Direct-to-chip liquid cooling

Direct-to-chip (DLC) cooling puts a copper cold plate on every hot die and pipes a water-based coolant through micro-channels a few hundred microns wide. Done right, it silently carries 80% of the rack heat. Done wrong, it becomes the highest-consequence maintenance surface in the building.

How direct-to-chip cooling works

Coolant enters the rack manifold at ~35 °C, splits into parallel paths per node, and passes through cold plates bonded to GPUs, CPUs and NVSwitches. It exits at ~45 °C into a return manifold, then to the CDU where a brazed-plate heat exchanger transfers heat to the facility loop.

What to specify on a new AI cluster

• Coolant. Demineralised water + OEM inhibitor, or PG25 for freeze-risk sites.
• Quick-disconnects. Blind-mate dripless couplings, stainless body, EPDM seals rated for the inhibitor package.
• Manifold material. Stainless or brass — never mixed metals without dielectric unions.
• Flow. Sized for design ΔT at worst-case ambient primary supply.
• Filtration. 50 μm bag on fill, 5 μm inline, per-CDU.

Common failure modes we see

Micro-channel fouling

Bio-slime and inhibitor breakdown clog cold-plate channels. Symptoms: rising junction temps on a subset of GPUs while flow meters report normal. Fix: chemistry program with quarterly ORP + conductivity checks.

QD seep, not burst

The classic DLC failure is a slow seep at a quick-disconnect after a hot-swap. Detect with leak rope under every rack drop and per-rack dew-point sensors.

Air ingress

A poorly located expansion tank or missing air separator lets micro-bubbles recirculate and destroy pump seals within weeks.

Leak strategy

We design DLC halls with three concentric containment rings: cold-plate O-ring integrity, dripless QDs, and a tray-and-rope leak detection layer under each row. A Class-1 leak (any liquid on the floor) triggers a P0 dispatch — see our dispatch protocol.

FAQ

Does DLC void GPU warranties?

Not when installed with OEM-approved cold plates. All major hyperscalers now deploy OEM-integrated DLC on H100/GB200 nodes.

What percentage of heat still goes to air?

Typically 15–20% — PSUs, NICs, DIMMs and switches. Plan for RDHx or in-row CRAHs to catch the balance.