When Air Cooling Stops Being an Option

Liquid cooling is frequently presented as an efficiency upgrade. At the densities accelerated computing now runs at, that framing is misleading: past a certain point it is not an optimisation, it is the only way the hardware runs at its rated clocks.

The threshold

Around 40kW in a single rack, air cooling becomes impractical. The constraint is volumetric: removing that much heat requires moving more air than can physically pass through the cabinet at acceptable velocities and acoustic levels. Push past it and the temperature rise across the rack drives inlet temperatures on rear equipment out of specification.

What happens next is not a failure, which is why it is easy to miss. The hardware throttles. Clocks drop, sustained throughput falls, and the cluster quietly delivers less than it was specified to deliver, with nothing in the monitoring obviously broken.

The two options

Liquid-to-liquid requires facility water and rejects heat outside the room. It is the higher-capacity answer and the one that scales: a 300kW in-row CDU supports a row of eight-way nodes.

Liquid-to-air rejects heat into the room, which means the room has to be able to take it. It suits a single rack in a site without facility water, and it is a genuine solution rather than a compromise, provided the room-level cooling has been checked rather than assumed.

What it actually costs

The CDU is the visible cost and usually not the largest one. Coolant loop installation, leak detection, a maintenance regime and staff who know the fill and purge procedure all carry ongoing cost. Plan for the operational side, not just the capital purchase.

Against that, liquid cooling permits densities that would otherwise require more racks, more floor space and more distribution. On a large enough deployment the comparison frequently favours liquid on total cost, not only on capability.