Air or Liquid Cooling for an AI Data Hall

What the client needed

A colocation operator was fitting out a new 2 MW data hall in Brisbane for AI workloads, with tenants asking for racks of 100 kW and more. The base design was air-cooled with hot aisle containment.

The question they needed answered: Can air cooling handle AI rack densities in Brisbane's climate, and if not, which liquid cooling approach makes the most sense?

What we did

• Modelled the hall with CFD for the air-cooled base design at mixed rack densities, including 130 kW AI racks.
• Compared air cooling, direct-to-chip liquid cooling with air assist, and single-phase immersion on annual PUE, energy cost, water use and heat reuse, using Brisbane weather data.
• Assessed floor loading, tenant equipment compatibility, maintenance and fluid management.

What we found

Air cooling works up to about 30 to 40 kW per rack. At 130 kW, hot spots of 31 to 34°C appear at the dense racks and their neighbours. More air handling doesn't fix it, because the limit is moving enough air through the rack.

Direct-to-chip liquid cooling handles the density, saves about $450,000 a year in energy against air, and is supported by the major AI server makers.

Warm-water loops at 35 to 45°C let dry coolers reject heat for most of the year in Brisbane, so water use is close to zero.

Immersion is more efficient again, but narrows the range of tenants. It suits a single-tenant or owner-operated hall better than colocation.

Retrofitting liquid cooling into a live colocation hall later means piping work around operating tenants. Making the hall liquid-ready now is the cheaper path.

What we recommended

• Use direct-to-chip liquid cooling with air assist for the AI hall, designed around warm-water loops and dry coolers.
• Provide coolant distribution units and secondary piping to every row, so tenants can raise density without rework.
• Reserve space and connections for heat recovery to neighbouring users.