CITIC Construction Investment: Focus on Investment Opportunities in the Liquid Cooling Heat Dissipation Sector

According to the Zhitong Finance APP, CITIC Construction Investment Securities has released a research report stating that 2025 will be a year of significant improvement in the penetration of NVIDIA's AI chip liquid cooling. At the same time, with the increase in single-chip power consumption, the subsequent liquid cooling market scale is expected to grow significantly. As ASIC cabinet solutions gradually adopt liquid cooling and domestic manufacturers launch super node solutions, along with the improvement in the maturity of the liquid cooling industry chain, the penetration of liquid cooling in the ASIC market and domestic market is also expected to increase rapidly, further opening up market space. It is recommended to pay attention to the liquid cooling sector. 01 Liquid Cooling Heat Dissipation Series Report II: Diamond Materials - The Choice for Efficient Heat Dissipation Breakthrough As the semiconductor industry advances to more advanced processes, chip sizes are shrinking while power is surging, making the "hotspot" issue prominent. Excessive surface temperature of chips can lead to decreased safety and reliability, creating a demand for efficient heat dissipation solutions. Diamond is an ideal heat dissipation material, with a thermal conductivity of up to 2000W/m·K, which is 4-5 times that of copper and silver, and several to dozens of times that of semiconductor materials such as silicon and silicon carbide. It also possesses a high bandgap, extremely high current carrying capacity, excellent mechanical strength, and radiation resistance, making it advantageous in harsh scenarios such as high power density and high temperature and pressure. Its application forms include diamond substrates, heat sink plates, and diamond structures with microchannels, which can meet the core heat dissipation needs of semiconductor devices, server GPUs, and more. In terms of preparation, chemical vapor deposition (CVD) is the mainstream method, capable of producing single crystal, polycrystalline, and nano-diamonds, with domestic and foreign companies having developed related products. With the increase in computing power demand and the development of third-generation semiconductors, the future market space for diamonds in high-end heat dissipation is vast. The chip "hotspot" issue urgently needs to be addressed. As the semiconductor industry gradually advances towards 2nm, 1nm, and even angstrom-level processes following Moore's Law, sizes continue to shrink while power continues to increase, bringing unprecedented thermal management challenges. Chips generate a large amount of heat during operation, and if heat dissipation is not timely, the chip temperature will rise sharply, affecting its performance and reliability. When heat cannot be effectively dissipated within the chip, localized areas will form "hotspots," leading to performance degradation, hardware damage, and cost increases. Diamond is an excellent heat dissipation material. Traditional metal heat dissipation materials (such as copper and aluminum) have good thermal conductivity, but their thermal expansion coefficients struggle to meet high thermal conductivity and lightweight requirements. As a heat dissipation material, diamond has a thermal conductivity of 2000W/m·K, which is 13 times, 4 times, and 43 times that of silicon (Si), silicon carbide (SiC), and gallium arsenide (GaAs), respectively, and is 4-5 times higher than that of copper and silver. When high thermal conductivity is required, diamond is the only selectable heat sink material. Diamond has three main application methods as a heat dissipation material: diamond substrates, heat sink plates, and introducing microchannels into diamond structures. Diamond has significant advantages as a semiconductor substrate material. 1) High thermal conductivity: Diamond has the highest thermal conductivity among known materials, effectively dissipating heat in high power density devices. 2) High bandgap: Diamond's bandgap is about 5.5eV, allowing it to operate stably in high temperature and high voltage environments, making it particularly suitable for high temperature/high power electronic devices. 3) Extremely high current carrying capacity: Diamond's current carrying capacity far exceeds that of traditional semiconductor materials, making it suitable for high current applications 4) Excellent mechanical strength: The hardness and wear resistance of diamond allow it to maintain stable performance under harsh working conditions, increasing the reliability and lifespan of devices. 