Behind AI compute centers, a competition in the liquid cooling supply chain worth hundreds of billions

Behind AI compute centers, a thousand-billion-yuan "cooling war" Source: Xinghai Intelligence Bureau Published: 2026-10-25 16:56 Compute power NVIDIA generate poster "Heat dissipation" has truly begun to become the choke point constraining AI development.

At the end of August, on NVIDIA's latest earnings call, Jensen Huang confirmed that the latest generation of hardware, Vera Rubin, had already begun shipping, with mass shipments starting in September.

On the surface, this was just an ordinary earnings call, but what is noteworthy is this: this time, the Vera Rubin rack has not a single fan, the per-rack power consumption has hit over 190 kilowatts, and the cooling solution has been switched directly to 100% full liquid cooling.

Previously, although GPU compute power and power consumption were still climbing year after year, air cooling on the whole could still keep up. But according to industry experience, the per-rack heat dissipation power that air cooling can handle is roughly only 20–30 kilowatts. Now, however, with the arrival of a new generation of more power-hungry GPUs, air cooling is starting to fall short.

For example: the HPE version of the GB200 NVL72 has a per-rack power consumption of 132 kilowatts, of which 115 kilowatts is carried away by liquid cooling, with air cooling handling only the remaining 17 kilowatts.

In fact, it is not just NVIDIA; even domestic servers have begun to adopt liquid cooling more and more for heat dissipation.

On August 30, just a few days after NVIDIA's earnings call, China's first immersion liquid-cooled compute pod also went into operation in Guiyang, Guizhou.

The head of the local compute center introduced that with traditional air cooling, for every kilowatt-hour of computing task, an additional 0.3 to 0.5 kilowatt-hours of electricity must be spent on heat dissipation; with full immersion liquid cooling, for every kilowatt-hour of computing task, only an additional 0.15 kilowatt-hours of electricity is needed for heat dissipation.

Compared with liquid cooling, traditional air cooling not only consumes electricity but also consumes a lot of water, because many data centers rely on cooling towers to evaporate water for heat removal. However, many traditional data centers not only consume electricity but also consume large amounts of cooling water.

Although in the past, humans also thought of finding some naturally cooler places, such as the caves of Guizhou, Ningxia, Inner Mongolia, and cold regions such as Sweden.

But these naturally cool places are, after all, limited.

Natural water resources and cold regions are limited, and as data center power consumption continues to climb, liquid cooling has become an inevitable technological route.

But the problem is: true liquid cooling is by no means as simple as soaking servers in coolant.

The key to liquid cooling

If one were to pick a few of the most critical parts of liquid cooling technology, those three parts would be the cold plate, the CDU, and the quick connector; these are also the three highest-value segments in the liquid cooling supply chain.

This is because these three segments happen to hold down the "entrance, heart, and joints" of the liquid cooling system; none can be missing, and each has extremely high technical barriers and irreplaceability.

To use the barrel theory as a metaphor, they respectively determine the heat dissipation ceiling, reliability, and stability of the liquid cooling system.

First is the cold plate.

This thing is actually the component that is directly attached to the chip and is responsible for "absorbing" heat into the liquid. There is a value rule in the liquid cooling supply chain: the closer to the GPU, the harder it is to make, and the higher the value.

This is also why the cold plate accounts for 32%-40% of the cost of the entire liquid cooling system, making it the biggest piece of the pie. The reason the cold plate is hard to make is not actually the material of that "plate," but whether heat can be quickly transferred from the material surface to the liquid.

This is because the thermal resistance of transferring heat from a solid wall to a liquid is dozens or even hundreds of times that of heat conduction between solids. The cold plates used for high-end AI chips have a microchannel structure, with densely packed water channels finer than a hair (on the 0.2 mm level). The processing yield of a single cold plate directly determines the cost.

Next is the CDU—coolant distribution unit.

This part is equivalent to the "heart and brain" of the liquid cooling system, distributing flow, regulating temperature, and ensuring redundancy.

Without the CDU, liquid cooling is just a pile of water pipes: the CDU is responsible for distributing coolant to each rack and each node in precise proportions, with flow control accuracy required to be ±2%. If the deviation is a little too large, a node may overheat and go down.

Because the CDU must collect data such as temperature, pressure, flow rate, and conductivity in real time, it cannot use mechanical pumps but instead requires a fluid control system with AI algorithms.

The difficulty here is that these changes are not instantaneously synchronized.

Temperature changes have thermal inertia, pressure waves propagating in pipelines have a water hammer effect, and the CDU needs to dynamically adjust pump speed and valve opening, and also cut off the faulty circuit within milliseconds in the event of a leak.

This requires extremely precise physical coupling.

Finally, the quick connector.

This is the "joint" of the entire system. It may seem inconspicuous, but it determines the reliability of the entire liquid cooling system.

