Liquid Cooling

Liquid cooling refers to any thermal strategy that uses a fluid loop — water, dielectric fluid, refrigerant — to carry heat away from IT equipment. It includes rear-door heat exchangers, in-row chilled-water units, direct-to-chip cold plates, single-phase immersion, and two-phase immersion.

Liquid cooling is increasingly required for AI training and inference workloads at the upper end of GPU and accelerator density. Air-cooled facilities top out around 30 kW per rack in practice. Direct-to-chip designs comfortably support 100–150 kW per rack. Two-phase immersion has demonstrated 200+ kW per rack in operating environments.

Choosing a liquid cooling strategy is an engineering, operational, and commercial decision in equal measure. Direct-to-chip is the most common AI retrofit pathway because it preserves rack and IT form factors. Immersion is the highest-density option but requires a structurally and operationally different hall. Both are common in 2025+ engineering scopes.

Operationally, liquid cooling shifts the failure modes operators have to manage — from primarily airflow and humidity to primarily leak detection, fluid chemistry, secondary loop pressures, and the manifolds that connect IT to the cooling system. Application-aware operations and skill-mapped staffing are not optional in a liquid-cooled environment.