How Do AI Data Centres Actually Stay Cool?

Link one: capturing the heat

Air. The traditional method blows cool air through the servers. To make it work well, racks are arranged so that cool air arrives on one side and hot exhaust leaves on the other, in alternating cold and hot aisles. The US Department of Energy's best practice guide, written by NREL, notes that strict hot aisle and cold aisle configurations can significantly increase the airside cooling capacity, because hot exhaust is kept from mixing back into the supply air.

Liquid at the chip. In direct-to-chip cooling, a metal cold plate sits on each processor and liquid flowing through it carries the heat away. The joint AI Data Center Energy Performance Framework from PNNL, ASHRAE and NEMA describes this as an inherently more efficient heat transfer medium than air, and advises matching the cooling design to the heat density of AI equipment rather than to older, lower-density baselines.

Both. Most current AI racks are hybrids. In NVIDIA's GB200 NVL72 systems, for example, the processors and main network chips are liquid cooled while the remaining parts are cooled by air.

Link two: carrying the heat out of the hall

Warm liquid from the racks flows to a heat exchanger, where it passes its heat to the building's main cooling loop without the two liquids mixing. Warm air is drawn back to cooling units in the hall. Either way, the heat now travels in the building's own water loop towards the plant that will release it.

Link three: releasing the heat outdoors

Free cooling. When the outside air is cool enough, the plant can reject heat with little or no mechanical refrigeration. The NREL guide says that most nights and during mild winter conditions, the lowest cost option is an air-side economiser, and that water-side free cooling suits climates with cool wet-bulb temperatures for many hours a year. A proper evaluation of the local climate comes first.

Cooling towers. A cooling tower sprays the warm water through moving air so that a small part of it evaporates. The US Department of Energy explains that the evaporative process extracts the heat and dissipates it to the surrounding atmosphere. This saves electricity, but the evaporated water is consumed, and the tower needs a steady supply of make-up water to replace it.

Dry coolers and chillers. Heat can also be released into the air through closed coils, like a car radiator, with chillers adding refrigeration on hot days. This uses little or no water at the site but more electricity in hot weather.

What liquid saves, and how water is counted

A June 2026 report from the IEA 4E Technology Collaboration Programme estimates energy savings in the region of 8% in servers and 30 to 40% at the facility level from liquid cooling, depending on the design. It also finds that use of liquid cooling is still low, held back by a lack of standardisation, high initial costs and concerns about long-term reliability. The same report notes that the common efficiency measure, Power Usage Effectiveness, systematically understates the gains from liquid cooling.

Liquid also opens a door to reuse. The PNNL, ASHRAE and NEMA framework notes that direct-to-chip and warm-water liquid loops provide higher-grade heat than air exhaust, which makes it more practical to put that heat to use nearby.

Because some cooling designs consume water, the industry tracks it with its own measure, Water Usage Effectiveness. The US Department of Energy describes it as how efficiently a facility uses water in relation to the energy used by the IT equipment, in litres per kWh.

What is not certain

• Savings from liquid cooling depend on the design, the climate and what it replaces. The IEA 4E figures are estimates across scenarios, not results from one facility.
• The best outdoor choice depends on local climate and water. This article does not assess any particular site.
• Cooling technology for AI is changing quickly, and new designs may change these trade-offs.

Sources and method

• Van Geet and Sickinger, NREL, for the US Department of Energy Federal Energy Management Program. Best Practices Guide for Energy-Efficient Data Center Design, revised July 2024. https://www.energy.gov/sites/default/files/2024-07/best-practice-guide-data-center-design_0.pdf
• PNNL, ASHRAE and NEMA, AI Data Center Energy Performance Framework: Energy and Thermal Efficiency. Undated web page, read 29 October 2026. https://www.ashrae.org/technical-resources/topics-and-initiatives/ai-data-center/energy-and-thermal-efficiency
• IEA 4E Technology Collaboration Programme, Electronic Devices and Networks Annex, Liquid Cooling in Data Centres, 22 June 2026. https://www.iea-4e.org/edna/publications/liquid-cooling-in-data-centres/
• US Department of Energy, Federal Energy Management Program, Cooling Water Efficiency Opportunities for Federal Data Centers, published 9 January 2019, updated 28 April 2026. https://www.energy.gov/cmei/femp/cooling-water-efficiency-opportunities-federal-data-centers
• NVIDIA, GB200 NVL Multi-Node Tuning Guide, Understanding Your Grace-Blackwell Systems, last updated 21 April 2025. https://docs.nvidia.com/multi-node-nvlink-systems/multi-node-tuning-guide/system.html

Method: each link was opened on 29 October 2026 and every figure was taken from the page cited beside it. No figure was estimated.