Battle of Cooling: Who Will Master the Future of High-Power AI Servers?

Can you withstand the heat dissipation test of 200kW + cabinets? Can you control the core supply chain and long-term operations? Can you find a place to survive between high-end monopolies and segmented breakthroughs? Currently in the domestic market, different heat dissipation manufacturers have provided different answers. On one hand, companies represented by Shenzhen Envicool Technology are attempting to become system-level suppliers through a full chain solution; companies represented by Guangzhou Goaland Energy Conservation Tech and Sanhe Tongfei Refrigeration are focusing on a core component, aiming to be professional manufacturers with extreme cost-effectiveness. In the short term, the full chain solution continues to dominate the high-end AI server market, maintaining high customer stickiness and high performance advantages. However, in the medium to long term, companies must find a balance between performance improvement and cost control, while also maintaining sensitivity to new technologies to deal with potential disruption. Chips are becoming "hotter," and effective cooling is key Heat dissipation has always been an important issue in the chip industry, focusing on how to increase computing power while effectively reducing power consumption and heat generation. In the era of CPUs and regular ASICs, heat dissipation mainly relies on PMIC (power management integrated circuit) for fine voltage and frequency adjustments on the chip side, combined with air cooling or simple liquid cooling to control power consumption at the level of hundreds of watts. However, in the era of rapid AI development, it seems that air cooling or simple liquid cooling is no longer effective. To meet the increasing computing power demands in recent years, AI servers need to have excellent high-performance computing capabilities. Accordingly, they require chips with higher computing power and integration, leading to exponential increases in power consumption and heat generation. This trend is particularly evident in NVIDIA, as a leading player in the chip field, where the computational power density of its AI servers has almost doubled in a "generational leap" manner. From the H100 era with single-cabinet power consumption of 30-80kW, to the GB200/GB300 breaking through 100kW, to the new generation Rubin cabinet reaching 200kW or even higher levels, the increase in computing power is no longer just about the performance growth of individual chips, but the simultaneous increase in total cabinet power and heat levels. Once the chip temperature exceeds 70C, reliability will decrease by 5% with every degree increase in temperature. Being in a high-temperature state for an extended period can accelerate chip aging, significantly shorten its lifespan, and even lead to chip deformation or melting. To cool down larger and more comprehensive AI servers, the industry is exploring expanding the heat dissipation limits of air cooling systems on one side, and researching new liquid cooling technologies on the other. Air cooling systems rely on air flow to transfer the heat generated by the chip, with a relatively simple technology, lower price, and applicability to most devices, making it a widely used technology by many manufacturers. However, there is a clear physical limit to the efficiency of air conduction. Currently, 3D VC heat dissipation limits can extend to 1kW, but if the heat further increases, even with increased airflow speed and optimized flow paths, it becomes difficult to achieve effective heat dissipation. Localized hot spots can accumulate in a very short time, rapidly raising chip temperatures. In this situation, the system has to compromise. GPUs may reduce frequencies, limit power, or even shutdown to ensure safe operation. This may result in the inability to sustain output performance. Therefore, new solutions are needed to meet the heat dissipation requirements of 200kW+ cabinets. Liquid cooling systems use liquid as a heat transfer medium, utilizing its high heat capacity and thermal conductivity to quickly remove the heat generated by the chip. Among them, there are mainly cold plate liquid cooling and immersion liquid cooling. Comparing to air cooling, both have stronger heat dissipation capabilities, and once mass-produced, they significantly reduce the cost of heat dissipation. In the new generation Rubin cabinets, NVIDIA has adopted a 100% full liquid cooling design. This change is a response to real constraints: in the era of high power density, liquid cooling is not just a means to improve efficiency, but also a technical condition to support the continuous growth of computing power. When liquid cooling becomes the consensus, disagreements have just begun According to the "Cooling Systems Market Report 2026", the global heat dissipation market exceeded $158 billion in 2025 and is expected to reach $200 billion by 2030 with a compound annual growth rate of 4.8%. Heat dissipation systems play a critical role in controlling total energy consumption in data center infrastructure. In the overall energy consumption of data center infrastructure, heat dissipation systems account for 40%, just behind IT equipment at 45%. Therefore, while ensuring efficient heat dissipation, controlling heat