The Data Center Liquid Cooling War: Water-Cooled Screw Chillers Are Replacing Air Cooling

A war without gunpowder has been decided.
In May 2026, H3C dropped a bombshell at the NAVIGATE Summit: the high-density, the high-density, fully liquid-cooled S90000 system, with 576 cores and data center PUE falling below 1.05.
Around the same time, NVIDIA unveiled its trump card, the Vera Rubin NVL72, the first architecture to be 100% liquid cooled at the GTC conference. From the 700W of the H100 to the 2000W of the Rubin to the 500W of the Vermeer, the MCU's power consumption skyrocketed.
Just a few years ago, the primary refrigeration solutions for data centers were air-cooled precision air conditioners and water-cooled screw chillers. They formed the backbone of the industry, with PUE generally ranging between 1.4 and 1.65, while COP is almost constant between 3.5 and 4.5.
But now, they were kicked off the stage.
Not gradual decline, but structural phase-out.
Water-cooled screw chillers: Once the "king of energy efficiency," now a "policy trap." To understand how drastic the change is, we must first look at how powerful how powerful the "king" who was replaced was.
Water-cooled screw chillers, which have twin screw compressors at their core, have cooling capacities ranging from 700 to 1000 kW, with typical COP values ranging from 5.8 to 6.2, and some high-end models even approaching 6.5. They employ shell-and-tube condensers, flooded evaporators and electronic expansion valves, and are equipped with an intelligent microcomputer control system that support stepless energy regulation to accurately meet demand in the 10% to 100% load range.
The system has been used for decades in industries such as chemicals, electronics, pharmaceuticals and new energy, and is undoubtedly the "king of industrial refrigeration." A 500-ton water-cooled screw chiller, which operates 8,000 hours a year, is indestructible in a traditional data center.
But here's the problem:its ceiling is the floor of A.I. chips.
When a rack's power density jumps from the traditional 5-10kW to 60kW or even 130kW, and the chip's TDP jumps from 700W to 2000W, the cooling power of a water-cooled screw chiller is like using a bucket to control a hydrant-it hasn't been tried, but it's practically impossible.
It's more about policy. The National Development and Reform Commission has proposed a hard target for the 2025 PUE, namely the construction and renovation of a number of new and largedata centres by 2025, of which no more than 1.2 will be national hub projects. Shanghai new smart computing centers requires more than 50% of its liquid cooling shelves, and data centers data centers need 50%.
Traditional water-cooled screw chillers support air-cooling systems. It's not a question of being "slightly inferior," it's a question of being "totally unqualified."
There's nothing wrong with wrong Water-cooled screw chillers. They simply go with the flow.
What make the advantages of liquid cooling? Three words: fast, cold, economy. Liquid cooling is not a new technology. IBM began working on liquid-cooled computers in the 1960s and came up with the concept of liquid dielectric cooling in 1966. But it wasn't until the explosion of artificial intelligence computing power that liquid cooling truly became mainstream.
The reason is simple: the physical limits of air-cooled screw chillers and air-cooled and water-cooled screw chillers have been shattered by artificial intelligence chips.
Let's do some heat dissipation calculations. How much colder is air than liquid? The answer: liquids are 1000 to 3000 times hotter than air. This is not a "slightly better" distinction, but a devastating victory in one respect.
When NVIDIA's H200 chip generate more than 30% of heat at high load, and when the power consumption of microcomputers increases from 700W to 200 W, the traditional air cooling system's heat dissipation limit are easily breached like a wall of paper. Air cooling carries the risk of local hotspots; temperature fluctuations can lead to disruptions in training missions, checkpoint rollbacks and increased network synchronization latency. For high-performance clusters, any unplanned downtime can cost millions of dollars.
How does liquid cooling solve this problem? Each method has its own advantages.
The first method: Cold plate liquid cooling --the mainstream solution and the fastest way to industrialize.
Using "ice chips" on the chip, liquids flow through microchannels, taking heat directly with them. The solution requires minimal modifications to existing server architecture and is easy to deploy, making it the top choice for AI infrastructure deployment from 2025 to2026. Both H3C's S90000 and Sugon's C7000-F phase-change cold plate liquid cooling solutions follow this approach.
Cold panel liquid can achieve a PUE of less than 1.2, some optimization schemes can even reach 1.1. Compared with conventional air cooling, liquid cooling improves heat dissipation efficiency by approximately 50% and saves over 30% of wasted energy. A large to medium-sized smart computing center with a capacity of 10 million kilowatts can save about $30 million a year in electricity costs.
The second method: Immersion liquid cooling-the ultimate solution to maximize heat dissipation efficiency.
This involves completely submerging the server in insulating coolant, with the chip coming into direct contact with the liquid. Heat dissipates directly through the vaporization of a liquid or the increase of its specific heat. The solution is thousands of times more efficient than air cooling, with a PUE stability controlled between 1.05 and 1.1, noise below 50 decibels, and power density per unit greater than 100 kilowatts.
