The explosion of large AI models has made chips increasingly powerful, but the resulting "high fever" problem has also become more severe. If cooling can't keep up, even the best chip will slow down after just two minutes, and computing power will be reduced. It can be said that cooling capability has become an "invisible bottleneck" that determines whether AI computing power can truly be unleashed.

This "invisible business," once hidden in the corners, is now booming with the surge in token demand. Data shows that the global data center cooling market is expected to nearly fivefold in the next eight years, reaching a scale of 869 billion RMB.

In this cooling revolution, liquid cooling technology has undoubtedly become the "standard" for newly built intelligent computing centers. At the same time, domestic substitution in traditional air cooling and other fields also holds huge business opportunities. The crazier AI computing power is, the more hot this cooling sector is.

| Giants are spending hundreds of billions to compete for the "cooling down" business

In July this year, global water treatment giant Ecolab spent $4.75 billion to acquire Canadian liquid cooling company CoolIT Systems. A few months ago, electrical equipment giant Eaton also spent $9.5 billion to acquire thermal management company Boyd Thermal. The two deals totaled $14.25 billion, and the buyers were not competing for large models or GPUs, but for a group of companies producing cold plates, fans, pumps, heat exchangers, and thermal conductive materials.

Money suddenly poured in because AI servers are getting hotter and hotter.

AI computing power has an unavoidable physical law: a large portion of the power consumed by chips eventually becomes heat. The more chips there are, the longer the runtime, and the hotter the data center. From the NVIDIA H100 to GB200 and then Rubin, the power consumption per card jumped from 700W to 1200W, and then to 2300W. The maximum for air cooling is about 800 watts per chip; beyond this threshold, no amount of fans can hold it down. By the Rubin generation, NVIDIA directly announced that the entire system must come standard with liquid cooling—not a matter of choice, but of physics.

If cooling is insufficient, even the most expensive GPUs cannot consistently deliver performance. With more stable temperature control, the chip can maintain higher performance during continuous inference tasks. High temperatures also accelerate the aging of electronic components, packaging materials, and connecting parts. For a server worth several million yuan or even more, reducing one failure and extending the effective lifespan mean more token output. This year, the demand for large model inference has increased, making smooth token production the top priority for major vendors.

Last year, the industry debated whether liquid cooling is really the trend. By this year, no one is discussing this issue anymore. Nowadays, everyone is discussing how to design cold plates, choose coolant, match thermal interface materials, and improve reliability. Liquid cooling orders in the North American market are growing rapidly, and some Chinese mainland and Taiwanese suppliers are already facing issues with placing orders.

According to available information, Yinlun Co., Ltd., which produces microchannel cold plates, is expanding overseas chip packaging and server customers with liquid cooling plates, heat exchangers, and other products; Invek is a core supplier of Google's TPU liquid-cooled CDUs. The overseas business of leading domestic companies is becoming the second growth curve.

Institutions estimate that the domestic liquid cooling market will reach nearly 30 billion yuan by 2026, and is expected to surpass 100 billion yuan by 2030, with a compound annual growth rate of over 45% from 2025 to 2030.

| Four ways to make money with liquid cooling

Many people think the liquid cooling industry is just about selling equipment to make money, but in reality, the way to make money along this industry chain is far more complex than imagined. Currently, there are four main 'money-making' models in the industry:

First: selling equipment (the most competitive "Red Ocean")

This is the most direct model, such as selling liquid-cooled distribution units (CDUs). A CDU of 25 to 30 kilowatts sells for 200,000 to 300,000 yuan on the market. Vedi Technology is a prime example of this model.

But the ceiling for this business is very low, and competition is extremely fierce. Although many domestic companies are seeing revenue growth, their profits have not increased in tandem. Because CDU and cold plates are essentially close to traditional manufacturing, trapped in the "volume but no profit" internal competition quagmire.

Second: Becoming a Tier 2 Supplier (a "shortcut" to quietly getting rich)

There are many companies in Jiangsu and Zhejiang specializing in radiators and pump sets, but without major brands, it's difficult to directly enter the supply chains of NVIDIA or North American customers. Therefore, they chose to become Tier 2 suppliers to server and rack companies in Taiwan or North America, using OEM production to "save the country indirectly."