dissipation costs is particularly important. Different manufacturers in the market have provided different answers. On one hand, some companies are attempting to become system-level suppliers by offering full chain solutions, while others are focusing on a specific core component, striving to be professional manufacturers with extreme cost-effectiveness. As the leading supplier of precision temperature control and energy-saving solutions in China, Shenzhen Envicool Technology (002387.SZ) was the first to introduce the highly reliable Coolinside full chain liquid cooling solution. Covering a full range of products from cold plates, quick couplings, manifolds, CDUs, cabinets, pipelines, cold sources, etc., the full chain liquid-cooling products and services of Shenzhen Envicool Technology are constantly extending from the server manufacturing factory to the "factory to site" delivery scenario at data center operation sites. The full chain solution not only eliminates the problem of interface mismatch caused by multiple component suppliers, effectively reducing costs and improving efficiency, but also the design of integrated equipment and standardized UQD interfaces can maximize the resolution of liquid cooling leakage issues. At the 2025 China Data Center Liquid Cooling Technology Conference, Shenzhen Envicool Technology stated that it has delivered over 1.2GW of liquid cooling projects to date, with all projects maintaining a zero-liquid leakage record in actual operation. Once Shenzhen Envicool Technology enters the market, customer stickiness is high. Involving the entire system architecture, the cost of switching customer products is extremely high. In the first half of 2025, Shenzhen Envicool Technology's data center temperature control business achieved operating income of 1.351 billion yuan, a year-on-year increase of 57.91%, with liquid cooling-related revenue from computing equipment and data centers exceeding 200 million yuan. Additionally, the company has become the core liquid cooling service provider for three global computing giants, NVIDIA, Google, and Huawei, becoming a "leader in the full chain of liquid cooling." Compared to system-level manufacturers, another part of the industry is choosing to focus on a specific core component. Guangzhou Goaland Energy Conservation Tech (300499.SZ), as an early domestic manufacturer to enter liquid cooling from the water side, has mastered both cold plate and immersion cooling technologies as its mainstream solutions. The company is the only domestic manufacturer that has all three types of solutions - cold plate, immersion, and containerized solutions. Power Usage Effectiveness (PUE) is a critical indicator for measuring data center energy consumption, and Guangzhou Goaland Energy Conservation Tech has been able to reduce it to below 1.1 using existing technologies, significantly outperforming the industry average of 1.5. Immersion liquid cooling involves completely submerging the heat-generating electronic components in an insulating cooling liquid. Compared to the current mainstream cold plate liquid cooling, this solution, which directly contacts electronic components, has the highest heat dissipation capacity and the lowest PUE range. Leveraging its technical advantages, in the first half of 2025, Guangzhou Goaland Energy Conservation Tech achieved revenues of 137 million yuan in high-power-density device thermal management business, accounting for 32.78% of total operating income for the same period. Although it is widely believed in the industry that immersion liquid cooling is the future trend, Guangzhou Goaland Energy Conservation Tech's layout contributes to sustainable development. However, the initial cost of this solution is extremely high, 2-3 times that of cold plate liquid cooling. Therefore, the cold plate liquid cooling solution remains the mainstream in the market. Sanhe Tongfei Refrigeration (300990.SZ) has a unique competitive advantage due to its industrial temperature control expertise. As a service provider in the field of industrial refrigeration solutions focusing on CNC equipment and power electronic device cooling, Sanhe Tongfei Refrigeration is able to offer advanced temperature control equipment for the semiconductor industry, including refrigeration units using fluorinated liquids, high-precision refrigeration units with temperature control of 0.02C, and efficient heat exchangers capable of withstanding temperatures up to 800C. Additionally, the company has a mature manufacturing system with a 220,000-square-meter production park and highly automated production lines that effectively reduce initial investment costs and production costs. The company places a high emphasis on technology research and innovation, with cumulative R&D investment reaching 122 million yuan in 2025, with a 24.02% year-on-year increase. The focus of research and development is on high-precision temperature control, intelligent systems, and liquid cooling technology in the data center field, expanding its clientele to include industry-leading clients like Kehua Data Co., Ltd. and Dongguan Taishuo Electronics Co., Ltd. However, data center temperature control, especially in AI computing centers, places high demands on the reliability of heat dissipation manufacturers, compatibility with other