Alibaba's immersion cooling system, deployed at the Zhang Bei data center, worked wonders after two years of operation: the cumulative failure rate of components of the liquid-cooled server is 52.9% lower than that of the air-cooled server. This not only saves electricity, but also reduces maintenance costs.
However, immersion liquid cooling also has high entry barriers --entry-high coolant costs, complex operation and maintenance, and a lack of industry standards. Therefore, in the short term, cooling of the cold plate is a priority, while immersion cooling is an adjunct.
Market data is not to be sniffed at: the $20 billion to $76.5 billion war. Numbers are the most honest judgment.
Global investment in liquid-cooled data centers grew steadily to $26.32 billion from 2017 to 2025. The Chinese market was more active,growing from 16.91 billion yuan in 2017 to 76.55 billion yuan in 2025, at a compound annual growth rate 21%%.
By 2026, that figure is accelerating. By 2026, China Galaxy Securities sees the liquid cooling market for AI enabled smart computing centers at about 20 billion yuan, servers and cloud liquid cooling at around 8-10 billion yuan, traditional industries at around 2 billion yuan and China liquid cooling server market at more than 30 billion yuan.
What about the traditional precision air conditioning market? In 2024, China's data center precision air-conditioning market size exceeded 5 billion yuan, growing 29% year-over-year. Does it seem like it's growing? But that's because existing markets are still digesting it. In the long run, liquid cooling technology will gradually compress the market share of precision air conditioners over the next 5-10 years, an irreversible trend.
In terms of competitive landscape, Sugon Digital Technology has a market share share of 55.7% to 61.3% in the liquid cooling infrastructure market, leading the market for 3-4 consecutive years. Using artificial intelligence-efficient algorithms, the Invik's iPower series precision air conditioners reduced PUE from 1.8 to 1.3 in real-world testing, saving more than $300,000 a year in electricity costs. Huawei's Thermal Management Controller (TMU uses AC / DC dual-power structure to achieve "zero interruption" cooling. Schneider Electric's SmartCool terminal air conditioning energy-saving solution Achieve 31 31% Time and power saving.
Instead of sitting on their hands, these traditional participants made a desperate transition to liquid cooling. However, the pace of transformation hasnot kept pace with the market being engulfed by liquid cooling.
Where is the "last battleground" for water-cooled screw chillers? Saying water-cooled screw chillers are being replaced does not mean they will disappear tomorrow.
In low-to-medium-density shelving --traditional data centers, factories and commercial buildings with single-unit shelving of less than 10 kW --water-cooled screw chillers --water-cooled screw chillers remain the most cost-effective option. They are technically mature, easy to maintain, have a low initial cost, and their COP of 5.8 to 6.2 is ideal for these scenarios.
As of early 2025, the total number of data center racks in the country is over 8 million, dominated bymedium and low density racks. These "old racks" won't be replaced overnight by liquid cooling; in this case, water-cooled screw chillers have a grace period of five to 10 years.
However, this grace period does not guarantee safety.
Shanghai, Shenzhen, Guizhou and other places explicitly require that the proportion of liquid cooling shelves in new data centers reach more than 50 percent, with the PUE upgraded to less than 1.4. This means that even if you're still using water-cooled screw chillers today, you'll have to use liquid cooling for your new project tomorrow.
The deeper logic is that liquid cooling displaces not only the cooler, but the entire cooling structure.
The traditional structure is "Cold Water Machine → Cooling Water → Precision Air Conditioning → cold air → chip" with long chain, high loss and a high PUE. Liquid cooling structure is "cold plate/immersion → liquid → chip" with short chain, low loss and a low PUE. In the new structure, coolers are being "de-cooled" -using high-temperature cooling water and liquid cooling panels to dissipate heat directly, bypassing traditional coolers altogether.
"Liquid cooling will be the next trend in AI computing power," Nvidia CEO Jensen Huang said bluntly. "In the future, 45 degrees Celsius of warm water will be enough to cool chips above 80 degrees Celsius without the need for coolers."
This is not a suggestion, but a conclusion.
In a word: not "better," but "necessary." The data center market in 2026 is undergoing relentless elimination process.
The performance of The water-cooled screw chiller has not been lost, andits 6.2 COP has become the ceiling of the air-cooled era. Because of the laws of physics, AI chips are not feasible: the heat dissipation capacity of air and cold water is simply not enough to make AI chips.
Liquid cooling is not a "better choice" but the "only choice." When the MCU uses more than 20 million kilowatts a year, when the policy requires PUE to be less than 1.25, when the 10MW Smart Computing Center uses more than 30 million kilowatts a year,you have no choice.
From water-cooled screw chillers to water-cooled screw chillers to liquid chillers chillers, data center cooling technology has been developed for more than 60 years. Each iteration was driven by the demands of computing power.
But this time, it's not the demand that pushes it to its limits, it's the laws of physics themselves.
The era of water-cooled screw chillers is not over,but its golden age has arrived. The next decade belongs to liquid cooling. This is not because liquid cooling is complicated, but because air-cooled screw chillers and air-cooled and water-cooled screw chillers have truly reached their limits.