Such companies actually live quite comfortably. Because Taiwanese companies are deeply tied to Nvidia, domestic companies can indirectly enter the North American market through them, so orders are not a concern.

Third: Operations and maintenance services (a steady, long-term "good business")

Traditional data centers fear water entering the server room, but liquid cooling systems have a complete liquid circulation system that requires regular cooling fluid testing and checking for corrosion, impurities, microorganisms, and leakage risks.

Once this system is put into operation, long-term maintenance needs will be continuous. This model requires low initial investment and mainly relies on personnel and service systems, enabling the formation of very stable long-term income.

Fourth: Joint R&D from the source (the most profitable future "dimensionality reduction strike")

Industry insiders generally believe that the future cooling will truly make big money from cooling with models like the one Jizhi Cold Source is exploring: no longer waiting for the chip to be finished to "supplement" cooling, but instead defining the entire cooling solution from the chip design stage together with chip manufacturers and packaging factories.

| Domestic cooling products are experiencing a "comeback" moment

In the past, everyone focused on chips, algorithms, and applications, thinking cooling chips was just an inconspicuous "dirty and exhausting job." But in reality, achieving the ultimate in performance, noise, lifespan, and cost of cooling fans requires a very high technical barrier. Now, with the explosive growth of AI computing power, this once "supporting actor" has finally made a comeback.

Why pursue "domestic substitution"?

The rise of domestic cooling fans follows a very clear logic:

Safety: In the past, cooling components were heavily reliant on Japan and Taiwan. In today's complex international environment, once the supply chain changes, downstream companies are in a very passive position.

Profit: The cooling industry was previously overlooked, but its gross profit margin is actually not low. As long as it can match the level of its Taiwanese peers, the profit margin is considerable.

The dividends of domestic substitution have already begun to be realized. It is reported that a Guangdong-based cooling fan company has revenue approaching 1 billion yuan and is expected to become the first fan company in mainland China to reach this scale.

Even more impressive is its growth curve: after entering AI data center businesses such as robotics, car systems, robot dogs, and Alibaba Cloud, revenue in new areas has nearly doubled every quarter.

Besides fans, upstream cooling materials are also being rapidly upgraded. From traditional aluminum, copper, and silver, it rolled all the way to diamond. Some companies producing synthetic diamonds in Henan suddenly became popular as a result. Diamond's thermal conductivity far exceeds that of metals, making it highly anticipated and expected to become the optimal heat dissipation substrate for high-power AI chips.

Kingtech's Perspective | From Optional to Essential Demand

With leading manufacturers like NVIDIA making liquid cooling standard, the competition in AI computing power has extended to physical cooling capabilities. The liquid cooling market is moving from early "pilot demonstrations" to the stage of "large-scale mass production." This is not just a technological iteration, but also a rigid restructuring of data center infrastructure. In the future, companies with full-stack liquid cooling solutions and deep ties to leading computing power vendors will be the first to benefit from the industry's explosive growth.

Avoiding "hardware involution"

In the liquid cooling industry chain, simple equipment assembly (such as CDU) is prone to falling into a red ocean price war. Investors should focus on core components with high-precision manufacturing barriers (such as microchannel cold plates, quick couplings, magnetic levitation compressors), as well as "liquid cooling operation and maintenance service providers" that offer long-term stable cash flow.

Meanwhile, as domestic AI chip power consumption rises, precision thermal component companies with domestic substitution logic and successfully entering the supply chains of leading domestic manufacturers will see a double hit on performance.

Forward-looking Layout of "Next-Generation Cooling Materials"

When traditional metal heat dissipation approaches physical limits, innovation on the material side will become the key to breaking through. Third-generation semiconductor materials such as diamond and silicon carbide are crossing over from "industrial teeth" to "cooling patches" for AI chips.

Investors are advised to closely monitor "specialized, refined, distinctive, and innovative" enterprises that have mass production capabilities in cutting-edge materials such as CVD polycrystalline diamonds and silicon carbide thermal substrates, and have already entered the advanced packaging industry chain. These companies not only possess extremely high technological moats, but are also expected to achieve significant valuation premiums in the future era of ultra-high-density computing power.