Direct-to-Chip Liquid Cooling for AI GPU Racks
Direct-to-chip liquid cooling puts a cold plate on the hottest silicon — the GPUs and CPUs — and carries the heat away in a liquid loop instead of air. It is the standard cooling approach for modern AI racks above roughly 40 kW, and mandatory for GB200/GB300-class systems.
What it is
Direct-to-chip (DLC) liquid cooling replaces the air heatsink on each processor with a cold plate. A coolant — treated water or a water/glycol mix — flows through the plate, absorbs the heat directly at the die, and carries it out of the rack through supply and return manifolds. Because liquid moves heat far more effectively than air, DLC can reject the 700–1000 W dissipated by a single modern GPU without the airflow a rack simply cannot provide at that density.
A DLC deployment is a loop, not a single box. The server cold plates connect to an in-rack manifold; the manifold connects to a coolant distribution unit (CDU) that isolates a clean "technical" loop from the facility water; and the facility loop rejects heat to a chiller or dry cooler outside. Direct-to-chip typically still leaves ~10–20% of rack heat (memory, NICs, PSUs) to be handled by air or a rear-door heat exchanger, so it is usually deployed as part of a hybrid design.
When you need it
Rack power density is what forces the move to direct-to-chip. Air cooling caps out around 15–20 kW/rack, and rear-door heat exchangers extend that to about 40 kW. Past that point, direct-to-chip is the only option that keeps modern accelerators in thermal spec.
- Rack density above ~40 kW — beyond what air or a rear-door exchanger can reject.
- GB200 / GB300 NVL72 racks (~120–140 kW) that ship with integrated direct-liquid cooling by design.
- B200 liquid GPU racks (~80 kW/rack) and dense H100/H200 HGX builds pushed past air limits.
- Workloads running GPUs at sustained high utilization, where TDP is continuous, not bursty.
- PUE / efficiency targets — DLC captures heat at the source, cutting fan energy and enabling warm-water cooling.
Quick facts & specs
Vertiv · Vertiv CDU 1.3 MW
| Type | Liquid-to-liquid CDU |
|---|---|
| Cooling capacity | 1,300 kW |
| Loops | Primary + secondary |
| Pump redundancy | N+1 |
| Approach temp | ≤ 3°C |
| Install | In-row, floor-mounted |
AI Rack Cooling Calculator
Size cooling capacity and heat rejection per rack, get an air-vs-liquid recommendation and CDU sizing.
Source it
Add the equipment to your quote list — pricing, availability and lead times are confirmed after you ask, usually within a business day.
Frequently asked
What is direct-to-chip liquid cooling?
Direct-to-chip (DLC) cooling mounts a liquid-cooled cold plate directly on each GPU and CPU. Coolant flows through the plate, absorbs heat at the die, and carries it out of the rack to a coolant distribution unit (CDU) and the facility loop. It is far more effective than air and is the standard approach for AI racks above ~40 kW.
When do I need direct-to-chip cooling instead of air or a rear-door exchanger?
Density decides it. Air handles up to ~15–20 kW/rack and a rear-door heat exchanger up to ~40 kW. Above 40 kW/rack — dense B200 racks, and all GB200/GB300 NVL72 systems — you need direct-to-chip liquid cooling. GB200-class racks ship with it integrated because no air-cooled configuration exists at ~120 kW.
Does direct-to-chip cooling remove all the heat from a rack?
No. Cold plates capture roughly 80–90% of rack heat — the GPUs and CPUs. The remaining 10–20% from memory, NICs and power supplies is still rejected by air or a rear-door heat exchanger, so DLC is usually deployed as a hybrid with a residual air path.
What equipment does a direct-to-chip deployment need?
DLC-ready servers with cold plates and in-rack manifolds, a coolant distribution unit (CDU) to isolate the technical loop, quick-disconnects and hoses, leak detection, and a facility heat-rejection plant (chiller or dry cooler). We scope and supply the full loop on one purchase order.