Scenarios — DataCool Technologies

— SCENARIO 01

Live-hall retrofit

APPROACH
DLC + REAR-DOOR HX

Adding a dense inference cluster to an air-cooled hall.

The situation

A Tier-III hall built for ~7 kW racks needs to host a new GPU inference cluster at 30–40 kW per rack. The room can’t go offline, and full immersion isn’t on the table.

How we’d approach it

Direct-to-chip cold plates on the GPU servers, in-row CDUs, and rear-door heat exchangers on adjacent racks to take the residual air load. A dedicated technology cooling loop tied into the existing chilled-water plant, filled and commissioned without dropping the surrounding rows.

What it targets

A PUE in the 1.15–1.2 range for the new cluster, no whitespace downtime during the cut-in, and lower fan noise as air load comes off the room.

Rack density

30–40kW

Target PUE

≤1.2

Whitespace downtime

0hrs

Heat to liquid

~80%

— SCENARIO 02

Greenfield AI pod

APPROACH
SINGLE-PHASE IMMERSION

A greenfield training pod in a tight footprint.

The situation

A new India-based training pod needs on the order of a megawatt of GPU compute in a small envelope, at the lowest practical PUE. A conventional air design would need several times the floor plate.

How we’d approach it

Single-phase immersion tanks populated with current GPU servers from a vendor-approved hardware list, single-phase dielectric coolant supplied and commissioned in country, and a CDU loop with redundant pumps rejecting heat to roof-mounted dry coolers.

What it targets

Around 100+ kW per tank, no CRAC units on the whitespace, and a much smaller floor plate than the equivalent air-cooled build.

Density / tank

100+kW

Target PUE

~1.1

Floor saved

~65%

Air handlers

How we’d approach your build.

Adding a dense inference cluster to an air-cooled hall.

A greenfield training pod in a tight footprint.

Let’s scope the scenario closest to yours.