Recently, at the NAVIGATE 2026 summit, H3C unveiled the S90000 high-density all-liquid-cooled server, which can reach up to 576 cores per system. H3C said its new high-density power delivery and full-liquid-cooling architecture can cut data center PUE below 1.05.

At the March 2026 GTC conference, NVIDIA also made clear that Vera Rubin NVL72 would be its first 100% liquid-cooled architecture. GB300/Rubin, Google TPU v7, Meta Minerva and other global computing power heavyweights have all moved to liquid-cooling solutions. As single-chip power draw jumps from 700W in H100 to 2,000W in Rubin and even above 5,000W in Vermeer, traditional air cooling has already hit a physical ceiling. More importantly, policy no longer leaves data centers a way to avoid liquid cooling. Liquid cooling is no longer an optional efficiency upgrade, but a required choice for both computing power deployment and regulatory compliance.

Why is liquid cooling necessary?

The full rollout of liquid cooling is the inevitable result of technology, business and policy forces working together.

The explosive rise in chip power consumption is the fundamental reason liquid cooling has become industrialized. NVIDIA’s single-chip TDP has climbed from 700W on H100 to more than 1,000W on B200, then to 2,000W on Rubin and 5,000W-plus on later platforms; Google’s TPU v7 reaches 980W, while v8P breaks 1,500W. Once rack power density exceeds 60kW and moves toward 130kW, the cooling limits of air systems and the resulting PUE deterioration become irreversible. By reducing or even replacing power-hungry fan arrays and using high-temperature liquid for natural cooling, liquid-cooling systems can bring data center PUE down significantly into the 1.05-1.1 range.

Beyond the hard limits of heat dissipation, liquid cooling is also a lifeline for uninterrupted AI training. Large-model training usually requires thousands of GPUs to run nonstop for weeks or even months. Air cooling carries the risk of local hotspots, and temperature swings can interrupt training jobs, roll back checkpoints, increase network synchronization delays and drag down cluster efficiency. Liquid cooling provides a more even and stable thermal environment, allowing chips to run at lower junction temperatures and sharply reducing failure rates. For 10,000-GPU clusters, any unplanned downtime can mean losses of millions of dollars. Stability comes first.

As land and power grow scarcer, the space-density advantage of liquid cooling matters just as much. It lets operators deploy more computing power in less physical space. For the same 1MW of power, air cooling may require hundreds of racks and a large dedicated air-conditioning room. Liquid cooling can cut the number of racks by 50%-70%, greatly improving computing output per square meter (FLOPS/m²). That is especially critical in first-tier cities and hub nodes, where land and power quotas are tight.

If technology and business demand are the internal drivers of liquid cooling, tighter policy and regulation have closed the door on air cooling entirely. In July 2021, the Ministry of Industry and Information Technology issued the Three-Year Action Plan for the Development of New Data Centers (2021-2023), which set a goal for newly built large and above data centers to bring PUE below 1.3 by the end of 2023, and for East Data West Computing hub nodes and cold regions to target below 1.25. In November 2021, the National Development and Reform Commission issued an implementation plan on pushing green, high-quality development of data centers and 5G under carbon peaking and carbon neutrality goals, further specifying that by 2025 newly built large and super-sized data centers should bring PUE below 1.3, and national hub nodes below 1.25. The eight hub nodes in the East Data West Computing project require eastern regions to keep PUE at or below 1.25 and western regions at or below 1.2, with even stricter efficiency targets. This year, four government departments jointly issued an action plan to promote two-way empowerment between AI and energy, stating clearly that all new large AI data centers must adopt liquid cooling, that existing air-cooled systems must be fully retrofitted before 2028, and that liquid-cooling penetration has become a hard requirement in computing power base approvals.

