Liquid cooling vs air cooling

Liquid cooling moves heat off a chip through a coolant loop instead of blowing air across a heatsink. In a direct-to-chip design, a cold plate replaces the heatsink on each GPU or CPU, and coolant runs through channels in that plate. A coolant distribution unit then pushes the heat into a facility water loop or a rear-door heat exchanger. Air cooling is the older default: fans pull air over finned heatsinks and the room's air handlers take it from there.

Power density is the dividing line. Air handles chips up to roughly 400 to 500 W and racks in the 15 to 20 kW range before airflow and fan noise stop being practical. Current accelerators pass that on their own, since an H100 SXM draws up to 700 W and a B200 close to 1,000 W. Nvidia's GB200 NVL72 ships liquid-cooled by design at around 120 kW in a single rack.

Air is cheaper, simpler to service, and fine for CPU-only or single-GPU machines. Past roughly 40 kW per rack the facility side has to come first. Piping, a CDU, and a heat rejection path all need to exist before the hardware lands.

Sources

Last verified August 29, 2026.