The Zhitong Finance App learned that CITIC Securities released a research report saying that demand for liquid cooling is spreading from GPUs and CPUs to heat sources for the entire cabinet, such as power supplies and optical modules, and the industry continues to expand; the main investment line for AI cabinet liquid cooling is switching from single-chip cooling to full-cabinet thermal management. As Rubin and Google TPU V8 enter the full mass production cycle, the liquid cooling industry is about to enter a period of concentrated fulfillment of orders and performance. It is recommended to explore investment opportunities in the liquid cooling industry chain around four main lines: 1) Air cooling manufacturers transform to liquid cooling, and integrating existing customers and systems is the greatest asset. 2) Connector manufacturers extend liquid-cooled connectors. After entering BOM, they have strong adhesiveness and increase in value from generation to generation. 3) Thermal management vendors switch to liquid cooling, material/component capabilities extend to the system level, and value jumps. 4) Other manufacturing industries use cross-border liquid cooling, which has the greatest flexibility.
CITIC Securities's main views are as follows:
Rubin paradigm: From “air-liquid mixing” to “zero fans”, the value of a single cabinet has reached the next level
Nvidia officially defines Rubin as “the world's first AI infrastructure to achieve 100% liquid cooling”. There is no fan in the system. It uses 45℃ water intake and single-phase direct touch chip liquid cooling. The demand for mechanical cooling has dropped dramatically, and it can follow the natural cooling path of a “chillerless unit”; the power consumption of the VR200 NVL72 single cabinet rises to 190-230 kW. According to Nvidia's CES, GTC, and earnings call, Vera-Rubin was released in January 2026. Production will climb in late May, mass production will begin in Q3, delivery will begin in Q4, and large-scale mass production from the end of 2026 to 2027. We expect the pace of mass production to advance.
More than Nvidia, full liquid cooling is an industry-level architecture switch
Google TPU v7 (Ironwood) has 9,216 liquid-cooled chips per pod, cluster FP8 computing power 42.5 exaFlops, single-chip TDP of about 980W. 100% liquid cooling is mandatory as standard. The third-generation liquid cooling infrastructure is already in production over 1GW; the TPU v8 uses fourth-generation liquid cooling. According to the AAI conference, the AMD Helios/MI455X integrates 72 GPUs, uses a liquid-cooled manifold+ quick connector for direct liquid cooling. Shipment began at the end of 26Q3, and Q4 release. Schneider's reference design 246kW frame can be expanded to 10.4MW. An industry-level switch means that the demand for liquid cooling spreads from a “single chip” to the entire cabinet chain.
Overseas infrastructure gaps are obvious, and domestic manufacturers are going overseas smoothly
According to TrendForce data, the global AI data center liquid cooling penetration rate increased from 14% in 2024 to 33% in 2025, and is expected to reach 40% in 2026. Looking at the geographical structure of the global liquid cooling market, North America is the largest and Southeast Asia is growing the fastest. It is the region with the tightest local production capacity and high acceptance of China's supply chain. Overseas leading orders have increased, and the Middle East and Southeast Asia have added new main lines. Orders from domestic manufacturers have been verified and completed, and the overseas trip is progressing smoothly.
National standards have been implemented to promote mass production, and gross margin restoration is imminent
2026 is the year of centralized implementation of liquid cooling standard systems, driving the industry from “project formulation” to “standardized mass production” and clarifying the problems of incompatible interfaces and different specifications. The liquid cooling equipment industry will continue its downward trend in 2026H1. It is expected that the second half of the year will show scale effects, gross margin is expected to recover, and the competitive focus will shift from “whether it can be supplied” to “whether the system can be integrated”.
Risk factors:
Risk of demand falling short of expectations; risk of profit pressure due to increased competition; risk that Rubin's mass production and delivery pace falls short of expectations; risk of technology route iteration; risk of overseas geographical and trade uncertainty.