Across most provinces in China, matching PUE controls and reward-and-penalty measures are already in place. Starting in 2026, Beijing will charge tiered electricity prices for data centers with PUE above 1.35, adding 0.2 yuan per kWh for each doubling beyond the limit and 0.5 yuan per kWh once it exceeds the limit by more than one times. Shanghai keeps PUE for new data centers strictly within 1.25, with 1.25 as a hard threshold for AI data centers. Sichuan’s Tianfu cluster offers a one-time subsidy of 20 million yuan for new projects with PUE below 1.25, and 30 million yuan if PUE is below 1.15. Tianjin requires large new data centers to keep PUE at or below 1.3 and small and medium-sized ones at or below 1.5; projects included in the Beijing-Tianjin-Hebei hub node must be at or below 1.25. By the end of 2026, all existing projects with PUE above 1.5 must be retrofitted to bring it below 1.4, or expansion will be restricted.

In short, AI data centers must use liquid cooling not because it is “better,” but because air cooling can no longer support current and future AI computing density at the physical level. Liquid cooling addresses heat-dissipation limits, energy compliance, space efficiency and operational stability at once. It is the only path for AI data centers to move from “usable” to “efficient and sustainable.”

Three Technical Paths in Liquid Cooling

Today’s liquid-cooling market is developing along three parallel tracks: cold plate, immersion and spray cooling.

Cold plate liquid cooling uses coolant flowing through cold plates attached to the chip to capture heat. It requires the least change to existing server architectures and is currently the most mature and widely deployed approach, accounting for about 65% of the market. In November 2025, H3C, together with Intel, Envicool and Unisound, released the H3C UniServer R4900 G7 all-domain cold plate liquid-cooled server, the industry’s first server to achieve full cold-plate coverage based on a dual-socket Intel Birch Stream AP architecture.

In cold plate liquid-cooling systems, the liquid cold plate is the biggest cost item at 32%, followed by quick connectors at 28% and the CDU at 25%; together, the three core components account for 85% of system cost. With advances in microchannel and pump-driven two-phase cold plate technology, heat removal capacity has already surpassed 2,500W and PUE can fall below 1.10, making it the clear volume driver for industrial-scale deployment in 2026.

Immersion liquid cooling fully submerges servers in coolant to capture heat, delivering higher efficiency and accounting for about 34% of the market. It comes in single-phase and two-phase variants. In single-phase systems, the coolant stays liquid and circulates to remove heat through sensible heat transfer. In two-phase systems, the coolant boils and vaporizes at the chip surface, carrying heat away through latent heat before the vapor rises, condenses and returns. Because single-phase systems are simpler and more cost-controlled, they currently see higher deployment in large AI data centers such as those run by Alibaba Cloud and ByteDance. Two-phase systems, with their higher cooling ceiling, are theoretically better suited to ultra-high-power-density chips such as future single-chip 1,000W-plus platforms, but they remain relatively small in commercial scale because of coolant cost and sealing challenges.

A China Telecom pilot showed that immersion cooling can lower core temperatures by 35% versus air cooling, with PUE as low as 1.1. Sugon’s megawatt-scale phase-change immersion liquid-cooled C8000 V3.0 cabinet delivers more than 900kW per rack, with cooling capacity 3-5 times that of conventional solutions and over 200W/cm² of heat dissipation. Using self-developed domestic coolant, it enables year-round natural cooling. It also marked the first large-scale use of diamond-copper thermal material, lifting thermal conductivity by 80% and helping chip performance improve by 10%.

As the third route, spray liquid cooling still accounts for only about 1% of the market, but it has shown a distinctive ability to cool with precision. The technology targets specific heat-generating points on the server by spraying coolant directly onto them. The Ministry of Industry and Information Technology has included “chip-level precision spray liquid cooling technology” in its 2025 recommended catalog of energy-saving and carbon-reduction equipment for the industrial and information technology sectors, saying it can bring PUE down to 1.10.

