{"id":2385,"date":"2026-11-02T01:06:52","date_gmt":"2026-11-02T06:36:52","guid":{"rendered":"https:\/\/rackbank.com\/blog\/?p=2385"},"modified":"2026-11-02T01:09:02","modified_gmt":"2026-11-02T06:39:02","slug":"air-vs-liquid-cooling-gpu-racks","status":"publish","type":"post","link":"https:\/\/rackbank.com\/blog\/air-vs-liquid-cooling-gpu-racks\/","title":{"rendered":"Air vs Liquid Cooling GPU Racks: Sizing by kW per Rack"},"content":{"rendered":"
\n\n\n\n
TL;DR<\/strong>
\nUptime Intelligence places optimised perimeter air at up to 20 to 25 kW per rack and close-coupled cooling, including rear doors, at up to 50 kW.
\nAbove 50 kW, liquid cooling is typically used; as rack power approaches 150 kW or more, cooling becomes total or close to total liquid.
\nNVIDIA describes its GB300 NVL72, a rack-scale system with 72 GPUs, as fully liquid-cooled, so the colocation questions worth testing cover coolant distribution unit (CDU) capacity who owns each loop.<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n

When deploying their own GPU servers in a colocation facility, teams need to evaluate whether air or liquid cooling is better suited to their GPU racks and whether the facility can support the cooling requirements of the chosen configuration.<\/p>\n

Our guide to what makes a data centre GPU-ready for AI workloads<\/a> covers power, network and operational readiness at a broader level. This article takes a closer look at cooling, examining GPU rack power requirements, cooling configurations and the questions teams should ask colocation providers before deployment.<\/p>\n

Rack Power Bands and the Cooling Each One Needs<\/h2>\n

Uptime Intelligence finds that perimeter air cooling is generally effective up to 20 kW to 25 kW per rack with optimised airflow<\/a>, or 10 kW to 15 kW in older systems. Close-coupled air, which places cooling units and piping next to the racks, typically supports up to 50 kW<\/strong>. Liquid cooling, hybrid or total, is typically used above 50 kW, and for rack powers approaching 150 kW<\/strong> or more, total or close to total liquid cooling is required.<\/p>\n

\n\n\n\n\n\n\n\n\n
Rack power band<\/th>\nCooling method<\/th>\nWhat the hall needs<\/th>\nGPU deployments in this band<\/th>\n<\/tr>\n<\/thead>\n
Up to 20 to 25 kW<\/td>\nPerimeter air with hot or cold aisle containment<\/td>\nCRAH (computer room air handler) or fan wall units, containment<\/td>\nLower-powered GPUs or accelerators, which Uptime says can keep a rack below 20 kW<\/td>\n<\/tr>\n
Up to 50 kW<\/td>\nClose-coupled air: in-row units or rear-door heat exchangers<\/td>\nWater piped to the row or to each rack door<\/td>\nA few high-density rows inside a perimeter-cooled hall<\/td>\n<\/tr>\n
Above 50 kW<\/td>\nTypically direct-to-chip cold plates, plus air for residual heat<\/td>\nCoolant distribution units (CDUs), piping in the white space, air or rear doors for the remainder<\/td>\n64-GPU racks of liquid-cooled servers, which Uptime puts at 80 kW to 90 kW<\/td>\n<\/tr>\n
Approaching 150 kW or more<\/td>\nTotal or close to total liquid<\/td>\nCold plates extended to memory and storage, with more than 90 percent of heat to liquid<\/td>\nRack-scale systems as they near this band; Uptime cites the Blackwell NVL72 at 132 kW<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n

Uptime’s 2026 survey reports that average modal rack densities are still rising slowly<\/a>, while a growing number of operators report peak densities of 30 kW or higher. A useful check for any hall: which band was it designed for, and which band does your rack fall in?<\/p>\n

Sources: Uptime Intelligence, AI and cooling: methods and capacities (Feb 2025); Uptime Institute, Density choices for AI training are increasingly complex (Jun 2025); Uptime Institute 2026 Global Data Center Survey press release (Jul 2026).<\/em><\/p>\n