equipment, etc. Downstream customers have strict and lengthy certification processes for suppliers. As Sanhe Tongfei Refrigeration transitions from the industrial temperature control field to data center liquid cooling, with a lack of long-term operational data and products and technology still in the validation stage, there is significant uncertainty in the company's prospects. Expansion and restructuring of liquid cooling manufacturers for AI servers In conclusion, the development of full chain solutions is still the goal that companies are striving for. Although companies like Guangzhou Goaland Energy Conservation Tech and Sanhe Tongfei Refrigeration are currently focusing on specific components, both companies have stated in their periodic reports that they have the ability to provide comprehensive solutions and produce a full range of products. This creates new opportunities for value enhancement for the companies themselves. Companies are no longer mere manufacturers and suppliers of individual components, but integrators of complete AI server infrastructure solutions. This model can increase customer switch costs and stickiness, add room for premium pricing, and lay the foundation for manufacturers to have a higher voice in the data center ecosystem. Companies are no longer just manufacturers and suppliers of individual devices, but providers of AI server infrastructure. Of course, the development path of full chain solutions also comes with hidden risks. On one hand, companies need to provide a large amount of after-sales and maintenance services, increasing service costs. For example, Shenzhen Envicool Technology's full chain solution promises "5 full guarantees throughout the cycle, worry-free for 5 years." Achieving this goal requires a significant investment in manpower and resources. In the 2024 annual report, Shenzhen Envicool Technology announced a change in accounting policy, shifting after-sales service costs to costs, causing a significant decrease of over 2 percentage points in the company's gross profit margin for 2024. This indicates that after-sales services account for a significant portion of Shenzhen Envicool Technology's operating costs. If costs cannot be effectively controlled, it will be challenging to support the company's long-term development. On the other hand, the company's layout and expansion will closely follow the footsteps of core customers, reducing autonomy and increasing volatility. In the NVIDIA GB200/GB300 NVL72 project, Shenzhen Envicool Technology secured 50% of the initial demand in the Chinese region, with 1200 units ordered by the third quarter of 2025. However, the company's positioning in liquid cooling is heavily dependent on NVIDIA's order demands. If customer orders fall short of expectations, it will create significant pressure on the company. Temperature control technology is in a phase of rapid iteration, and technical routes may change, with the full chain development solution not being the only approach. Whoever can achieve low-cost production of immersion liquid cooling first will strengthen their competitive advantage in the market. This also presents an opportunity for small and medium-sized enterprises to break through existing patterns. Additionally, emerging technologies are brewing, such as Amazon Cloud's direct cooling chip technology attempting to integrate heat dissipation solutions as part of chip delivery, designing motherboards and rack architectures that include heat dissipation channels, bypassing traditional server manufacturers and forming their vertical architecture. If their research and development are successful and widely applied, it may restructure the value chain of the liquid cooling industry. In this context, industry competition is no longer solely based on heat dissipation performance. Companies must find a balance between technology research and development, cost control, and customer ecological layout. Closing Thoughts With the rapid growth of AI computing power, heat dissipation has gradually become a critical factor limiting computing power growth, and the market for AI server heat dissipation is vast. In this scenario, liquid cooling solutions remain the mainstream choice in the coming years. Thus, market competition is intensifying, and manufacturers' paths are beginning to diverge. Currently, system-level manufacturers like Shenzhen Envicool Technology are maintaining high customer stickiness and high performance advantages in the high-end AI server market with their full chain solutions. However, the aforementioned solutions also have risks such as high operational costs, reduced production autonomy, and rapid technological iteration. On the other hand, specialized manufacturers like Guangzhou Goaland Energy Conservation Tech and Sanhe Tongfei Refrigeration are using their technical features to create competitive advantages, increase research and development investments, and bet on future technological development. However, they face challenges like lengthy customer certification processes, high R&D costs, and the risk of changing technology routes. In the medium to long term, players in the liquid cooling race must find a balance between performance improvement and cost control, while also maintaining sensitivity to new technologies to address potential market risks.