Modern AI GPUs generate 700W+ each in dense rack configurations, pushing air cooling past its practical limits. Direct-to-chip liquid cooling circulates coolant through cold plates on each GPU, removing heat more efficiently. Immersion cooling submerges entire servers in dielectric fluid. As racks exceed 100 kW densities, liquid and immersion cooling are becoming standard in new AI data centers, with the cooling infrastructure itself consuming 30–40% of total facility power in traditional air-cooled designs versus under 10% with advanced liquid cooling.