same time, relevant standards are also being established.
According to TrendForce, currently, the four major North American cloud service providers are continuously increasing their investment in AI infrastructure, starting a new wave of data center expansion in local areas, Europe, and Asia, and simultaneously building facilities compatible with liquid - cooling architectures. For example, Google and Amazon Web Services have launched modular buildings with liquid - cooling wiring capabilities in the Netherlands, Germany, Ireland, etc. Microsoft has carried out liquid - cooling pilot deployments in the mid - western United States and multiple locations in Asia and plans to use liquid - cooling systems as the standard architecture starting from 2025.
On August 20th, Intel announced the establishment of the Universal Quick - Connect Interchange Alliance, marking a key step in the interoperability of the liquid - cooling ecosystem. The first - batch certified partners of the alliance include five enterprises: Envicool, Danfoss, Liminda Technology, LanKe Electric, and Zhengbei Connection. The establishment of this alliance aims to build a unified standard for the liquid - cooling industry. By formulating specifications for universal quick - connectors, the products of various manufacturers will achieve the function of inter - insertion and interchange. This measure is expected to solve the compatibility problem of liquid - cooling systems and promote the large - scale development of the entire industry.
And the suppliers in the liquid - cooling supply chain generally report that their current production capacity can no longer meet the market demand.
Both Shuanghong Technology and Chi Sheng Technology, cold - plate manufacturers, have reported strong demand for liquid - cooling solutions, and both companies have started to increase their manufacturing capabilities. Chi Sheng Technology said that even at full production capacity, its current order volume still exceeds supply, which has prompted it to continuously expand its factory in Vietnam.
Shuanghong Technology is also confident about the liquid - cooling prospects in the second half of the year. The company is currently vigorously expanding its production capacity in Thailand and plans to enter North America by setting up a production base in Mexico.
Delta Electronics, a liquid - cooling supplier, also mentioned that its liquid - to - air (L2A) cooling solution for data centers has received strong demand. This solution currently occupies a high market share, and the supply is still tight. It is expected that new customers will join in the second half of the year. Benefiting from the increase in the shipment volume of liquid - cooling solutions, the profit of Delta Electronics' Infrastructure Business Unit increased by 4.17 times year - on - year in the second quarter.
04
China's liquid - cooling industry has a bright future
China is also actively participating in this industry transformation from air - cooling to liquid - cooling.
According to Leqing Think Tank, the current "East - West Computing" project requires that the PUE of newly - built data centers should be ≤1.25, which is forcing the popularization of liquid cooling. As of June 2025, 150 data centers have been built in the first batch of eight national computing power hub nodes, among which the proportion of liquid - cooled data centers exceeds 30%.
According to the latest calculation by the China Academy of Information and Communications Technology, the market scale of liquid cooling in China's intelligent computing centers was 18.4 billion yuan in 2024 and will reach about 130 billion yuan by 2029, that is, it will increase by more than 7 times in 5 years.
Driven by both the market and policies, many enterprises in the fields of temperature control and thermal management have started to layout the liquid - cooling track and achieved remarkable results.
Envicool is a provider of precision temperature - control and energy - saving solutions and products. Its main business includes temperature - control and energy - saving products for computer rooms, temperature - control and energy - saving products for cabinets, bus air conditioners, rail transit train air conditioners, and services.
At the same time, Envicool is deeply involved in the liquid - cooling track. One of its core advantages is its membership in the NVIDIA industry chain. Its UQD products have been included in the NVIDIA MGX ecosystem partners.
Envicool has launched the Coolinside full - chain liquid - cooling solution, covering the delivery scenarios from factories to data centers, and has provided products for major large - scale data centers in China such as ByteDance, Tencent, Alibaba, and China Mobile.
Wind statistics show that within 19 trading days starting from July 24th, Envicool's cumulative increase reached 113.28%; if calculated from the lowest point of the year on April 9th, it has risen by more than 190%.
GLAN was established in 2001 and has been committed to the R & D, design, production, and sales of thermal management equipment and control systems since its establishment. The company's products are mainly divided into high - power power electronics thermal management products, high - power density device thermal management products, and engineering operation and maintenance services.
GLAN is one of the few domestic enterprises that simultaneously master the core technologies of cold - plate and immersion - type liquid cooling and can provide data center liquid - cooling solutions using two mainstream technologies to meet the heat - dissipation needs from single cabinets to super - large - scale data centers.
GLAN has currently established long - term and stable cooperative relationships with well - known domestic and foreign customers such as ByteDance, Alibaba, Tencent, GDS, Inspur, ZTE, H3C, CenturyLink, GE, Siemens, and ABB in the data center business. Among them, in 2024, it successfully completed the delivery of the subcontracting project of the cold - plate secondary side of the first domestic super - computing center and the delivery of cold - plate liquid - cooling equipment and related supporting equipment for a data center project in Malaysia. In 2025, it delivered immersion - type liquid - cooling products to ByteDance.
Shinry is a leader in the field of special - purpose air conditioners. Its main business includes air conditioners for data services, industrial air conditioners, special air conditioners, and public and commercial air conditioners.
The company has been ahead of the curve in layout of liquid - cooling technology since 2011 and is an expert in liquid - cooling system solutions. As of the first half of 2024, it has 67 patents in the field of liquid cooling. Its technical solutions cover cold - plate and immersion - type technologies and can provide end - to - end overall solutions from dry coolers, CDUs to pre - fabricated pipelines. Its core product, the "Air - Liquid Homologous Large - Scale Liquid - Cooling System", can support high - density heat - dissipation needs of over 200kW per single cabinet and can reduce the PUE value of data centers to below 1.1.
Shinry has established in - depth cooperative relationships with many core customers. Huawei is one of its most important strategic partners. The large - scale deployment of Huawei's AI Ascend chips has directly driven the demand for Shinry's liquid - cooling products. In addition, the company's customers also include operators such as China Mobile, China Telecom, and China Unicom, as well as major domestic Internet and technology giants such as Tencent, Alibaba, ByteDance, Baidu, and Kuaishou.
This article is from the WeChat official account "Semiconductor Industry Insights" (ID: ICViews), author: Junxi, published by 36Kr with authorization.