Guosheng: Liquid cooling transitions from expectations to mass production, opening a double inflection point for NV + CSP.

Guosheng released a research report stating that from 2025 to 2026, with the full-scale deployment of NVIDIA's high-power computing platform and the launch of GW-level AI data centers by overseas CSPs, the core contradiction of liquid cooling has shifted from "whether to use" to "who can deliver, who can scale up". From H100/H200 to GB200 NVL72, and then to the subsequent GB300 and Rubin/Rubin Ultra, NV is continuously increasing the power density of single server racks through platform-level upgrades. In addition to the NV platform, Google, Meta, Amazon, and other leading CSPs are accelerating the development of in-house ASICs and customized AI servers to optimize computing costs and efficiency. Guosheng's main points are as follows: Liquid cooling has entered a "systematic expansion period" driven by the upgrade of computing platforms and capital expenditures of cloud vendors. From 2025 to 2026, with the full-scale deployment of NVIDIA's high-power computing platform and the concentrated construction of GW-level AI data centers by overseas CSPs, the core contradiction of liquid cooling has shifted from "whether to use" to "who can deliver, who can scale up". Liquid cooling is no longer a technological choice, but one of the most deterministic and clear incremental processes in AI computing infrastructure. The first major turning point comes from NV: liquid cooling is "written into" the delivery standards of computing platforms. From H100/H200 to GB200 NVL72, and then to the subsequent GB300 and Rubin/Rubin Ultra, NV is continuously increasing the power density of single server racks through platform-level upgrades. The power of a full GB200 rack has reached 125-130kW, and the GB300 will further increase this, with the next-generation Rubin platform expected to exceed 200kW. Traditional air cooling is increasingly difficult to support on a scale replication level from an engineering and reliability perspective. NVIDIA is not actively "promoting liquid cooling," but is passively standardizing and popularizing liquid cooling in the evolution of computing platforms, with a high level of certainty in demand for liquid cooling. From a quantitative perspective, NV's roadmap has opened up a space of billions of dollars for liquid cooling. Under conservative assumptions, it is estimated that the market space for liquid cooling systems corresponding to NVIDIA's high-power platforms (GB200/GB300) in 2026 is close to $10 billion. With increasing shipments, the liquid cooling market space corresponding to NV's roadmap in 2028 is expected to increase to over $15-20 billion, and the liquid cooling space brought about by NV alone could exceed one trillion RMB in a single year. The second major turning point comes from CSPs: in-house AI servers and ASIC clusters open up the second growth curve. In addition to the NV platform, Google, Meta, Amazon, and other leading CSPs are accelerating the development of in-house ASICs and customized AI servers to optimize computing costs and efficiency. In-house clusters also exhibit high power density and high integration characteristics, continuously increasing their reliance on liquid cooling. Guosheng estimates that by 2026, the market space for liquid cooling corresponding to CSPs' proprietary systems will exceed $4 billion, and by 2027, it is expected to increase to around $7 billion, further opening up the liquid cooling market space. From an industry perspective, the competition focus of liquid cooling has shifted from technical advantages to supply chain access and engineering delivery capabilities. Liquid cooling is not just a single component, but a system-level engineering that runs through chips, servers, and data centers. In actual projects, the liquid cooling supply chain is often jointly determined by chip manufacturers, CSPs, and ODMs. Being included in the mainstream platform whitelist, and having the ability to deliver at scale and provide long-term reliability verification, are the core factors determining market share. The liquid cooling industry will accelerate the formation of a "strong-getting-stronger" pattern, with alpha focusing on leading manufacturers with full-chain capabilities and deep ties to mainstream computing platforms. Targets: Core recommendation: Shenzhen Envicool Technology (002837.SZ) (full-chain liquid cooling + system-level leader + early advantage in going global). Suggested attention: Guangdong Shenling Environmental Systems (301018.SZ) (capability to expand into overseas markets), Guangzhou Goaland Energy Conservation Tech (300499.SZ) (international layout), Feilong Auto Components (002536.SZ) (core upstream), Shenzhen Cotran New Material (300731.SZ) (back-end outsourcing), Guangdong Hec Technology Holding (600673.SH) (fluoride + components), etc. Risk warning: Progress in computing may be slower than expected, AIDC development may be slower than expected, technological innovation risk, industry competition risk, trade restrictions risk, market size estimation risk.