Immersion cooling submerges entire servers in tanks of dielectric fluid that doesn't conduct electricity. The fluid absorbs heat from all components simultaneously. Two approaches exist: single-phase where fluid stays liquid, and two-phase where fluid boils to absorb more heat. Two-phase handles the highest power densities, over 1600W [[Thermal Design Power - TDP]] per chip. Dielectric fluids are engineered to be non-conductive and non-flammable. Common choices include 3M Novec fluids and mineral oils. The fluid circulates through heat exchangers where external water or air cooling dissipates the captured heat. Uniform cooling across all components. GPUs, memory, power supplies, network cards all benefit. Captures more total waste heat than [[Direct-to-Chip Cooling]], enabling higher potential for energy recovery and reuse. Downsides: requires specialized servers and tanks. Maintenance means lifting components out of fluid, draining is time-consuming. Initial deployment is more complex than retrofitting existing racks with DTC systems. Best for maximum density deployments. Bitcoin mining operations pioneered adoption because they run at constant high load. Now [[Hyperscalers]] are testing for AI training clusters where power density exceeds 100kW per rack. Companies like Submer, GRC, and Asperitas lead the market. Immersion works where nothing else does, but you pay in operational complexity. --- #deeptech #datacenters