5) Radiation resistance: The radiation resistance of diamond makes it suitable for use in high-radiation environments such as space and nuclear energy. 02 Liquid Cooling Series Report 1: Thermal Interface Materials - Building High-Speed Heat Dissipation Channels for Chips and Other Electronic Components With the development of high-density chips and packaging technology, the thermal power consumption of electronic components continues to rise. The thermal power consumption of NVIDIA GPUs has increased from 700W for the H100 to 1200W for the B200, and the thermal flux density of mobile phone chips has exceeded 15W/cm², leading to a sharp increase in cooling demand. The market size of thermal interface materials (TIM) in China has grown from 975 million yuan in 2018 to 1.875 billion yuan in 2023, with a compound annual growth rate of 13.97%, showing significant growth. In chip cooling, TIM1 and TIM2 form a "dual thermal conduction engine," with TIM1 directly contacting the chip, requiring low thermal resistance and high thermal conductivity, using fillers such as graphene and boron nitride, which have high thermal conductivity coefficients; TIM2 adapts to heat spreaders and radiators, balancing cooling efficiency and cost, with thermal conductivity coefficients typically ranging from 5-10W/m·K. The two work together to reduce contact thermal resistance and ensure stable chip operation. Additionally, TIM is widely used in consumer electronics and new energy vehicles, accounting for 46.7% and 38.5% respectively, with broad industry prospects as downstream demand upgrades. Increasing Cooling Demand for Electronic Components, TIM as the Core Cooling Component As high-density chips and packaging technology continue to develop, the cooling issues of electronic components have become increasingly prominent. Thermal interface materials (TIM), as core cooling products, are experiencing rapid market growth. TIM is widely used in various fields such as computers, consumer devices, telecommunications infrastructure, and automobiles, primarily to fill the tiny gaps between heat-generating devices and heat dissipation devices, reducing contact thermal resistance and improving cooling efficiency. TIM Has a Wide Range of Applications, Chip Cooling Demand Leading Product Iteration In chip cooling, TIM1 and TIM2 play the role of a "dual thermal conduction engine." The thermal power consumption of NVIDIA GPUs has increased from 700W for the H100 to 1200W for the B200, and the thermal flux density of mobile phone chips has exceeded 15W/cm², leading to a sharp increase in cooling demand. In the consumer electronics sector, as the performance and power consumption of devices such as smartphones and tablets increase, cooling solutions are continuously upgraded. From traditional thermal interface materials combined with graphite films to combinations of heat pipes and heat spreaders, the penetration rate of high thermal conductivity materials is gradually increasing. At the same time, electronic products such as VR/AR devices, solid-state drives, smart speakers, and wireless chargers are also demanding higher cooling performance, and thermal interface materials provide precise cooling solutions for specific scenarios. New Materials Assist TIM in Breaking Through Cooling Capabilities, Domestic Production Rate Expected to Continue to Rise In the future, with the continuous development of new materials, such as diamond materials with superior performance and high thermal conductivity graphene and other nanomaterials, the cooling capabilities of thermal interface materials will be further enhanced. Currently, the global thermal interface materials market is still dominated by overseas companies, but with the improvement of domestic material localization rates and breakthroughs in R&D barriers, domestic companies are expected to gradually increase their market share At the same time, with the continuous expansion of downstream markets such as consumer electronics and automotive electronics, the thermal interface materials industry will welcome broader development space. 