Option 2 · direct-to-chip liquid cooling
The hall is quiet. Almost no air is moving, because the heat never enters the room: it goes from die to cold plate to pipe, and only surfaces again as a warm updraft above the outdoor fluid coolers.
Coolant runs through machined channels in a plate clamped to the die. Heat conducts out of the silicon and into the water within centimetres of where it was made.
Processor → cold plate → coolant
Flexible hoses tie each server to the vertical rack manifolds. Warm return water from thousands of processors gathers into one closed technology loop.
CDU → manifold → server → manifold → CDU
A plate heat exchanger separates the technology loop from the facility loop. The two never mix; only heat crosses the metal. Several CDUs per hall keep the design modular.
TCS → heat exchanger → FWS
Facility water runs at roughly 35–40 °C supply. That is warm enough to dump straight to ambient air for most of the year, so the compressors mostly sit idle.
FWS → dry / hybrid cooler → air
Why the warm water matters. An air-cooled hall needs supply air in the low twenties, which means mechanical cooling whenever the outdoor wet bulb climbs. A cold plate is happy with 40 °C water, and 40 °C water can be cooled by simply blowing outside air over a coil. The chiller becomes emergency equipment rather than the core of the plant. Temperatures and the 85/15 split shown here are illustrative and depend on the server architecture.