Here is an example from home PCs, and everyone will understand. Friends who have installed water cooling on a PC certainly know that these water cooling setups often have sealed connectors connecting the water pipes and the cold head; custom water cooling commonly uses compression fittings, while all-in-one water cooling mainly uses pre-installed seals. If this thing is of poor quality, the final result is often that coolant "squirts" out, and in the end the entire motherboard, or even the CPU, is ruined along with it.

Similarly, there are hundreds of thousands of liquid cooling connection points in data centers. Each server node needs to be plugged and unplugged for maintenance, and each plugging and unplugging requires pressure, zero leakage, and repeatability of more than 5,000 times.

Home PCs can only have their water pipes unplugged after shutting down, while data centers require online maintenance. These connectors must be unplugged under pressure (with liquid flowing inside) without leaking a single drop.

This requires extremely precise sealing structures, self-locking mechanisms, and residual pressure release designs. Behind this are industrial technologies such as ultra-high-precision machining and special sealing materials.

Around these three key parts of liquid cooling, manufacturers in China and abroad have launched a series of technological breakthroughs, which in turn have driven a huge hundred-billion-level industrial chain.

Who is making the money in liquid cooling?

In the first part, we broke down that cold plates, CDUs, and quick connectors together account for 85% of the cost of a liquid cooling system.

But even more striking than the cost share is the market concentration in these three segments—the global high-end market has long been monopolized by a small number of companies, and they belong to three different "spheres of influence": Taiwanese players control cold plates, European and American players control CDUs and quick connectors, and some leaders have therefore maintained rather attractive gross margins.

For example: on a full-year 2025 basis, Vertiv's revenue was US$10.23 billion, with a gross margin of 36.3%; Dover, the parent company of quick connector maker CPC, had a gross margin of 39.8%; and the gross margins of Taiwanese thermal companies Auras and AVC were 25.81% and 27.38%, respectively.

A closer look at these overseas leaders reveals that although their moats differ, they arrive at the same destination by different paths—what they are essentially selling is not parts, but "certainty."

Because in the liquid cooling industrial chain, the price of parts is always a small matter; the real big matter is the risk of failure and leakage—an improperly designed cold plate may cause GPU throttling; a CDU with failed control may affect several racks; a leak in a quick connector costing tens or hundreds of dollars may threaten an entire set of servers worth millions of dollars. Therefore, what customers are really buying is not a copper plate, a water pump, or a connector, but the certainty that "this equipment will not bring expensive computing power to a halt."

Leading manufacturers often get involved before the GPU, server, and rack are finalized, designing jointly based on chip heat source distribution, rack flow rates, and facility water conditions.

To put it plainly, they know exactly how the GPU was designed from the very beginning.

For example: in 2024, Vertiv jointly released the GB200 NVL72 power and cooling reference architecture with NVIDIA, supporting up to 132 kilowatts per rack. It also participated in NVIDIA's COOLERCHIPS program; after three years of joint research and development, the two parties proposed the industry's first rack-mounted hybrid cooling solution that couples cold plate liquid cooling and immersion liquid cooling into the same system.

Similarly, CoolIT's cooperation with NVIDIA also covers the entire chain from laboratory R&D to mass production deployment: this includes not only "hardware adaptation" but also soft capabilities such as coolant formulation and health monitoring algorithms, making the barrier deeper than one might imagine.

The benefit of this is: matching cooling requirements with data center conditions, avoiding the detours of modifying cold plates, replacing pumps, and reworking after finalization.

The second distinctive feature is mass production consistency;

In liquid cooling systems, making a single sample leak-free is not difficult; what is difficult is keeping hundreds of thousands of products consistent in brazing, sealing, pressure drop, and flow rate. To achieve this, what matters is not subjective "striving for perfection" or "rigorous carefulness," because many defects simply cannot be seen even with great care.

The truly decisive advantage is the process knowledge formed by materials and testing, as well as reliable measurement. For example, Eaton's cold plates use processes such as CNC machining and vacuum brazing, and each unit is tested for leakage and flow performance before shipment, which directly turns visual judgment into instrument-based acceptance.

Finally, and most importantly—the responsibility undertaking mechanism.

To put it plainly, it means having someone to back you up when something goes wrong.

CDU manufacturers not only sell equipment, but are also responsible for water quality, filtration, redundancy, control, on-site commissioning, and fault response. For example, CoolIT's services cover on-site surveys, piping design, system pressure testing and flushing, as well as maintenance and fault repair after commissioning.

For example, when the CoolIT team intervenes to improve the process, it moves piping assembly and flushing to the factory in advance throughout the process. What customers buy is both equipment and an engineering team that finds and solves problems.