03 25Q2 North America CSP capital expenditure increased by 64%, continue to recommend the liquid cooling sector In 2025Q2, the total capital expenditure of the four major internet companies in North America reached USD 95.8 billion, a year-on-year increase of 64%, maintaining a high growth trend, and optimistic about the outlook for subsequent quarters and the whole year, among which Google and Meta raised their guidance for this year. Amazon's 2025Q2 capex was USD 32.2 billion, a year-on-year increase of 83%, and the company stated that the capital expenditure in the second quarter could represent the level of quarterly capital expenditure in the second half of the year; Microsoft's 2025Q2 (the fourth fiscal quarter of 2025) capex was USD 24.2 billion, a year-on-year increase of 27%, and it is expected that the capital expenditure in the next quarter (the first fiscal quarter of 2026) will exceed USD 30 billion (corresponding to a year-on-year growth of over 50%); Google's 2025Q2 capex was USD 22.4 billion, a year-on-year increase of 70%, and it raised its annual capital expenditure from USD 75 billion to USD 85 billion, also stating that investment in 2026 will continue to grow; Meta's 2025Q2 capex was USD 17 billion, a year-on-year increase of 101%, and it raised its annual capital expenditure guidance from USD 64 billion-72 billion to USD 66 billion-72 billion, indicating that it will significantly increase AI investment in 2026. The year 2025 will see a significant increase in the penetration of NVIDIA's AI chip liquid cooling, and as the power consumption of single chips increases, the subsequent liquid cooling market scale will grow significantly. With the gradual adoption of liquid cooling in ASIC cabinet solutions and the introduction of ultra-node solutions by domestic manufacturers, along with the improvement of the maturity of the liquid cooling industry chain, the penetration of liquid cooling in the ASIC market and domestic market is expected to increase rapidly, further opening up market space. It is recommended to pay attention to the liquid cooling sector. CITIC Construction Investment Securities believes that the demand for computing power driven by AI is strong, and continues to recommend the computing power industry chain, including the North American chain and the domestic chain, suggesting continued attention. 04 AI New Era: Forge Ahead, Ignite the Fire AIDC Domestic and International Cooling System Development Trends and Outlook High computing power demand drives the power density of computing centers. Air cooling systems adapt to higher thermal density cooling demands by bringing the cooling source closer to the heat source or sealing the cold/hot channels. As rack density rises above 20kW, various liquid cooling technologies have emerged to meet the cooling needs of high thermal density cabinets. The need for green and low-carbon development of computing centers continues to deepen. PUE (Power Usage Effectiveness, total energy consumption of the data center/actual energy consumption of IT equipment) is the most common performance evaluation index for computing centers and the main measure of the green performance of computing centers in the industry. The closer the PUE value is to 1, the higher the degree of greenness of the computing center. Liquid cooling technology is mainly divided into cold plate, immersion, and spray liquid cooling technologies, among which cold plate liquid cooling technology is the most mature and widely used liquid cooling solution The power density of single cabinets is gradually exceeding 30kW. According to Colocation America, the average power of global data center single cabinets reached 20.5kW in 2023, with the proportion of single cabinets exceeding 30kW continuously increasing. It is generally believed that 30kW is the upper limit for air cooling. With the rapid increase of 30kW+ power cabinets, the cooling method should gradually transition from air cooling to liquid cooling. The market size of liquid-cooled servers is continuously increasing, with cold plate liquid cooling still being the mainstay. From the perspective of market size, data from Guanyan Tianxia shows that the market size of China's liquid-cooled servers will reach 20.1 billion yuan in 2024, a year-on-year increase of 84.4%. It is expected that the market size growth rate will be 46.3% in 2025, reaching 29.4 billion yuan. In terms of market structure, the market share of cold plate liquid cooling is about 65% in 2024; the market share of immersion liquid cooling is about 34%, and the market share of spray liquid cooling is about 1%. Air Cooling: This solution is implemented by placing air cooling components (including Thermal Interface Material (TIM), Integrated Heat Sink (IHS), 3D Vapor Chamber (3DVC), and fans) at the front end of the computing platform. The DGX H100 air cooling solution (including 8 H100 GPUs): The front end of the computing platform is equipped with 3 rows and 4 columns of fans for cooling the 8-card H100; the DGX B200 air cooling solution (including 8 B200 GPUs): The front end of the computing platform is equipped with 4 rows and 5 columns of fans for cooling the 8-card B200. Components required for air cooling: Thermal Interface Material TIM (directly covering the GPU), Integrated Heat Sink IHS (connected to TIM), Multi-dimensional Two-phase Uniform Temperature Component 3DVC (composed of heat pipes and heat sinks, placed on top of IHS), and fans (placed at the front or back of the server). The name 3DVC comes from the 1-dimensional heat pipe, 2-dimensional heat sink, and the 3-dimensional intercommunication of heat pipes and heat sink cavities; VC (Vapor Chamber) comes from the process of liquid evaporation and condensation The principle of air cooling: The heat from the chip is conducted through TIM to the IHS, where the heat enters the 3DVC and evaporates the liquid in the 3DVC into vapor. The vapor is then conducted upward through heat pipes to the upper multi-layer heat sink. The fans at the front and back of the server, along with the air conditioning in the data center, condense the vapor back into liquid within the cavity, creating a continuous cycle. Therefore, air cooling consists of two parts: the multi-dimensional two-phase temperature equalization element above each chip and the fans and air conditioning that serve the entire server's cooling. The higher the thermal design power, the greater the height required for the air cooling components. The air-cooled design of the HGX H100 and HGX B200 is primarily composed of three parts: the power tray, the motherboard (CPU) tray, and the GPU computing tray. The height of the GPU computing tray accounts for two-thirds of the server's height. The height of the chip itself is nearly zero, with the main height coming from the air cooling components: the higher the chip's thermal power, the taller the heat sink required for effective cooling. From HGX H100 to HGX B200, the height of the air cooling components has increased by 50%. A significant amount of internal cabinet space is used for air cooling rather than actual effective cluster computing. Liquid cooling can effectively address the pain points of air cooling. It significantly increases the cooling power of the cabinet. The thermal capacity of water is 4000 times that of air, and its thermal conductivity is 25 times greater than that of air. For the same temperature change, water can store more heat and transfer it much faster than air. The water-cooled computing tray design of the GB200 utilizes an efficient heat exchange mechanism between the cold plate and the coolant, evenly transferring the heat generated by the chip to the surface of the cold plate. The coolant, flowing at high speed through the cold plate, can quickly carry away heat and distribute it evenly. It greatly improves the space utilization of data centers. The height of the air-cooled HGX H100 computing platform is approximately 6U (1U = 4.445 cm). In contrast, the air-cooled HGX B200 requires a height of 10U for the air cooling equipment to meet cooling demands. Comparatively, the height of the GB200 computing tray using DLC is only 1U. With the same deployment of 8 GPU chips, the HGX H100 has a height of 6U, the HGX B200 requires 10U, while the GB200 NVL72 only needs a total height of 2U for two computing trays. Space utilization is significantly improved. Cold plate liquid cooling technology: It indirectly transfers the heat generated by chips and other heating components to the cooling liquid enclosed in the circulation pipeline through a cold plate. The cooling liquid carries away the heat and transfers it to the primary side loop, where it is cooled by the cooling system, ultimately expelling the heat from the system. The cold plate liquid cooling system can be divided into two parts: the primary side (outdoor) circulation and the secondary side circulation (indoor). Among them, the secondary side circulation mainly achieves heat transfer through the rise and fall of the cooling liquid temperature, while the heat transfer on the primary side mainly occurs through the rise and fall of water temperature. In terms of cost, the primary side accounts for about 30% of the liquid cooling cost, while the secondary side accounts for about 70%. Cooling effect: Cold plates generally have the best heat exchange effect when used on flat surfaces, such as CPUs, GPUs, and memory modules, and are not suitable for other components like power supplies and IC capacitors. Overall, cold plate liquid cooling can remove 70-75% of the heat generated by devices in the rack, so a hybrid cooling method is required. Core components of the cold plate liquid cooling secondary side (indoor side): ① Liquid Cooling Plate: A heat dissipation device that absorbs and transfers heat through liquid circulation, widely used in high-performance computing and data centers. They are typically installed on