\"Four-step<\/figure>\n

Figure 1. The cooling ladder by rack power, as described by Uptime Intelligence: perimeter air up to 20 to 25 kW, close-coupled and rear-door up to 50 kW, direct-to-chip above 50 kW, and total or close to total liquid as rack power is approaching 150 kW or more.<\/em><\/p>\n

NVIDIA describes the GB300 NVL72 as a fully liquid-cooled, rack-scale architecture<\/a> that combines 72 Blackwell Ultra GPUs and 36 Grace CPUs. For the next generation, NVIDIA said in March 2026 that Vera Rubin NVL72 is in full production and on track to ship in the second half of 2026<\/a>. StorageReview reported on 25 September 2026 that Supermicro is now shipping Vera Rubin NVL72 racks with its direct liquid cooling stack and 1.8 MW in-row CDUs<\/a>. A rack power figure for Vera Rubin NVL72 is not yet published by NVIDIA on the pages cited here, so the power and cooling numbers for your exact configuration belong in the vendor’s site planning document at order time.<\/p>\n

Air vs Liquid Cooling GPU Racks: How Each Tier Removes Heat<\/h2>\n

The three tiers differ in where the heat meets water. Air carries it to a room or row unit, a rear door catches it at the back of the rack, and direct-to-chip cooling takes most of it straight off the processor through a cold plate.<\/p>\n

Perimeter and contained air<\/h3>\n

In perimeter cooling, CRAHs or fan wall units supply cold air, and hot or cold aisle containment manages the airflow. Uptime notes that at rack loads above the 20 kW to 25 kW range these systems may struggle to deliver enough airflow, leading to hot spots, throttling or equipment failures, especially at high utilisation or during maintenance. The upside is breadth: air cooling works with almost all IT hardware.<\/p>\n

Rear-door heat exchangers<\/h3>\n

A rear-door heat exchanger (RDHx) is a door 10 cm to 30 cm deep fixed to the back of the rack that removes heat from the server exhaust air. Uptime classes it as close-coupled air cooling: the servers are still air-cooled, but the doors connect by piping to a secondary chilled water loop, sometimes through a CDU. Manufacturer ratings set the ceiling per product; Motivair, for example, rates its ChilledDoor active rear door heat exchanger at up to 75 kW per rack<\/a>. Uptime describes close-coupled systems as a relatively easy way to add higher rack power in a few rows of a perimeter-cooled hall.<\/p>\n

Direct-to-chip cold plates<\/h3>\n

Direct-to-chip, or direct liquid cooling (DLC), circulates coolant to cold plates on the GPUs and CPUs. CDUs pump that coolant and control its flow, pressure and temperature, and liquid-to-liquid CDUs pass the heat to the facility cooling system, with each CDU or pair of CDUs serving several racks. Uptime adds that cold plate systems still depend on air for 5 to 30 percent<\/strong> of the heat, sometimes up to 50 percent. In its example, a 130 kW GPU rack with 70 percent of heat on cold plates leaves 39 kW for air, which requires rear doors. For the Blackwell NVL72, Uptime reports cold plates handling around 100 kW at full load while air exhaust exceeds 25 kW<\/a>.<\/p>\n

\"Flow<\/figure>\n

Figure 2. The direct-to-chip heat path: the facility water loop feeds a CDU, the CDU drives the technology cooling loop to the cold plates, and residual heat to air leaves through a rear-door heat exchanger or room air. Who owns each hand-off in the contract?<\/em><\/p>\n

In this post, the technology cooling loop means the circuit from the CDU to the cold plates in the rack; Uptime describes this piping as tertiary cooling rings to IT cold plates. Immersion is a fourth option in which servers sit in dielectric fluid. RackBank’s cooling portfolio includes direct-to-chip liquid cooling and our patented Varuna immersion cooling; our post on Varuna and water-conscious AI infrastructure<\/a> covers how immersion works, so it is not repeated here.<\/p>\n

Matching Cooling to Your Deployment, and Where Each Option Stops<\/h2>\n

Four common deployments show how the bands translate into a hall choice.<\/p>\n