Author: Peng Cheng
Recently, at the NAVIGATE 2026 Leadership Summit, H3C (New H3C Technologies) released the high-density fully liquid-cooled S90000 server, which supports up to 576 cores. H3C proposed that through innovative high-density power supply and fully liquid-cooled thermal management paradigms, the data center PUE (Power Usage Effectiveness) can be reduced to below 1.05.
Coincidentally, at the GTC conference in March 2026, NVIDIA also clarified that the Vera Rubin NVL72 is its first 100% fully liquid-cooled architecture. Global computing power giants such as GB300/Rubin, Google TPU v7, and Meta Minerva have fully shifted to liquid cooling solutions. As the power consumption of a single chip leaps from 700W for the H100 to 2000W for Rubin and even over 5000W for Vermeer, traditional air cooling has hit its physical ceiling. More critically, policy regulations have left no retreat for data centers to "not use liquid cooling." Liquid cooling is no longer an "optional energy-saving measure," but a dual "mandatory option" for computing power deployment and regulatory compliance.
01 Why Is Liquid Cooling Absolutely Necessary?
The comprehensive popularization of liquid cooling is the inevitable result of the combined forces of technology, business, and policy.
The exponential surge in chip power consumption is the fundamental reason for the industrialization of liquid cooling. NVIDIA's single-chip TDP (Thermal Design Power) has climbed from 700W for the H100 to over 1000W for the B200, and further to 2000W for Rubin and 5000W+ for subsequent platforms; Google's TPU v7 reaches 980W, and the v8P breaks through 1500W. When the power density of a single rack breaks through 60kW and even advances towards 130kW, the cooling limit and PUE degradation of air cooling systems have become irreversible. Liquid cooling systems can significantly reduce the data center PUE to the 1.05-1.1 range by reducing or even replacing high-energy-consuming fan arrays and utilizing high-temperature liquid cooling to achieve natural cooling.
Beyond the hard constraints of physical cooling limits, liquid cooling is also the lifeline for ensuring the continuity of AI training tasks. Large model training typically requires thousands of GPUs to run uninterrupted for long periods (from weeks to months). Air cooling carries the risk of local hotspots, and temperature fluctuations can lead to issues such as training task interruptions, checkpoint rollbacks, increased network synchronization latency, and decreased overall cluster efficiency. Liquid cooling can provide a more uniform and stable temperature environment, allowing chips to operate at lower junction temperatures, which significantly reduces the failure rate. For 10,000-card clusters, any unplanned downtime means millions of dollars in losses, making stability the top priority.
In today's context of increasingly scarce land and power resources, the spatial density advantage brought by liquid cooling is equally critical. Liquid cooling allows for the deployment of more computing power within a smaller physical space. For the same 1MW of power, air cooling may require hundreds of racks and large-scale air conditioning rooms. Liquid cooling can compress the number of racks by 50%-70%, significantly increasing the computing power output per unit area (FLOPS/m²). This is particularly crucial in first-tier cities or hub nodes where land and power quotas are scarce.
If technological and business demands are the internal driving forces for the development of liquid cooling, then increasingly stringent policies and regulations have completely closed the retreat for air cooling. In July 2021, the Ministry of Industry and Information Technology (MIIT) issued the "Three-Year Action Plan for the Development of New Data Centers (2021-2023)," specifying that by the end of 2023, the PUE of newly built large and above data centers should be reduced to below 1.3, and efforts should be made to reduce it to below 1.25 for hub nodes of the "East Data and West Computing" project and cold regions. In November 2021, the National Development and Reform Commission (NDRC) issued the "Implementation Plan for Implementing the Requirements of Carbon Peaking and Carbon Neutrality Goals to Promote the Green and High-Quality Development of Data Centers, 5G, and Other New Infrastructure," further clarifying that "by 2025, the PUE of newly built large and super-large data centers will be reduced to below 1.3, and national hub nodes will be reduced to below 1.25." The eight major hub nodes of the "East Data and West Computing" project require the PUE target in the eastern region not to exceed 1.25 and in the western region not to exceed 1.2, making energy efficiency indicators more stringent. This year, four departments jointly issued the "Action Plan for Promoting Two-Way Empowerment of AI and Energy," explicitly stating: newly built large AI data centers must adopt 100% liquid cooling, all existing air-cooled systems must be retrofitted before 2028, and the liquid cooling penetration rate has become a hard indicator for the approval of computing power bases.
