Liquid cooling is usually sold on efficiency, and the claim is real, but it is not where most people think it is. Being precise about it makes the business case defensible and stops the design being tuned for the wrong variable.

The saving is not the pumps. Pumping water costs real energy — a secondary loop moving 170 L/min against 2 bar is a meaningful parasitic load, and a CDU's pumps run continuously. What liquid cooling removes is far larger: the server fans that were moving air through a 90 kW cabinet, and the chiller compressors that were producing cold air for them to move. Fan power in a dense air-cooled rack can run to 10–15 % of IT load, and it grows non-linearly as density rises, because fan power scales roughly with the cube of flow rate. Removing most of it is a bigger win than the pumps give back.

The real prize is chillerless hours. A loop that runs at 32–45 °C supply can reject heat through dry coolers, without a compressor, whenever ambient wet-bulb allows. In much of the world that is most of the year. Every hour spent on free cooling instead of mechanical cooling is the entire chiller plant's draw removed from the meter. This is why supply temperature is the number worth arguing about at design time: it does not change how well the GPUs run, and it changes the operating bill enormously.

Efficiency metrics get slippery here, so be careful what you promise. PUE improves because both the fan and the chiller terms shrink. But PUE is a ratio, and the same design change that improves it also raises IT load per rack, so the absolute facility draw of a liquid-cooled hall is higher than the air-cooled one it replaced. A PUE of 1.15 on 20 MW is not a saving over 1.4 on 5 MW. Quote both numbers, or quote energy per unit of useful work instead — which for an AI facility is the honest metric anyway, and which is why the throughput measurements in the acceptance chapter matter to the finance conversation as much as to the engineering one.

Heat reuse is real but not free. Return water at 45–50 °C is a genuinely usable heat source for district heating, greenhouses, or preheating a building's own domestic hot water, and several operators do it. The catch is that reuse imposes a temperature floor on your return water, which reduces your freedom to widen ΔT, which pushes flow back up. It is a facility-level trade with real engineering consequences upstream, not a bolt-on. Decide early or not at all.