Selecting Cooling Architectures for High-Density AI Data Centres

Rising power concentration in artificial intelligence servers has made cooling architecture a primary design constraint for high-density data centres. This paper compares recent peer-reviewed studies of air cooling, direct-to-chip cold plates, single- and two-phase immersion, and their environmental effects. The findings show that liquid cooling expands the thermal operating envelope of high-power hardware, although the size of the benefit depends on the measurement boundary. Direct-to-chip systems can retain a substantial air-side load, warmer coolant may increase server power, and immersion performance varies with flow distribution, fluid properties, and hardware design. Facility-level outcomes also depend on climate, electricity supply, water source, coolant production, and equipment lifetime. Cooling options should therefore be compared under a common workload and system boundary. Optimised air remains suitable for moderate-density or legacy facilities, direct-to-chip cooling provides a practical route for many high-density racks, and immersion is best suited to cases where exceptional heat density and purpose-built infrastructure justify changes to hardware and maintenance.