Currently, the vast majority of provinces nationwide have introduced supporting PUE control, reward, and penalty measures. Starting in 2026, Beijing will impose differential electricity prices on data centers with a PUE exceeding 1.35, adding 0.2 RMB/kWh for exceeding the limit by up to one time, and 0.5 RMB/kWh for exceeding it by more than one time; Shanghai strictly controls the PUE of newly built data centers within 1.25, with 1.25 as a hard threshold for intelligent computing centers; the Sichuan Tianfu Cluster offers a one-time reward and subsidy of 20 million RMB for newly built projects with a PUE below 1.25, and 30 million RMB for those below 1.15. Tianjin requires the PUE of large newly built data centers to be ≤1.3, and ≤1.5 for small and medium ones; projects included in the Beijing-Tianjin-Hebei hub nodes need to be ≤1.25. By the end of 2026, all existing projects with a PUE >1.5 must complete retrofits, with the target reduced to below 1.4, otherwise capacity expansion will be restricted.
In short, intelligent computing centers must use liquid cooling not because liquid cooling is "better," but because air cooling can no longer support current and future AI computing power density at the physical level. Liquid cooling simultaneously solves pain points such as cooling limits, energy consumption compliance, spatial efficiency, and operational stability, making it the only path for intelligent computing centers to move from "usable" to "user-friendly and sustainable."
02 Three Technical Routes for Liquid Cooling
The current liquid cooling market presents a pattern of parallel development across three technical routes: cold plate, immersion, and spray cooling.
Cold plate liquid cooling refers to the cooling liquid flowing through the cold plate attached to the chip to complete heat capture. It requires the least modification to existing server architectures and is currently the most mature and widely used solution, accounting for approximately 65% of the market. In November 2025, H3C, together with Intel, Envicool, and Yilian, released the H3C UniServer R4900 G7 full-domain cold plate liquid cooling server. This is the industry's first server based on the dual-socket Intel Birch Stream AP architecture to achieve full-domain cold plate coverage.
In cold plate liquid cooling systems, the liquid cold plate accounts for the highest cost proportion at 32%, followed closely by quick disconnects at 28%, and the CDU (Coolant Distribution Unit) ranks third with a cost proportion of 25%. These three core components together account for 85% of the system cost. With the upgrade of microchannel and pump-driven two-phase cold plate technologies, its heat dissipation capacity has broken through 2500W, and the PUE can be reduced to below 1.10, making it the absolute main force for industrialization volume production in 2026.
Immersion liquid cooling completely immerses the server in the cooling liquid to complete heat capture. It has higher cooling efficiency and accounts for about 34% of the market, divided into two routes: single-phase and two-phase. In single-phase technology, the cooling liquid maintains a liquid circulation, removing heat through sensible heat (temperature increase); in two-phase, the cooling liquid boils and vaporizes on the chip surface, removing heat through latent heat (phase change), and the steam rises and condenses back. The single-phase route has a higher deployment proportion in large-scale intelligent computing centers (such as Alibaba Cloud and ByteDance) due to its simple system and controllable costs. The two-phase route, with a higher cooling limit, is theoretically more suitable for ultra-high power density chips (such as future single chips of 1000W+), but its commercialization scale is relatively small due to the constraints of cooling liquid costs and engineering sealing difficulties.
China Telecom pilots show that immersion cooling can reduce the core temperature by 35% compared to air cooling, with the PUE reaching as low as 1.1. Sugon Data Energy released the megawatt-level phase-change immersion liquid cooling whole rack C8000 V3.0. The power of a single rack exceeds 900kW, and the cooling capacity reaches 3-5 times that of traditional solutions, with a cooling capacity exceeding 200W/cm². Adopting self-developed domestic refrigerants, it can achieve natural cooling throughout the year. It also applies diamond-copper thermal conductive materials on a large scale for the first time, increasing the thermal conductivity by 80% and helping to improve chip performance by 10%.
As the third route, spray liquid cooling currently accounts for only about 1% of the market, yet it demonstrates unique value in precise cooling. This technology achieves heat capture by directionally spraying cooling liquid onto specific heating locations of the server. The MIIT's "Recommended Catalog of Energy-Saving and Carbon-Reduction Technologies and Equipment in the National Industrial and Information Technology Field (2025 Edition)" has included "chip-level precise spray liquid cooling technology" in data center energy-saving and carbon-reduction technologies, which can reduce the PUE to as low as 1.10.
