Dongwu Securities: Liquid cooling becomes the next generation of AIDC rigid infrastructure Rubin brings industry certainty

The Zhitong Finance App learned that Dongwu Securities released a research report saying that the liquid cooling industry is about to enter the engineering implementation stage. Rack-level AI computing will be within the cooling calculation and power supply system; Rubin has entered full mass production, and partner systems will be delivered starting in the second half of 2026. The bank expects that the performance of liquid cooling suppliers in 2026 will be low and high, while 2027 will enter a profit release phase driven by the annual conversion of Rubin orders, increased domestic production share, and product structure upgrades.

The main views of Dongwu Securities are as follows:

Liquid cooling has been upgraded from an optional cooling solution to a rigid infrastructure for AI data centers. Full mass production of VeraRubin marks the industry entering a period of concentrated performance release

Traditional servers mainly rely on air cooling and computer room air conditioning cooling, but rack-level AI computing requires GPUs to achieve low latency and high bandwidth interconnection through NVLink. At the same time, the power consumption of the CPU, switching chip, memory, and power supply rises simultaneously, and the cabinet has been transformed into a computing system with unified power supply, interconnection, and cooling. In high-density scenarios, continuing to increase air volume will rapidly increase energy consumption and airflow organization risks. As a result, liquid cooling and high-voltage power supply are necessary conditions. The Vera Rubin NVL72 single rack integrates 72 GPUs, 36 CPUs, and multiple exchange trays. The cold plate, manifold, quick connector, and CDU must be determined collaboratively during the design phase. Liquid cooling requirements also have heat flow density, power resources, and delivery rigidity: the cold plate is close to the chip, which can shorten the heat transfer path; the electricity saved by the refrigeration system can be converted into GPU deployment capacity; and the high-density cabinet must complete a complete set of liquid cooling tests before leaving the factory. Rubin entered full mass production at the end of May 2026. The third and fourth quarters of 2026 will take the lead in order and shipment growth, forming a full annual contribution in 2027.

Each GW of liquid-cooled IT load corresponds to an equipment market space of about 4.9 billion to 7.7 billion yuan. The core of industry growth comes from power scale expansion

The total value of liquid cooling in a single cabinet of Rubin-grade NVL72 is about 140,000 to 160,000 US dollars, of which the internal parts of the cabinet are about 60,000 to 70,000 US dollars, and the CDU and secondary side systems are about 80,000 to 90,000 US dollars. According to estimates of 150 kW and 220 kW in a single cabinet, each 1GW of liquid-cooled IT load corresponds to about 4545 to 6667 cabinets. In the future, the price per MW may decline as CDUs become larger and interfaces are standardized, but global AI data center power expansion is enough to hedge against unit price pressure. Profit quality differentiation is obvious in all areas: CDU has the greatest value and the highest revenue flexibility, but scale may reduce profit margins; cold plates need to be co-designed with chip packaging and motherboard structures, and processing and yield barriers are further improved after being upgraded to microchannels and jets; quick connectors are less valuable, but can be replicated across multiple ODM and server models after entering the platform BOM, and customer stickiness and return on capital are higher. Vera Rubin uses 45℃ warm water for single-phase direct touch chip liquid cooling, and the supply chain and operation and maintenance system are mature. The main increases in 2026-2028 will be due to flow channel upgrades, increased liquid supply temperatures, and additional liquid cooling cover devices. Although two-phase liquid cooling can handle higher heat flow densities, it still needs to solve problems such as local drying, pressure control, condensation return flow, and material compatibility, and is more likely to be a supplementary solution for the next generation of ultra-high heat flow density chips. As a result, profit pools will be concentrated on high-performance cold plates and high-reliability connectors; Manifold and pipelines are expected to achieve domestic production volume as soon as possible.

Domestic liquid cooling has moved from sample delivery verification to batch delivery. The share increase in 2026-2028 is expected to be faster than the industry's growth

Domestic enterprises are gradually entering the supply system of chip platforms, server ODM, and overseas cloud vendors from peripheral cooling sources and OEM processes. Leading companies such as INVIC have formed an end-to-end product system covering cold plates, quick connectors, Manifold, CDU, cabinets, piping, and cooling sources. Some products have passed Intel testing, and the UQD series has entered the NVIDIAMGX ecosystem, indicating that domestic suppliers have passed the stage of “only samples, no orders”. Domestic replacement will advance from manifolds, pipelines, and some CDUs to highly reliable quick couplings and cold plates to the core BOM, and eventually upgrade to a global liquid cooling platform with self-made core components, system integration, and overseas service capabilities.

Risk warning: the pace of mass production and delivery of the Rubin platform falls short of expectations; global AI data center capital expenditure growth is slowing down; profit pressure due to increased competition in the liquid cooling industry; risk of reliability and verification of key products; risk of overseas market expansion falling short of expectations; risk of liquid cooling technology route evolution exceeding expectations