NVIDIA Vera Rubin Full Liquid Cooling Era Reshapes AI Data Center Thermal Management
On June 21, 2026, NVIDIA officially announced that its next-generation Vera Rubin AI infrastructure will adopt 100% liquid cooling architecture. All chips and network components are cooled by a closed-loop liquid cooling system with a maximum coolant temperature of 45°C, eliminating cabinet fans entirely. This marks a fundamental shift from partial component heat dissipation to full cabinet structural upgrade. As a professional and reliable data center CDU supplier, leading industry vendors are accelerating the launch of high-matching thermal management equipment to support large-scale deployment of Rubin full liquid cooling racks.
The Vera Rubin NVL72 platform is built on the 3rd-generation MGX cabinet architecture, integrating 72 Rubin GPUs, 36 Vera CPUs, 18 computing trays and 9 NVLink switching trays in a single rack. Redesigned with cable-free, hose-free and fan-free structures, the new computing trays reduce assembly and maintenance time from nearly two hours to five minutes. To adapt to ultra-high-density rack layout requirements, the industry is rapidly promoting compact CDU liquid cooling solutions, which feature small footprint and high heat exchange efficiency to perfectly match the compact and integrated design of next-gen MGX racks.
The full liquid cooling heat dissipation path forms a complete closed loop consisting of cold plates, manifolds, quick connectors, CDUs, dry coolers and facility water systems. Different from traditional air cooling, the new architecture transfers heat directly from chips to outdoor terminal equipment through liquid circulation. The mature secondary loop CDU system realizes isolated heat exchange between the IT-side liquid circuit and the facility-side water circuit, precisely controlling system temperature, pressure and flow while avoiding cross-contamination of different cooling media.
According to Morgan Stanley’s valuation calculation, the liquid cooling system cost of a single Vera Rubin NVL144 rack reaches USD 55,710, a 17% increase compared with the GB300 NVL72 platform. Equipped with a sidecar CDU, the total rack thermal management value hits USD 122,100, with a single sidecar CDU valued at approximately USD 50,000. Facing differentiated power density and deployment demands of AI clusters, customized liquid cooling CDU has become the mainstream choice for high-end customers, supporting adaptive parameter tuning and structural optimization for Rubin full liquid cooling scenarios.
The industrial value chain presents a clear hierarchical pricing logic. Cold plates and CDUs occupy the highest value and are prioritized for customer certification and mass commercialization, while manifolds and quick connectors determine the overall operational reliability and O&M efficiency of full-liquid-cooled racks. To further optimize rack-level energy consumption, intelligent energy efficient CDU system solutions are widely adopted, which dynamically adjust flow and pressure according to real-time GPU load, effectively reducing overall data center PUE and operational power consumption.
Cold plate technology continues iterative upgrading for Rubin platforms. The microchannel spacing of Rubin GPU cold plates is reduced from 150μm to 100μm to enhance heat exchange area. In 2026, the industry upgraded the cold plate structure from integral solid copper plate to a combined design of 8mm copper substrate and 2mm copper shell, yet the mass production yield remains merely 50%. For the upcoming VR300 (Ruby Ultra) with a 3.6kW per-module thermal design power, NVIDIA is verifying two cutting-edge solutions: phase-change microchannel cold plates and microchannel cover plates, with the final solution to be confirmed from October to December 2026. In high-density cabinet scenarios, the optimized liquid cooling rack manifold achieves uniform flow distribution across multi-tray and multi-cold-plate branches, eliminating local overheating caused by unbalanced fluid delivery.
In terms of hybrid cooling architecture innovation, data centers are accelerating the integration of liquid cooling and rear-door heat dissipation. The advanced active rear door cooling solution cooperates with internal full liquid cooling loops to assist rapid heat removal at the rack rear end, further improving heat dissipation redundancy for ultra-high-power GPU clusters and ensuring stable full-load operation of Rubin cabinets.
Compared with active cooling modes, the low-energy-consumption passive RDHX cooling system makes full use of natural cold source conditions to assist liquid cooling heat exchange. It effectively reduces CDU and liquid loop operating load, becoming an important energy-saving configuration for medium and low-load periods in large-scale AI data centers and complementing full liquid cooling architecture to achieve optimal PUE control.
In the domestic market, ByteDance and Alibaba lead liquid cooling penetration among hyperscalers, with 30% and 40% of their business migrated to liquid cooling respectively, forming a domestic market scale exceeding RMB 600 million. Cold-plate liquid cooling dominates over 90% of the market, while immersion cooling remains in pilot stages due to high renovation costs and uncertain chip warranty policies. Domestic competition presents distinct differentiation: Invic boasts the most comprehensive product lineup with self-developed cold plates, coolant and quick connectors; Shenling Environment features low quotation and fast response with partial outsourced core components; Vertiv officially launched China-specific liquid cooling products in 2026 after missing market share in the early stage due to high-end overseas strategies.
From the perspective of technical iteration, VR300’s alternative solutions face different risks. The phase-change microchannel solution delivers higher reliability, while the microchannel cover plate scheme requires GPU grooving directly by TSMC, transferring yield risks to the cooling component side with uncertain cost-sharing mechanisms. In terms of cooling media, traditional propylene glycol is gradually being replaced by low-viscosity water-based media to adapt to refined microchannel flow channels. Phase-change fluorinated media are currently tested with overseas suppliers, while traditional refrigerant solutions have been phased out due to environmental and performance defects.
Looking ahead, forward-looking technologies such as diamond-copper composite cold plates and liquid nitrogen immersion cooling are under verification, while traditional full immersion solutions have been phased out due to high maintenance difficulty and excessive load weight. Industry core constraints have shifted from production capacity to yield bottlenecks. The 50% mass production yield of current microchannel cold plates becomes the core supply limitation, and the profitability of domestic manufacturers is further squeezed by fierce price competition. Therefore, overseas high-margin markets and sidecar CDU incremental scenarios have become the core growth directions for domestic liquid cooling vendors.
Globally acclaimed as a premier authority in liquid cooling, EXTRCOOL offers a broad portfolio of thermal management hardware—including dual-mode CDUs, manifold assemblies, cold plates, rear-door heat exchangers, dry coolers, and containerized units—all custom-engineered for AI and HPC data centers. Engineered for dependable performance, seamless system integration, and straightforward scalability, these solutions provide a comprehensive turnkey cooling package. We collaborate with data center operators and enterprise clients across North America, South America, Europe, Asia, and Oceania.
