AI data center — OneCompute

30–132 kW

AI rack-density range, per rack

<1.2

Target PUE with direct liquid cooling

2700 W

Power draw of a current top-tier accelerator

DLC

Direct liquid cooling — cold-plate and immersion

The density problem

Accelerators changed the power envelope.

Accelerator power consumption has climbed steeply generation over generation — roughly 400 W to 2700 W across recent generations — and rack densities have followed. Legacy data centers sit at 6–8 kW per rack; retrofitting half of them to AI workloads requires at least 30 kW per rack, and frontier training racks now reach above 130 kW. Building for that envelope is a design discipline, not an upgrade.

6 kW

Legacy rack density — never designed for accelerated draw.

30 kW

The minimum to retrofit a rack for AI workloads.

100 kW+

Frontier training racks, with liquid cooling.

AIDC design

Cooling and density scale together.

As the server profile shifts from air to liquid, rack density climbs — and the facility has to be built for it from the start.

Air-cooled node

Air cooling, roughly 10 kW per server — about 20 kW per rack with two servers.

Liquid-cooled node

Direct liquid cooling plus air conditioning — about 50 kW per rack with four servers.

Integrated rack

Direct liquid cooling — up to 132 kW per rack.

Multi-mode cooling

Air, cold-plate liquid, and immersion — matched to the workload. Cold-plate carries the highest share of modern AI deployments and is the path to a sub-1.2 PUE.