China Great Wall has launched the country's first domestically produced spray liquid cooling server, filling the technological gap; the chip-level spray technology of the Guangdong Heyi New Material Research Institute controls the PUE below 1.1 through independent flow channel design, which is particularly suitable for scenarios with limited space and dispersed heat sources, such as edge computing.
Although it is difficult to shake the mainstream position of cold plate in the short term, spray cooling is becoming an important supplement to cold plate and immersion cooling in fragmented scenarios such as edge AI and communication base stations.
Currently, the US market for the liquid cooling industry is dominated by a few cloud giants (Microsoft, Google, Meta, Amazon), concentrating on building 500MW-1GW level super AI factories. With NVIDIA's whole-rack ecosystem at the core, liquid cooling solutions are highly unified, and standardization is advancing rapidly. The supply chain presents a centralized characteristic of "direct supply from Tier 1 suppliers," opening a window for Chinese manufacturers to directly enter the overseas system.
China, on the other hand, is building a national intelligent computing center network under the guidance of national projects, relying on multiple domestic GPU and liquid cooling manufacturers, presenting the characteristics of "multi-point layout and diversified ecology."
From the Inner Mongolia hub node to the Sichuan Tianfu Cluster, from H3C's full-domain liquid cooling to Inspur Information's largest liquid cooling R&D base in Asia, China's liquid cooling industry is iterating rapidly amidst a hundred flowers blooming. Although this "diversified network" model faces challenges in standard unification and interoperability, it also endows the industrial chain with stronger resilience and iteration speed.
03 Four "Mountains" in the Deep Water Zone of Industrialization
Although the industrialization of liquid cooling is advancing by leaps and bounds in 2026, four "mountains" still lie ahead.
First is the lack of standards. Liquid cooling interfaces, cooling liquid formulations, and piping protocols have not been fully unified, and the ecological compatibility of UQD/MQD quick disconnects directly affects deployment efficiency.
Second is the risk of liquid leakage and the reconstruction of the operation and maintenance system. Liquid cooling systems pose new requirements for data center operation and maintenance, shifting from "fear of power outages" to "fear of water leakage." Insurance, operation and maintenance, and monitoring systems need to be upgraded simultaneously.
Third is the cost of retrofitting existing facilities. For already operating air-cooled data centers, the transition to liquid cooling involves structural renovations such as machine room load-bearing capacity, floor height, and waterproofing. In the short term, the hybrid architecture of "air cooling + liquid cooling" will exist for a long time.
Finally, there are capacity and certification barriers. Transitioning from a "demonstration supplier" to a "mass production supplier" requires crossing three thresholds: capacity preparation, cost control, and product consistency. The certification cycle for large overseas customers is relatively long, and domestic manufacturers need to complete the leap from "able to manufacture" to "dared to use" within the window period of 2026. Meanwhile, the "Implementation Plan for High-Quality Development of Energy-Saving Equipment (2026-2028)" issued by four departments has proposed that by 2028, the proportion of equipment reaching energy efficiency level 2 and above in newly added servers in the information and communication field must exceed 80%, setting a more long-term technical iteration coordinate for the liquid cooling industry.
04 "Water, Electricity, and Coal" Infrastructure in the Era of AI Factories
The essence of liquid cooling industrialization in 2026 is the upgrade of thermal management technology from "supporting components" to the "computing power base." When NVIDIA and Google define the deployment standards for next-generation AI chips with "liquid cooling only," when multiple provincial-level administrative regions force green transformation with PUE red lines, and when domestic manufacturers such as H3C achieve the leap from components to systems with "full-domain liquid cooling," liquid cooling is no longer an "optional choice" for data centers, but the "water, electricity, and coal" infrastructure supporting the operation of AI factories.
The future competition of intelligent computing centers is not only a contest of GPU performance but also an energy efficiency game of "token output efficiency per watt." As the liquid cooling penetration rate breaks through 37% in 2026 and advances towards the critical point of 50% in 2027, liquid cooling will grow from a "niche accessory" to a hundred-billion-level core track, just like optical modules in the past. In this thermal management revolution, domestic manufacturers that have mastered the processes of core components such as cold plates, quick disconnects, and CDUs, and can provide whole-rack system solutions, will become the biggest beneficiaries of AI computing power infrastructure. In 2026, the story of liquid cooling industrialization has just begun.