Hydronic loop design for AI data centers

A hydronic loop for a 100 kW+ rack is a mechanical system where every joint, elbow and valve is a candidate for a leak that kills a million dollars of GPU time. This is the blueprint we use.

Design principles that survive contact with reality

• Constant ΔT beats constant flow. Design pumps to modulate, not to run flat out.
• Every rack drop must be isolatable without taking the row down.
• No mixed metals without dielectric unions. Full stop.
• Air is the enemy. Design in air separators and vent points at every high spot.
• Expansion tanks sized for full ΔT swing plus 10% for coolant makeup.

Pipe and pump sizing

Target velocity 1.5–2.5 m/s in steel risers, under 1.5 m/s in copper drops. Pipe friction losses under 250 Pa/m keep the pump curve honest. Undersized risers are the number-one root cause of ΔP alarms we're called to.

Expansion, makeup and air management

• Bladder-type expansion tank sized for the full 10–40 °C swing.
• Automatic makeup with dedicated de-ionised water skid, isolated by check valve.
• Coalescing air separator at the highest point of the loop.
• Manual vent points at every riser cap.

Controls and setpoints

A modern CDU PLC controls: secondary supply temp (setpoint 35 °C ±0.5), pump VFDs to hold ΔP, flow limits per rack, leak-rope zones, and CDU-to-CDU failover logic. Publish everything to BMS via BACnet/IP or Modbus TCP.

Commissioning gates

1. Hydrostatic pressure test at 1.5× MAWP for 4 h.
2. Chemical clean and passivation.
3. Fill with de-aerated fluid, run air separator 24 h.
4. Balance flow to every rack drop within ±5%.
5. Load bank test to 110% design load.
6. Sign-off leak-detection matrix with facility ops.