China Great Wall has launched the country’s first domestically produced spray liquid-cooled server, filling a technical gap. The Chip-level Spray Technology from the Guangdong HeYi New Materials Research Institute uses an independently designed flow channel to keep PUE below 1.1, making it especially suitable for edge computing and other space-constrained, widely distributed heat-source scenarios.

Although it is unlikely to challenge cold plate systems in the near term, spray cooling is becoming an important complement alongside cold plate and immersion systems in fragmented settings such as edge AI and telecom base stations.

For now, the U.S. liquid-cooling market is dominated by a handful of cloud giants - Microsoft, Google, Meta and Amazon - which are building 500MW to 1GW super AI factories at scale. Centered on NVIDIA’s rack-level ecosystem, their liquid-cooling solutions are highly standardized, deployment is moving quickly, and the supply chain is characterized by direct delivery from tier-one suppliers, opening a window for Chinese vendors to plug directly into overseas systems.

China, by contrast, is building a nationwide network of AI data centers under state-led project guidance. Backed by multiple domestic GPU and liquid-cooling vendors, it shows a pattern of distributed deployment and diverse ecosystems.

From the Inner Mongolia hub node to Sichuan’s Tianfu cluster, from H3C’s all-domain liquid cooling to Inspur Information’s largest liquid-cooling R&D base in Asia, China’s liquid-cooling industry is iterating quickly amid a fragmented but vibrant landscape. This diversified network model faces challenges around standardization and interoperability, but it also gives the supply chain greater resilience and a faster pace of iteration.

Four Barriers in the Deep End of Industrialization

Even as liquid cooling races ahead in 2026, four major barriers still stand in the way.

The first is the lack of standards. Liquid-cooling interfaces, coolant formulas and pipeline protocols are not yet fully unified, and the ecosystem compatibility of UQD/MQD quick-disconnect connectors directly affects deployment efficiency.

Second is leak risk and the rebuilding of operations and maintenance systems. Liquid-cooling systems impose a new operating logic on data centers, shifting the fear from power loss to water leaks. Insurance, O&M and monitoring systems all need to be upgraded together.

Third is the cost of retrofitting existing facilities. For air-cooled data centers already in operation, a move to liquid cooling requires structural changes such as floor load-bearing upgrades, higher ceiling clearance and waterproofing. In the short term, hybrid air-plus-liquid architectures will remain in place for a long time.

Finally, there are capacity and certification barriers. Moving from a pilot supplier to a mass-production supplier means crossing three thresholds: capacity readiness, cost control and product consistency. Overseas major customers also have long certification cycles, so domestic vendors need to use the 2026 window to move from “able to build” to “trusted to use.” The Four Departments’ Implementation Plan for High-Quality Development of Energy-Saving Equipment (2026-2028) has already set a longer-term benchmark: by 2028, more than 80% of new servers in the information and communications sector must meet energy efficiency level 2 or above. That gives liquid cooling a clearer long-range technology roadmap.

The Water, Power and Fuel of the AI Factory Era

The essence of liquid-cooling industrialization in 2026 is the upgrade of heat dissipation from a supporting component to the foundation of computing power. When NVIDIA and Google define the deployment standard for next-generation AI chips as “liquid cooling only,” when multiple provincial-level regions use PUE red lines to force green transition, and when domestic vendors such as H3C use “all-domain liquid cooling” to move from components to systems, liquid cooling is no longer a data center add-on. It has become the water, power and fuel infrastructure that keeps AI factories running.

The competition among future AI data centers will not just be about GPU performance, but about the efficiency battle over token output per watt. As liquid-cooling penetration crosses 37% in 2026 and moves toward the 50% inflection point in 2027, it is set to grow from a niche accessory into a core market worth hundreds of billions of yuan, much like optical modules did years ago. In this cooling revolution, domestic vendors that control the core manufacturing of cold plates, quick connectors and CDUs, and can also deliver full-rack system solutions, will be the biggest winners in AI infrastructure. In 2026, the liquid-cooling story is only just beginning.