This mechanism ensures that after a failure occurs, someone will locate, respond to, and repair it as agreed, rather than leaving the customer alone to coordinate with various parts suppliers; these three mechanisms have in fact formed a mutually reinforcing cycle in the liquid cooling industry: the earlier a company enters co-design and R&D, the earlier it forms certification; with certification and orders, it can continuously improve mass production consistency in actual production; as consistency and yield improve, the company's track record gets better, and it continues to qualify for early intervention.

Faced with such barriers, Chinese companies are prying open gaps one by one with a playbook that is quite characteristic of Chinese manufacturing.

Three keys of Chinese manufacturing

Although China started relatively late in liquid cooling, it does indeed have its own advantages. However, such advantages are not the general "large production capacity," but three more specific systemic capabilities:

The first capability is cross-sector reorganization and entry by switching tracks.

For example, China's Sanhua Intelligent Controls, Yinlun, and Nabai Chuan had, before entering data center liquid cooling, respectively accumulated R&D and mass production experience in automotive thermal management components, heat exchangers, and battery liquid cooling plates. Capabilities in materials, brazing, sealing, and batch manufacturing can serve as the foundation for entering new scenarios.

Another example is Envicool. What it has accumulated in data centers, communications, and energy storage is not just a set of refrigeration equipment, but also capabilities in heat exchange, fluid transport, working fluid compatibility, and system control.

According to company disclosures, in 2020 its water-cooled units were already applied in batches to energy storage, and in 2024 it delivered a full-chain liquid cooling solution for a 5MWh energy storage system.

These accumulated capabilities have been integrated into the data center's Coolinside: from the cold plate attached to the chip, to the quick connectors, manifolds, and CDUs, and further to the working fluid, piping, and cold source, forming a system that can be jointly designed, tested, and delivered.

The value of this path lies in using technology accumulated in other scenarios to win validation opportunities on new platforms, and then gradually entering joint development.

The second strategy is to compensate for system gaps with single-point efficiency;

Specifically, this amounts to accelerating the accumulation of failure data, control experience, and process capabilities through more intensive project deliveries when individual technologies are not superior—that is, using a higher "rate of experience generation" to catch up with the longer experience time of overseas manufacturers.

For example: Inspur and JD Cloud jointly developed a liquid-cooled full cabinet and deployed it at scale, while connecting the production testing of each device to the manufacturing execution system and retaining traceable records.

The advantage of doing this is that the next batch of products can be optimized using the problems found in the previous batch, forming positive feedback.

The third type of strategy: trading efficiency and time for market.

This is the more common approach of domestic companies: rather than waiting for a single component to become the world's best, they simultaneously lay out cold plates, piping, manifolds, CDUs, quick connectors, and control systems. Even if individual technologies are temporarily not advantageous, they enter projects through faster joint development, more efficient efficiency, and lower overall delivery costs; afterward, they use real operating data to improve individual components in turn.

Inspur disclosed in 2023 that through prefabrication of piping and cold sources, construction time was shortened from one month to one week; a prefabricated liquid cooling project for an internet customer was delivered in four weeks.

A similar approach also appears in the cooperation between Kehua Data and MetaX. Kehua disclosed in 2026 that it built a cold-plate liquid cooling POD for the MetaX Innovation Center, integrating water pumps, sensors, and controllers into a hydraulic module, pre-installing and pre-adjusting them at the factory, and taking only 45 days from requirements alignment and design to delivery.

To be realistic, although China will still find it difficult to win in the short term on the battlefield of "setting standards," the biggest difference between liquid cooling and GPUs is that it does not have a moat like CUDA;

This means that as long as key components meet performance targets, the price is lower than others'; supply capacity is stronger; and delivery is faster; you can take orders away from others within standards set by others;

Envicool provides a more concrete progress: an official NVIDIA article has listed it as a Blackwell infrastructure ecosystem partner, and the company's 2026 semi-annual report also explicitly discloses that UQD and MQD products have entered the MGX ecosystem.

At present, Envicool has exhibited the Deschutes 5 CDU custom-developed for Google at the 2025 OCP Global Summit. This at least means that Chinese manufacturers are already able to develop products around the specific needs of leading customers.

This is where China truly has a chance to "win" in liquid cooling: not through a miraculous comeback by a single component, but by turning industrial chain collaboration into computing infrastructure that is affordable, stable to use, and fast to deliver.

This article is reproduced from media reports or online platforms and represents the author's personal position or views. Our reproduction is solely for sharing and does not mean that we endorse or agree with it. If the source attribution is incorrect or your lawful rights and interests are infringed, please contact customer service promptly, and we, as a neutral platform service provider, will promptly correct, delete, or handle it in accordance with the law. Comments: No user comments yet. Latest News More Related Recommendations

Infineon renews contract with Wolfspeed: the "next hurdle" for 200mm silicon carbide 2026-10-25 17:11

Semiconductor-related: the United States designates helium as a critical material 2026-10-25 17:05

Powerchip raises mature-process prices again, with an increase of up to 20% expected in 2027 2026-10-25 17:02