AppsGuidesAI DictionarySmall FixesOn RecordAboutFollow the lab

Liquid cooling

Liquid cooling carries heat away from AI chips with water or special fluids instead of air, because the newest chips run too hot for fans alone.

Chips and scaleCurious

In plain words

Every AI chip turns the electricity it uses into heat. For decades data centers blew cold air over their computers. But AI racks now pack so many power-hungry GPUs together that air cannot carry the heat away fast enough.

Liquid is far better than air at soaking up heat. So the newest AI servers pump a coolant, often water or water mixed with glycol, through metal plates pressed onto each chip. The warmed liquid flows away, hands its heat to the building's cooling system, and comes back cool, in a closed loop.

The numbers explain the shift. A typical server rack draws around 20 kilowatts. A rack of 72 of Nvidia's newest GPUs draws about 120 to 140 kilowatts, six or seven times more in the same floor space.

Everyday example

A cold bottle after a hot match

After a cricket match in May heat, a fan blowing warm air barely helps. Pressing a bottle of cold water against your neck works much faster, because water pulls heat away quicker than air. Now imagine the bottle is a pipe with cold water always flowing through it, carrying the heat off and coming back cold. That is direct-to-chip cooling: a cold plate on each chip, with fresh coolant always arriving.

Where the example breaks: Sweat cools you by evaporating, so the water is used up. A closed loop inside the servers reuses the same liquid, although the building may still use water outside to get rid of the heat.

Go deeper

How it actually works

Liquid cooling is a family of methods. They differ in how close the liquid gets to the chip.

  1. Direct-to-chip cold plates

    A metal block with tiny channels sits on top of each GPU and processor. Coolant runs through it and carries the heat to a coolant distribution unit (CDU), which passes it to a separate building loop through a heat exchanger. Memory and other parts may still need some air.

  2. Immersion

    Whole servers sit in a tank of fluid that does not conduct electricity. In single-phase immersion the fluid stays liquid and is pumped to a heat exchanger. In two-phase immersion it boils on the hot chips, and the vapour turns back to liquid on cold coils above.

  3. Rear-door heat exchangers

    A halfway step for air-cooled racks: a radiator on the back of the rack catches the hot air as it leaves. Vertiv, a cooling maker, rates passive doors for roughly 5 to 25 kW per rack and active ones up to about 50 kW.

  4. Getting heat out of the building

    The warm liquid still needs somewhere to dump its heat: dry coolers, chillers or evaporative cooling towers. Towers consume water: Schneider Electric cites an estimate that a 100 MW site cooled this way can use about 2 million litres a day. Hot, humid climates such as much of India's usually still need chillers too.

  5. An old idea, back again

    Supercomputers used liquid long before AI. The Cray-2 of 1985 had its circuit boards immersed in a non-conductive liquid called Fluorinert. AI has brought the idea into ordinary data centers.

Typical server rack
About 20 kW; racks of Nvidia's earlier H100 GPUs, over 40 kW (Tom's Hardware, 2025)
Nvidia GB200 and GB300 NVL72
About 120 to 140 kW per rack, cooled by cold plates on the chips in a closed loop
Early immersion cooling
Cray-2 supercomputer, 1985, with its boards in Fluorinert
India
The IndiaAI Mission's GPU tender asks for a power usage effectiveness (PUE) below 1.35, against an Indian average of about 1.5 to 1.6 (Schneider Electric, July 2026)

PUE is a data center's total power divided by the power its computers use. 1.0 would mean no energy at all goes on cooling and other overheads.

Questions people ask

Why do AI data centers need liquid cooling?

Because AI chips use so much power, packed so tightly, that fans cannot remove the heat fast enough. A rack of Nvidia's newest GPUs draws about 120 to 140 kW, against about 20 kW for a typical rack. Liquid carries that heat away far more effectively.

Does liquid cooling use more water?

Not inside the servers: the coolant there runs in a closed loop and is reused. Water use depends on how the building gets rid of the heat. Evaporative towers consume water, while liquid-cooled systems can run on warmer water, which reduces the need for power-hungry chillers.

What if the liquid leaks onto the chips?

Leaks are a real worry, so systems use sealed loops and leak detection. Water and water-glycol mixes carry heat best. Some systems use special fluids that do not conduct electricity instead: they carry less heat than water but cause less damage if they leak.

What is the difference between direct-to-chip and immersion cooling?

Direct-to-chip cooling pipes liquid through a plate on each chip, and the rest of the server stays in air. Immersion cooling puts the whole server in a tank of non-conductive fluid, which can remove nearly all of its heat.

Was this clear?

1 is confusing, 5 is crystal clear. One tap, no account needed.

Spotted something wrong, or know a better example? Suggest a change

Further reading

Prepared by Mundhra Labs. Reviewed by Aditya Mundhra. Checked November 2026.