Liquid cooling

AI racks pack so much power into a small space that fans and cold air can no longer carry the heat away, so water-based cooling is piped right up to the chips.

Layer 1 · Physical & Energy · Build and run time · reviewed 2026-10-30

Read first: Power & cooling for AI data centres.

D1 · StoryFor young children and total beginners

Which cools a hot pan faster: blowing on it, or running it under water?

AI chips get very hot, and they are squeezed close together.

Blowing air over them is like blowing on hot soup. It works for a little heat, but not for a lot.

So builders run cool water through metal plates that touch the chips, and the water carries the heat away.

Like: Cooling a hot pan under the tap instead of blowing on it.

D2 · PictureFor curious beginners

Nearly all the electricity a chip uses becomes heat. Air cooling pushes cold air through the rack with fans, but air holds little heat, so very dense racks would need more airflow than is practical.

Water carries heat far better than air. In direct-to-chip cooling, a liquid flows through a cold plate pressed onto each processor, then out to a heat exchanger. In immersion cooling, whole servers sit in a non-conductive fluid.

The newest AI systems are designed as liquid-cooled racks from the start, so a building that can only do air cooling may not be able to host them. Most liquid-cooled racks still use some air for parts such as power supplies.

Like: Cooling a hot pan under the tap instead of blowing on it.
Where it breaks: Tap water runs once and goes down the drain. Data-centre loops are usually closed: the same liquid circulates, and the heat still has to be dumped somewhere outside, by cooling towers or dry coolers, which may use water of their own.

D5 · FrontierFor experts

D2 presents liquid cooling as the fix for dense racks. Which liquid, at what temperature, and where the heat goes are all still open.

Water temperature. Warmer facility water lets operators run chillers less and reuse more heat, but leaves less margin for hot chips. Industry water classes now carry their upper temperature limit in their names (US DOE, 2024).

Two-phase fluids. Two-phase immersion relies on engineered fluorinated fluids, and a major maker announced in 2022 that it would leave all PFAS manufacturing, including fluorinated fluids, by the end of 2025. Fluid supply and chemical regulation shape which approach can scale (3M, 2022).

Waste heat. EU rules expect data centres above 1 MW to use their waste heat or other heat recovery unless they show it is not feasible. Warm-water heat is low grade and needs a heat user nearby (Directive (EU) 2023/1791).

Water versus energy. Liquid loops run warm enough for dry coolers, which save water; paired with evaporative economisers they save more energy but use substantial water (Berkeley Lab, 2024).

Examples as of 2026-10-30

Direct-to-chip cold plates in NVIDIA GB200 and GB300 NVL72 racks
Coolant distribution units from Vertiv and Schneider Electric (Motivair)
Direct-to-chip cooling units from LiquidStack (now part of Trane Technologies)

Sources

How it connects

Liquid cooling removes the heat for Power & cooling for AI data centres.