Immersion Cooling Technologies in High-Density Computing: Review and Future Perspectives

Air cooling is being pushed to its limits in data centres, with racks now pushing tens of kilowatts and localised areas of the accelerator reaching at 100 W cm-2. The architecture most frequently suggested to replace it is immersion cooling, where electronics are in direct contact with a dielectric liquid. Still, the literature backing this replacement is very disjointed—results are reported at varying spatial scales, reported in different system boundaries, and with incomplete property and hydraulic data, making direct comparison risky. This review does not aim to be yet another descriptive survey, but to tackle this fragmentation. 90 studies published from 2015 to 2026 were identified via a documented search protocol, screened for explicit criteria and analysed within five levels of chip, package, server, tank, and facility scales. The comparative performance of single-phase (SPIC) and two-phase (TPIC) immersion cooling is systematically compared using twelve engineering criteria and the quantitative data extracted are integrated into thirteen comparison tables between coolant properties, heat-transfer coefficients, thermal resistances, pressure drops, enhancement penalties, energy metrics with explicit boundary tags, cost and environmental indicators, and long-duration test evidence. Three findings emerge. First, the reported performance envelopes of SPIC and TPIC overlap significantly more than is suggested by their mechanisms: Forced SPIC has achieved heat-transfer coefficients of 0.4-2 kW m-2 K-1, and jet or locally actuated SPIC has achieved values between 2 and 15.3 kW m-2 K-1, which spans much of the range of heat-transfer coefficient for plain-surface nucleate boiling from 1.5 to 12 kW m-2 K-1; thus, the choice of architecture depends less on the heat-transfer coefficient achieved, and more on the critical heat flux margin, vapour management and complexity of operation. Second, apparent contradictions in the literature can be traced to five identifiable sources: declared system boundaries were different, unreported coolant properties, bulk flow rate as a substitute for local availability of coolant, mismatched scales, and steady uniform heating as a substitute for realistic workloads, and once these issues are controlled, the disagreements are largely eliminated. Third, the evidence base is very unbalanced: what is known about thermal performance—measured from the chip to the server—is well developed, but multi-year cooling ageing, material compatibility, and facility-scale demonstration and standardised test protocols are relatively limited or non-existent—this is the current limiting constraint for deployment, not thermal performance. On this basis, we suggest a prioritised research agenda and minimum reporting requirements for future immersion-cooling studies.