Zhongtai: AIDC prosperous outbreak, liquid cooling is the trend

Zhongtai released a research report stating that the chip power consumption index has increased significantly, with PUE energy efficiency constraints and TCO advantages highlighted, and the penetration rate of liquid cooling has rapidly increased. The energy efficiency red line is tightening, with national requirements for new and renovated large/extra-large data centers to have a PUE 1.25, and hub node projects to have a PUE 1.2. Local requirements are even higher, and liquid cooling solutions can effectively help reduce PUE gradually. From an economic perspective, data center costs (TCO) consist of construction costs (Capex) and operating costs (Opex). Liquid cooling solutions can significantly reduce energy consumption and save Opex costs, highlighting long-term TCO competitiveness. It is expected that from Q2, liquid cooling will have sector-specific opportunities, with acceleration expected in May and June. System solution providers and various component manufacturing companies have the opportunity, so it is recommended to pay attention to opportunities in the liquid cooling sector. Zhongtai's main points are as follows: The chip power consumption index has increased significantly, with PUE energy efficiency constraints and TCO advantages highlighted, and the penetration rate of liquid cooling has rapidly increased. The power density of standalone machines is rapidly increasing, with NVIDIA cabinets iterating quickly, reaching over 1 MW in power, while 40-60 kW/Rack has reached the limit for air cooling, making liquid cooling solutions the inevitable choice. The energy efficiency red line is tightening, with national requirements for new and renovated large/extra-large data centers to have a PUE 1.25, and hub node projects to have a PUE 1.2. Local requirements are even higher, and liquid cooling solutions can effectively help reduce PUE gradually. From an economic perspective, data center costs (TCO) consist of construction costs (Capex) and operating costs (Opex). Liquid cooling solutions can significantly reduce energy consumption and save Opex costs, highlighting long-term TCO competitiveness. Cold plate liquid cooling is the mainstream route, while upgrading computing platforms drives the technological process of cold plates, CDUs, and UQDs among other core components simultaneously. In terms of overall solutions, single-phase cold plates, with their high maturity, compatibility with existing server architectures, and complete maintenance system, are the most mainstream liquid cooling solutions for AI servers at present. Moving forward, immersion and chip-level liquid cooling are aimed at optimizing for higher PUE and higher heat flow density scenarios, which may lead to accelerated penetration as power density increases. The evolution of cold plate technology solutions is becoming more diverse, including advancements in microchannels, 3D printing integration, upgrades in materials like copper and diamond-like materials, and breakthroughs in dual-phase liquid cooling technology to continue pushing the heat dissipation barriers, driving liquid cooling systems towards higher heat exchange efficiency, lower thermal resistance, and increased reliability. At the CDU (Cooling Distribution Unit) end, there is a focus on reinforcing pump driving capabilities, pursuing lower temperature differentials, and finer temperature and pressure flow control, evolving from single heat exchange equipment to high-power, intelligent, centralized liquid cooling control hubs. UQDs (Quick Disconnects) and Manifolds are being upgraded towards higher reliability, standardization, and intelligence liquid cooling modules to meet the demand for lower leakage, higher flow rates, dynamic regulation, and fast maintenance capabilities. In addition, in terms of liquid cooling media, traditional fluorinated liquids are facing pressure due to PFAS regulations and the discontinuation of 3M products. Domestic electronic fluorinated liquids and PFAS-Free environmentally friendly media are expected to be rapidly introduced. The liquid cooling sector is growing rapidly, driven by the large-scale adoption of NV+ASIC chips and domestically produced AI clusters, with the global market expanding continuously. NV chip computing cabinets are continuously being upgraded generation to generation, with significant increases in the value per cabinet for liquid cooling. For example: NVL72 cabinets have significantly higher value compared to GB200 cabinets, eliminating air cooling solutions completely at the chip level and adopting an independent cold plate liquid cooling architecture. The number of cold plates has doubled to 126 pieces, a 133% increase in quantity, with the value per cabinet increasing by 57%. The UQD quantity has doubled to 270 pairs, resulting in an 81% increase in value per cabinet. The total value of the heat dissipation module is estimated to be $98,000, with the cold plate/UQD/Manifold/CDU value representing 41%/14%/12%/31%, respectively. The Rubin platform uses a fully liquid-cooled design, with liquid cooling coverage expanding to include more components in the cabinets, completely removing fans within the compute trays and using third-generation cable-free, hose-free modular designs, extending liquid cooling coverage from GPUs/CPUs to server-side NICs, network-side device modules/switches, and power supplies, among other heat-generating components. Key liquid-cooled components have been upgraded, including: 1) using microchannel cold plates, with internal channels shrinking to micrometer scale, increasing manufacturing complexity; 2) using higher density CDUs, simultaneously increasing CDU heat exchange power and flow control precision; and 3) integrating Manifolds. The ongoing optimization of the Rubin platform includes trends towards phase change cold plates, revolutionary cold plate materials, microchannel cover plates, and other changes. It is estimated that from 2026 to 2027, the liquid cooling space in corresponding to NVidia AI server cabinets will reach RMB 63.1 billion to 90.5 billion. North American CSPs are promoting self-developed ASIC chips and domestically made AI clusters as standard liquid cooling, with continuous global expansion: Mainstream North American CSPs are accelerating the development of self-developed ASIC chips and as ASIC server power increases, the adoption of liquid cooling is accelerating. For example, Google's TPU v7/v8 platforms are ramping up production from 2026 to 2027 and are mandatorily equipped with liquid cooling. It is predicted that TPU shipments will reach 4.13 million to 9.87 million units in 2026 to 2027, with v7 and v8 chips accounting for 85% to 100%, bringing a liquid cooling space of RMB 25.7 billion to 76.5 billion. Domestic computing power is bridging the gap in chip manufacturing technology through "clustered, systematized" means, accelerating the deployment of super nodes to break through. Huawei's CloudMatrix 384 cluster single cabinet density reaches 42 kW, with a total power consumption of 559 kW. The substantial increase in power consumption from domestic chip super nodes, coupled with increased application proportions and early liquid cooling configurations due to increased computing power, will lead to rapid deployments of liquid-cooled data centers by domestic internet giants. Assuming that from 2026 to 2027, domestic new AI data center racks will be 5.0/7.5GW, with liquid cooling penetration rates of 38%/54%, the liquid-cooled space for data center racks in China will be RMB 9.8 billion to 21.5 billion from 2025 to 2027. The supply chain landscape has changed, with domestic liquid cooling manufacturers accelerating their market entry. The NV supply chain is mainly dominated by European/American and Taiwanese investments, while domestic companies have adopted diverse models to enter the market: The NV liquid cooling supply chain currently includes: Cold plates mainly from AVC (nearly 30% market share), Cooler Master (market share over 25%), and companies like Digital Revolu... (Note: The text exceeds the character limit, hence it is incomplete.)