Data Center Cooling Explained: Air, Direct-to-Chip Liquid and Immersion

Data center cooling has gone from a facilities detail to a strategic decision that determines which chips a building can host, how much power it wastes and how much water it consumes. This guide explains air, direct-to-chip liquid and immersion cooling, what each costs and who supplies it. It is part of our Data Center Economy hub; for the power side of the equation read why power is the new data center bottleneck, and for how cooling feeds into running costs see data center operating costs, PUE and staffing.

Why is cooling suddenly so important for data centers?

Cooling matters because AI chips concentrate far more heat in each rack than conventional servers. Average racks drew under 8 kW in Uptime Institute surveys, while Nvidia’s current AI racks draw roughly 15 to 17 times that, overwhelming the air systems most facilities were built around.

Every watt of electricity a server consumes becomes heat that must be removed. For decades, operators handled this with computer room air handlers, raised floors and hot-aisle containment. Uptime Institute’s 2024 Global Data Center Survey found average rack densities were rising slowly but still below 8 kW, and noted that relatively few facilities had racks above 30 kW. About one-third of operators said they were developing capacity for high-density cabinets.

AI changed the curve abruptly. Graphics processors used for training and inference draw hundreds to over a thousand watts each, and Nvidia packages 72 of them with CPUs and networking into a single rack-scale system. The result is a step change in heat density that no amount of fan speed can handle efficiently.

Cooling is also an energy cost. The IEA estimates cooling can account for around 7% of electricity use at efficient hyperscale facilities but more than 30% at less efficient enterprise sites. Uptime Institute’s surveys have shown average power usage effectiveness (PUE) stuck around 1.54 to 1.56 for several years, meaning a typical facility uses roughly half again as much energy as its IT equipment, much of it for cooling.

How much power do AI racks draw today?

Current Nvidia rack-scale AI systems draw roughly 120 to 140 kW. Operator and analyst estimates place the GB200 NVL72 near 120 kW and the GB300 NVL72 at about 135 to 140 kW, with peaks reported up to 155 kW. Nvidia’s roadmap points to 600 kW racks in 2027.

These figures are not vendor specifications, and they vary with configuration and cooling design, but they are consistent across operators and analysts. Both the GB200 and GB300 NVL72 require direct liquid cooling; air cooling was never a viable option at these densities. The Vera Rubin NVL144 generation, arriving in the second half of 2026, is expected to land at a similar 120–130 kW per rack.

The bigger jump comes next. Nvidia has previewed its Rubin Ultra system in a new Kyber rack design at around 600 kW, roughly five times today’s systems, targeted for 2027, with a sidecar unit for power and cooling that effectively occupies two rack footprints. Nvidia is also moving to 800-volt direct current power distribution, which it says can improve end-to-end efficiency by up to 5% and support racks from 100 kW to over 1 MW.

Source: Uptime Institute Global Data Center Survey 2024; Spheron and Introl summaries of operator and analyst estimates (2026). Nvidia rack figures are not official vendor specifications.

How does traditional air cooling work, and where are its limits?

Air cooling blows chilled air through servers and removes the hot exhaust using computer room air handlers, chillers and cooling towers. It works well up to roughly 20 to 30 kW per rack, and with rear-door heat exchangers can stretch higher, but it becomes inefficient and impractical beyond that.

Classic air-cooled halls use a chilled-water plant or direct-expansion units to cool air, which is pushed through a raised floor or overhead ducts into cold aisles. Containment systems keep hot and cold air from mixing. In dry or temperate climates, operators use outside air or evaporative cooling for much of the year to save energy, which is one reason water use has become a sustainability issue.

Air has a fundamental physics limit: it carries far less heat per unit volume than water. Moving enough air through a 120 kW rack would require enormous fan power and airflow, with noise and hot spots that threaten reliability. Rear-door heat exchangers, which put a water coil on the back of the rack, extend air cooling’s range and are a common retrofit step, but they cannot handle the heat flux of modern GPUs directly.

Air cooling is not disappearing. Even in liquid-cooled AI racks, networking gear, storage and power supplies often remain air-cooled, and most of the world’s installed base of enterprise and colocation space still runs on air. Uptime Institute’s 2025 cooling survey described liquid cooling adoption across the broad industry as slow and steady, with most operators still relying on air for most IT equipment.

What is direct-to-chip liquid cooling?

Direct-to-chip liquid cooling pumps a coolant, usually treated water or a water-glycol mix, through metal cold plates mounted on processors and memory. The liquid carries heat to a coolant distribution unit, which transfers it to the facility water loop.

This is the approach Nvidia has standardized for its rack-scale systems and the one most hyperscalers and colocation providers are deploying at scale. A typical architecture has three loops: the technology cooling loop inside the rack, the coolant distribution unit (CDU) that isolates it, and the facility water system that rejects heat through dry coolers, chillers or cooling towers. Because the liquid captures most of the heat right at the chip, facility water can run warmer than in air-cooled designs, which enables more hours of free cooling and less chiller energy.

Direct-to-chip systems typically capture the bulk of rack heat, with the remainder handled by air. That is why most AI halls are hybrid: liquid for GPUs and CPUs, air for everything else. Single-phase cold plates are most common, while two-phase systems, in which a dielectric fluid boils on the cold plate, are emerging for the highest densities. Uptime Institute reported in 2025 that operators are warming to dielectric cold plates.

The main challenges are plumbing and operations: leak detection, fluid quality, filtration and flushing, quick-disconnect fittings, and training technicians who are used to air-cooled servers. Those operational services have become a business in their own right, which is why equipment makers have been buying specialist firms, as described below and in our article on the data center supply chain.