servers or electronic devices to cool the equipment through liquid flow. ② Quick Disconnect Device (QD): Allows for quick and convenient connection and disconnection of liquid pipelines without leakage. ③ Coolant Distribution Unit (CDU): Responsible for the distribution, regulation, and monitoring of the coolant. They ensure that each server receives an adequate amount of coolant to maintain suitable operating temperatures. The CDU is divided into L2A, which includes: RPU (pump, water tank), heat sink, and fan; and L2L, which includes: RPU and brazed plate heat exchanger (BPHE). ④ Internal piping of the cabinet (Manifold): The internal piping includes Rack Manifold and Row Manifold, which are the pipeline systems used to distribute coolant in the liquid cooling system. The Rack Manifold is responsible for distributing coolant to various servers in the rack. The Row Manifold is responsible for distributing coolant to each row of servers within the rack. Value breakdown of the cold plate liquid cooling cabinet (taking the GB200 NVL72 cabinet as an example). The overall value of the internal liquid cooling system of the GB200 NVL72 cabinet is approximately $84,000, accounting for about 2.8% of the cabinet cost (assuming the total cost of the cabinet liquid cooling is $3 million). In terms of components, the GB200 NVL72 cabinet contains 126 chips, of which the Compute Tray contains 108 chips (72 GPUs + 36 CPUs) The corresponding value of the cold plate is approximately $32,000; the corresponding value of the Switch Tray chip is approximately $3,600, totaling 43%. The unit price of the cooling distribution unit (CDU) is $30,000, accounting for 35.8%. The value of quick connectors accounts for approximately 10.5%, and the manifold accounts for about 4.8%. Overall, the liquid cooling plate and CDU together account for 78.8% of the total liquid cooling cost, making them the core components of the liquid cooling solution. Immersion liquid cooling is a cooling technology that directly immerses all or part of heat-generating electronic components (such as CPUs, GPUs, memory, and hard drives) in a chassis filled with a non-conductive inert fluid medium. It consists of two cycles: the primary side cycle utilizes outdoor cooling equipment (such as cooling towers or chillers) and heat exchange units (such as CDU) for heat exchange, discharging the heat from the cooling liquid; in the secondary side cycle, the CDU exchanges heat with the IT equipment in the liquid cooling box, transferring heat to the cooling liquid. Based on whether the cooling liquid undergoes a phase change during the circulation cooling process, it is divided into single-phase immersion liquid cooling and two-phase immersion liquid cooling. ① Single-phase immersion: The secondary side cooling liquid, as the heat transfer medium, only undergoes a temperature change during the heat transfer process without any phase change, relying entirely on the sensible heat change of the substance to transfer heat. ② Two-phase immersion: The secondary side cooling liquid, as the heat transfer medium, undergoes a phase change during the heat transfer process, relying on the latent heat change of the substance to transfer heat. Due to the direct contact between the heat-generating components and the cooling liquid, immersion liquid cooling has higher heat dissipation efficiency, lower noise compared to cold plate and spray cooling, and can solve the heat dissipation problem of high-density cabinets. ① In single-phase immersion liquid cooling, the dielectric cooling liquid (with a higher boiling point) remains in liquid form, and electronic components are directly immersed in the liquid, transferring heat from the electronic components to the liquid. Typically, a circulation pump is used to flow the heated cooling liquid to the heat exchanger, where it is cooled and circulated back to the container. ② Phase change immersion liquid cooling uses phase change cooling liquid (with a lower boiling point) as the heat transfer medium. In operation, when the temperature of the cooling liquid rises to the boiling point corresponding to the system pressure, the cooling liquid undergoes a phase change from liquid to gas, absorbing heat through vaporization to achieve heat transfer. Considering factors such as initial investment costs, maintainability, PUE effects, and industry maturity, cold plate and single-phase immersion cooling technologies have advantages over other liquid cooling technologies and are the mainstream solutions in the industry. Cold plate liquid cooling can achieve a smooth transition from traditional air cooling modes and is applied more in the data center field.