Air cooling or liquid cooling for AI?

If your soup is too hot, you can blow on it. If you are in a hurry, you can stand the pan in cold water. Both work. Water simply carries heat away much more effectively than air does.

Data centers face the same choice with their chips. This piece explains how air and liquid cooling work, what each suits, and why denser AI machines are pushing many sites toward liquid.

How air cooling works

Most data centers built in the past were cooled by air. Fans pull cool air through the front of each server, across the chips, and push warm air out of the back.

Racks are usually arranged so the warm air from one row does not feed straight into the next. The warm air is then cooled again, by air handlers, chillers or outside air, and sent back round.

Air cooling is well understood. Equipment is widely available, staff know how to run it, and servers can be swapped without touching any pipes.

Where air reaches its limit

Air does not hold much heat. To remove more heat, you have to move more air, faster. That means bigger fans, more noise, more energy and more space for airflow.

AI racks pack more power into the same footprint than most older computing. (We explain why AI draws so much power in Why AI uses so much electricity, in plain words.) At some point, there is simply not enough room to push enough air through a dense rack. The chips then run hot and slow themselves down.

The exact point depends on the hardware and the building. But as AI racks get denser, more of them cross it.

How liquid cooling works

Liquid cooling brings a fluid much closer to the heat. There are two main approaches.

Direct-to-chip cooling. A metal plate sits on top of each hot chip. Liquid flows through the plate, picks up heat, and carries it away through pipes. The rest of the server may still use some air.

Immersion cooling. Whole servers sit in a bath of a special fluid that does not conduct electricity. The fluid absorbs heat from every part, then is cooled and circulated.

In both cases, the warm liquid goes to a heat exchanger. There the heat passes to a second loop that takes it outside, or to a place it can be reused. (See What can be done with a data center's waste heat.)

What each one suits

Neither method is right for everything. They suit different jobs.

Many sites use both. Liquid handles the hottest chips, while air takes care of everything else in the room.

What changes for the site

Moving to liquid is not just a change of equipment. It touches the whole design.

Plumbing. Pipes, pumps and heat exchangers have to be planned, installed and maintained.

Leak handling. Systems need detection, isolation valves and procedures in case something drips.

Skills. Staff need to handle fluids as well as electronics.

Hardware. Servers must be built or fitted for the chosen method. Not every machine works with every system.

Water use. Liquid in a closed loop is not consumed. Whether the site uses water depends on how the heat finally leaves the building.

This is one reason standard units help. If cooling is designed into the unit, the site does not have to be rebuilt each time hardware changes. (See Why standard units matter and What is a modular data center?)

Questions to ask

If you are buying capacity for AI, or hosting it, these questions show whether cooling will keep up.

1. Which cooling method does the site use, and for which racks?
2. What is the highest rack density it can cool today?
3. Can it add liquid cooling later without major rebuilding?
4. How are leaks detected and contained?
5. Does the final heat rejection use water, and how much?
6. Is any heat reused, and where does it go?

What we are building

We're building cooling around the density AI hardware needs.

We're building standardized compute close to power, connectivity and demand, joined by one programmable network for AI inference. The network is designed to place workloads by policy, on latency, data residency, model and load, and every workload is designed to generate an auditable record of commitment, placement and delivery.

If you need capacity for dense AI hardware, use Request capacity below.