News Categories

Powered by rapidly escalating AI chip heat flux, modern data center thermal management has shifted from energy-saving optimization to mandatory functional support. Advanced liquid cooled server designs are now essential to sustain stable operation of high-density GPU clusters that can no longer be cooled by traditional air systems.

With the single-chip power consumption of AI large model training servers breaking the kilowatt threshold, traditional air cooling has reached physical limits, making liquid cooling a rigid necessity for high-density smart computing centers. Marked as the breakout year for domestic liquid cooling, 2026 witnesses explosive market growth with dual development of cold plate and immersion cooling. As core thermal hardware, the high-performance coolant liquid CDU for data center supports large-scale industrial implementation amid ongoing challenges including cost control, standardized iteration, leakage reliability and domestic material innovation.

With the power consumption of NVIDIA GB300 chips exceeding 1400W per card, traditional air cooling systems have encountered fatal bottlenecks and survival crises. Air cooling, with a maximum rack heat dissipation capacity of merely 20kW, fails to support 100kW+ AI server clusters, resulting in a 50% drop in equipment reliability and service life, as well as frequent system crashes and downtime. Advanced liquid cooling distribution unit delivers precise and balanced coolant circulation, solving uneven heat dissipation issues caused by ultra-high chip power density and stabilizing the overall operation of high-load AI computing facilities.

Most industry practitioners still define liquid cooling as a simple chip heat dissipation method. However, NVIDIA delivered a landmark move at GTC 2026 that overshadowed its new chip launches. For the first time, NVIDIA has taken direct control of the liquid cooling supply chain from traditional OEMs and established an officially certified supplier roster covering four core components: cold plates, quick connectors, manifolds, and CDUs. Liquid cooling is no longer a generic thermal solution, but a strictly certified computing infrastructure locked by NVIDIA’s authoritative qualification system. The upgraded liquid cooling data center ecosystem has become the indispensable foundation for supporting ultra-high-power next-generation AI supercomputing platforms, completely overhauling the traditional heat dissipation logic of artificial intelligence infrastructure.

Liquid cooling has evolved from an optional upgrade to a mandatory requirement for AI computing infrastructure, triggering explosive market expansion. Statistical data shows the global liquid-cooled server market and China-based AI liquid-cooled server market maintain rapid year-over-year growth. Penetration keeps climbing, with global AI data center liquid cooling penetration projected to hit 40 % in 2026 while high-power training servers reach over 70 %. Key driving forces include soaring chip power consumption, strict PUE policy requirements, and NVIDIA’s Rubin platform adopting 100 % full liquid-cooled architecture with 45℃ warm-water circulation. Leading domestic and global players shape competitive patterns across cold-plate, immersion and spray-based technical routes. The mature data center air liquid cooling CDU serves as the thermal control hub, supporting diversified liquid-cooled architectures and helping facilities meet strict energy-saving and low-PUE targets.

A liquid cooling manifold acts as the core fluid distribution component in rack thermal management systems. It takes over cooled coolant from the CDU, distributes the medium evenly to cold plates of individual rack servers for heat exchange, and collects heated fluid to form a complete closed cooling loop. Most AI cabinet designs adopt a dual-side manifold layout with bottom inlet and return water structure. During operational analysis of vertical upward fluid flow, changes in gravitational potential energy and flow velocity jointly affect internal static pressure, which easily causes vertical flow deviation in conventional liquid cooling racks. The professionally structured GPU rack liquid cooling manifold solves basic hydraulic unbalance problems and lays a foundation for uniform fluid distribution in high-density GPU racks.

Also known in the industry as a distribution header, branch pipe, or water collector, a manifold is the core piping component responsible for coolant distribution and collection in high-performance computing data centers and energy storage liquid cooling systems. It acts as the flow traffic hub of the entire liquid cooling circulation loop. A poorly designed data center liquid cooling manifold leads to uneven branch flow: some servers receive insufficient coolant and suffer local overheating and thermal throttling, while other branches get excess flow, destabilizing overall temperature control and, in severe cases, causing computing equipment downtime.

With Rubin entering full mass production, liquid cooling value per rack has jumped from $41,500 for GB200 to $55,700 for NVL144. The market's first reaction was that cold plates and CDUs, accounting for nearly 90% of system value, would be the biggest beneficiaries. But a closer look at high density data center cooling component breakdown tells a different story. Cold plates and CDUs have a mature supplier landscape and certain but limited growth, while the top three components by earnings elasticity are three previously overlooked small parts: Manifold, quick connectors, and plate heat exchangers. Using teardown data from Soochow Securities and Morgan Stanley, we break down where the money in Rubin liquid cooling is really flowing.

The water pump is the core component of a CDU, demanding high standards in sealing, zero leakage, and long-term stable operation. Feilong Shares, a core supplier for N/G projects, has transferred automotive electronic water pump technology to IDC liquid cooling, with electronic pumps in batch supply to major domestic integrators and under verification by overseas cloud leaders. Dayuan Pump Industry brings deep canned motor pump expertise with ultra-low leakage, having passed Google audits and supplying CDU equipment for data centers. Other key CDU internals include plate heat exchangers (10–20% of CDU cost) led by Yinlun Shares, valve components by Sanhua Intelligent Controls, frequency converters by Megmeet, and ultrasonic flow sensors by Audiowell. Together these form an energy saving air liquid cooling CDU that regulates pump speed for efficient thermal control while monitoring flow and detecting leaks.

Liquid cooling is the most deterministic "water seller" segment in the AI compute infrastructure chain. Whether it is NVIDIA GB300 or Huawei Ascend, Alibaba's $80 billion AI investment or Xiaomi's AI chip, the only certain beneficiary of the compute arms race is thermal management. Chip power consumption rises 50–100% per generation, and liquid cooling penetration has surged from 15% in 2024 to an expected 45%+ in 2026. Cold plate liquid cooling is the current mainstream, and the liquid cooling distribution unit, or CDU, is the single highest-value component. Per-rack liquid cooling solution value is approximately 300,000–500,000 RMB/MW for cold plate and 800,000–1.2 million RMB/MW for immersion. The CDU accounts for 25–30% of system value, cold plate assemblies 15–20%, and liquid cooling pumps 8–12%.