What is immersion cooling and when does it make sense?

Immersion cooling submerges entire servers in a non-conductive dielectric fluid, either single-phase (the fluid stays liquid and is pumped through heat exchangers) or two-phase (the fluid boils and condenses). It can handle very high densities and remove nearly all heat without fans.

Immersion’s appeal is simplicity at the thermal level: there are no server fans, heat transfer is excellent, and very low PUEs are achievable because almost all heat leaves in liquid form. Two-phase immersion has demonstrated PUEs close to 1.0 in some deployments. Nvidia’s Kyber-generation descriptions also point toward fully liquid, fanless designs.

Yet immersion remains a niche relative to direct-to-chip. Servers must be designed or adapted for submersion, maintenance means lifting hardware out of tanks, tanks change floor-loading and layout, and some dielectric fluids, particularly certain fluorinated chemicals used in two-phase systems, face environmental and regulatory scrutiny. Vendor warranty and support models for immersed hardware are still maturing.

Immersion makes most sense for purpose-built facilities, edge sites with no room for air handling, crypto-mining and high-performance computing operators, and environments with dust or harsh conditions. For mainstream hyperscale AI, direct-to-chip has won the current generation; immersion’s role could grow if rack power heads toward 600 kW and beyond.

How much does liquid cooling cost compared with air?

Liquid cooling raises upfront mechanical costs per rack but usually lowers cost per kilowatt of IT load at high densities, because one liquid-cooled rack replaces many air-cooled racks and reduces fan and chiller energy. Exact costs vary widely by design and are rarely disclosed.

Market data show where spending is heading. Dell’Oro Group has repeatedly raised its liquid cooling forecast; in an update reported by Fierce Network it projected liquid cooling revenue of about $4.3 billion in 2028, roughly a third of a data center thermal management market it puts at about $13 billion, with cumulative liquid cooling revenue of about $15 billion between 2024 and 2028. Air cooling growth, by contrast, is expected to slow to mid-to-high single digits. For overall build costs, including mechanical systems, see our guide to data center construction cost per MW.

On the operating side, liquid cooling’s economics come from three places: less fan energy inside servers, warmer facility water that allows more economizer hours, and higher density that reduces building footprint per megawatt. Offsetting this are higher maintenance and fluid management costs and the expense of redundancy in pumps and CDUs.

For most new AI builds, the cost comparison is academic: the hardware simply will not run without liquid cooling. The real economic question is whether to build fully liquid-ready halls now, at higher cost, or hybrid space that can be converted later. Many developers are choosing flexible designs that can serve both air-cooled cloud tenants and liquid-cooled AI tenants.

Who are the main data center cooling vendors?

The leading suppliers include Vertiv, Schneider Electric, Eaton, Johnson Controls, Trane and Carrier on the facility side, plus specialists such as CoolIT, Motivair, Boyd and immersion firms. Electrical and thermal majors have been buying liquid cooling specialists to secure capability.

Vertiv is the most visible pure play. It reported 2025 revenue of about $10.2 billion and ended 2025 with a $15 billion backlog, up 109% year on year. In Q3 2026 it posted net sales of $4.31 billion and raised its full-year 2026 sales guidance to $14.45–14.85 billion. Vertiv has spent more than $1.2 billion since late 2025 on liquid cooling and fluid management businesses, including the roughly $1 billion purchase of PurgeRite, which provides flushing and filtration services, Strategic Thermal Labs and ThermoKey. It also acquired UK-based CoolTera in December 2023 and is working with Nvidia on 800 VDC power architecture.

Schneider Electric agreed in October 2024 to acquire a controlling interest in Motivair, a liquid cooling specialist, for about $850 million, and has partnered with Nvidia on cooling reference designs. Specialist suppliers such as CoolIT and Accelsius have demonstrated cooling for racks of around 250 kW, according to industry reporting. Chipmakers and server makers, including Nvidia, Dell, HPE and Supermicro, increasingly ship systems with integrated liquid cooling, which shifts some value from facility vendors to IT vendors.

How does cooling affect data center water use?

Cooling is the main driver of direct water use at data centers, mainly through evaporative cooling towers. Lawrence Berkeley National Laboratory estimated US data centers used about 17 billion gallons of water directly in 2023, while closed-loop liquid designs can reduce evaporative use to near zero.

The LBNL estimate also found far larger indirect water use, about 211 billion gallons, from generating the electricity data centers consumed, and projected that direct use by hyperscale facilities could reach 16 to 33 billion gallons a year by 2028. Company disclosures show the scale: Google reported over 6 billion gallons of data center water use in 2023, and Meta about 776 million gallons. The IEA notes that a 100 MW data center can use up to about 2 million litres a day under evaporative designs. Our article on data center water, land and jobs statistics has more local detail.

This creates a trade-off. Evaporative cooling saves electricity but consumes water; closed-loop systems save water but can use more electricity on hot days. Microsoft announced in December 2024 that all new data center designs from August 2024 would use chip-level cooling with a closed loop that does not evaporate water for cooling, avoiding more than 125 million litres a year per facility. Pilots in Phoenix, Arizona, and Mount Pleasant, Wisconsin, were expected to operate from 2026, with broader rollout from late 2027. Microsoft reported its average water usage effectiveness had fallen to 0.30 litres per kWh, a 39% improvement since 2021.

Regulators are paying attention. The EU’s data center reporting rules require sites above 500 kW of IT load to disclose water input and water usage effectiveness alongside PUE, and water stress has become a frequent point of local opposition. Our guide to data center sustainability covers corporate water-positive pledges in depth.

Frequently Asked Questions