Liquid cooling for data centers

Air-cooled halls tap out around 30–40 kW per rack. NVIDIA GB200 NVL72 racks land at 120–132 kW. This guide is the compressed version of what our field crews wish every operator knew before their first liquid deployment.

Why liquid, why now

A 700 W GPU dumps roughly the same heat as three household space heaters running continuously. Multiply by 72 GPUs per rack and you have a small industrial process heat load in a single tile. Air cannot move that heat off a chip fast enough at acceptable fan power. Liquid, at ~3,500× the volumetric heat capacity of air, does — and it does it at temperatures warm enough to reject to a dry cooler for most of the year.

• Density. 60–150 kW/rack is now a standard AI cluster spec.
• Efficiency. Warm-water loops (W3/W4) enable free cooling year-round in most EU/NA sites.
• Heat reuse. 45 °C return water is directly usable for district heating and greenhouses.

Cooling architectures at a glance

Direct-to-chip (DLC)

Cold plates bolted to GPUs and CPUs, fed by manifolds and quick-disconnects. Captures 70–85% of rack heat into liquid; the rest goes to rear-door exchangers or in-row CRAHs.

Rear-door heat exchanger (RDHx)

A liquid-cooled coil bolted to the rear of the rack. Passive units handle up to ~35 kW; active (fan-assisted) units push to 70 kW. Best as a companion to DLC or as a first liquid step in a mixed-density hall.

Immersion

Single-phase dielectric fluid in an open bath. Very high density and efficiency, but a hard rewire of the operations model — no hot-swap in the traditional sense and dielectric handling becomes a chemistry program.

Anatomy of the secondary loop

A well-behaved secondary loop is a boring loop. Ours are engineered for constant ΔT (typically 10 K) and stable ΔP across all racks, with N+1 pumps, air separators, expansion tanks sized to the coolant temperature swing, and blind-mate couplings at every rack drop.

1. Primary chilled/warm water from the facility TCS.
2. CDU with brazed-plate HEX and N+1 variable-speed pumps.
3. Insulated stainless supply/return risers.
4. Manifold row with balancing valves and flow meters.
5. Quick-disconnects at the rack, cold plates on the silicon.

Coolants and water chemistry

The default fill is demineralised water with an inhibitor package or a 25% propylene glycol blend (PG25) where freeze protection is required. Skip the automotive coolants; the wrong inhibitor will attack copper or trigger biofouling inside cold plates whose channels are 200 μm wide.

Retrofit strategy for an air-cooled hall

We typically stage retrofits in three phases: (1) install RDHx on the two hottest rows to buy thermal headroom, (2) drop a CDU per row and pipe risers, (3) migrate rack-by-rack to DLC as tenants refresh silicon. Downtime windows stay under four hours per rack when the manifold work is pre-fabricated.

Frequently asked questions

What supply temperature should I design for?

For GB200-class clusters: 32–40 °C supply, 10 K ΔT, gets you free cooling almost everywhere and lands within NVIDIA's spec window.

How much water does this consume?

A closed loop consumes essentially none. Site WUE depends on whether you use adiabatic assist on the dry coolers during summer peaks — see our WUE playbook.

Do I still need CRAHs?

Yes, for the 10–20% of heat that stays airborne (PSUs, switches, NICs) and for humidity control. But you can shrink them dramatically.