Applications

AI training and inference halls

Eight-way accelerator sleds at sustained utilisation put 100 to 140 kW into a single rack. Air cannot take that out of the room at any sensible fan power, and it cannot take it off the die at all above roughly 700 W per package. Direct-to-chip cooling moves 70 to 80 per cent of the load into water and leaves the rest to a much smaller air path.

The design question stops being "can we cool it" and becomes "what is our supply water temperature, and how much of the year can we reject heat without mechanical cooling".

High-performance computing and research

Research systems live with their cooling for a decade and are measured on total energy, not peak performance. Warm-water loops running at 40 °C supply let dry coolers carry the load for most of the year in most European and North American climates, and make heat recovery a realistic conversation rather than a slide.

Colocation retrofit

Most buildings that now need liquid cooling were not designed for it. A sidecar CDU takes facility water — or in a liquid-to-air configuration, no facility water at all — and serves a single row without touching the building's chilled water plant.

It is not the most efficient answer. It is often the only one that can be delivered inside a live building on an operating floor, and it can be removed again.

Industrial and edge deployments

Compute at the edge of a plant or a substation rarely has a clean room to sit in. Sealed liquid loops keep dust, salt and swarf out of the heat path, and let the enclosure be closed rather than filtered.