Formula · 2 min read
Heat capture split (liquid vs air)
Direct-to-chip liquid cooling captures 60-80% of server heat, not all of it: Pliquid = capture × PIT. A 40 kW rack at 75% capture sends 30 kW to liquid and 10 kW to air — and the air share still needs the room to remove it.
Formula
Cold plates sit only on the big chips (CPUs, GPUs, accelerators). Memory, drives, VRMs and power supplies keep air-cooling themselves. Full immersion is the architecture that captures ~100% — at the cost of different hardware, fluids and service model (comparison).
Worked example
The 10 kW air share is why an L2A cabinet still moves substantial air, and why room AC sizing uses the full IT load, not just the liquid part.
Try it
Common mistakes
- Assuming 100% capture. Only immersion gets close. Sizing flow, CDU capacity and room AC on 100% overstates the liquid side and — worse — hides the air share that still heats the room.
- Using TDP sums as measured load. Thermal design power is a worst-case rating; typical draw is lower and varies with workload (TDP → rack kW).
- Forgetting the air share entirely. The 20-40% that stays on air still needs cabinet airflow, room AC capacity and an air path out of the room.
Related
Coolant flow rate (uses the liquid share) · Room AC load (uses the total) · Rack builder applies per-server capture
Sources
60-80% direct-to-chip capture is the range used across vendor documentation and this site's data layer; the rack builder uses 75-80% for GPU nodes and 50% for CPU nodes.