liquid cooling
A cooling method that uses water or another liquid instead of air to dissipate heat from computer chips, especially those running AI servers.
In plain terms
The chips that power AI consume enormous amounts of electricity, and almost all of it turns into heat. A single cabinet of AI servers can produce more than 100 kilowatts of heat, as much as fifty to sixty electric heaters running at full power. That much heat can no longer be blown away with a fan.
Liquid cooling works like a car engine's cooling system: fluid circulates next to the hot component, absorbs the heat, and carries it away to be dissipated. Water can absorb thousands of times more heat than air for the same volume, allowing more computing power to fit in a smaller space.
Compared to air cooling, the cooling itself consumes less electricity, the server room is generally quieter, and the hot water can even be used for district heating. In return, the setup is more complex and expensive: it requires pipes, pumps, heat exchangers, and leak detection.
How it works
- A metal cold plate is mounted on the chip, with tiny channels inside where liquid flows to absorb the chip's heat.
- Heated liquid is pumped out of the server rack to a distribution unit, which separates the closed internal circuit from the building's water system.
- In a heat exchanger, the internal circuit transfers heat to the building's cooling circuit, which releases it outdoors through cooling units on the roof or in the yard, or passes it on for heating purposes.
- The cooled liquid returns to the chips, and the cycle restarts; sensors continuously monitor temperature, pressure, and any leaks.
- Immersion cooling also exists: the entire server is submerged in a tank filled with a special non-conductive liquid, so every component is in direct contact with the coolant.
An everyday example
When you ask a chatbot a question or an AI agent works on a task for you, the answer is generated in a distant data center. Next to the chips working there, water is likely circulating in pipes to prevent overheating. You can see the same on a small scale at home: behind the transparent side panel of gaming computers, you often see the water cooling pipes and radiator.
Who is affected
A common misconception
Limits and risks
- Deployment is expensive and complex, and converting an existing server room designed for air cooling is often harder than building a new one.
- Leakage risk is real: water meeting electronics can cause serious damage, requiring continuous monitoring and careful maintenance.
- Cooling only removes heat; it does not eliminate power consumption. Liquid cooling makes data centers more efficient but does not make them green on its own.
- The heatsink mounted on the chip removes only the majority of the heat: the server's other components are typically still cooled by air, so fans and room cooling do not disappear entirely.
- In some immersion cooling variants, fluorine-containing (PFAS, so-called 'forever chemicals') fluids are used; due to their environmental impact, the largest manufacturer, 3M, has phased out production by the end of 2025. Servicing servers submerged in such fluid is also more cumbersome.
Similar, but not the same
- air cooling – Air cooling removes heat with cold air blown by fans, which is cheaper and simpler, but is no longer sufficient for the most powerful AI chips.
- immersion cooling – Immersion cooling is a type of liquid cooling in which a cold plate is not just attached to the chip, but the whole machine sits in liquid.
- air conditioning – Air conditioning cools the room's air, whereas liquid cooling removes heat directly from the heat-generating component.
Related concepts
data center, cooling technology, AI infrastructure
direct chip cooling, immersion cooling, two-phase cooling, rear-door heat exchanger
GPU, energy efficiency (PUE), water consumption, waste heat recovery, AI factory
A brief history
Liquid cooling is not an invention of the AI era: mainframe computers were already cooled with water in the 1960s, and the components of the Cray-2 supercomputer, introduced in 1985, were entirely immersed in coolant. Liquid cooling remained in supercomputers, but cheaper air cooling became standard in everyday data centers. Liquid returned to these facilities in the 2020s due to high-consumption AI chips, now on a massive scale.
Where things stand
At the end of 2026, direct chip cooling is already a basic requirement in new data centers built specifically for AI, because the most powerful server racks cannot be air-cooled. However, most existing traditional server rooms still use air cooling, and immersion cooling remains a solution used in a narrower range of cases.
In the news
- Data center power and cooling: 800 V DC, supercapacitors and liquid coolingOct 26, 2026 · Ipar
Also known as
liquid cooling · water cooling · liquid cooling · water cooling · direct-to-chip cooling · DLC (direct liquid cooling) · immersion cooling · immersion cooling
Sources
This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.