Reference

Anatomy of a liquid-cooled AI loop.

From facility water to silicon die — every interface that matters, with the temperatures, components, and failure modes we plan for.

Supply manifold

Cold plates

Return loop

CDU interface

01
Facility supply
18–32°C
Treated water from the chiller plant or dry cooler delivered to the data hall through insulated risers.
02
CDU primary loop
32–45°C
Coolant distribution unit exchanges heat between facility water and the closed secondary loop. Pumps, filters, and pressure control live here.
03
Manifold supply
40–45°C
Brass or stainless manifolds distribute warm coolant to each rack via vertical drops with isolation valves.
04
Quick-disconnect
—
Dripless blind-mate QDs let racks hot-swap without coolant loss. The single most failure-prone joint in the loop.
05
Cold plate
45 → 55°C
Microchannel plates clamped to GPU and CPU dies. 1–2L/min per plate, pulling 700–1200W per device.
06
Return manifold
54–58°C
Heated return collects from all racks and feeds back to the CDU — and increasingly into heat-reuse loops.

Flow schematic

Two loops, one heat path.

18°C

Facility

32°C

CDU

45°C

Secondary

55°C

Return

32°C

CDU

ΔT 10°C secondary · 5–8 m³/h per CDU · N+1 pump redundancy

Critical leak response, pump diagnostics, and pressure stabilization. Engineers en route in under 4 hours.