Immersion cooling hits 100+ kW/rack with PUE as low as 1.03. How hyperscalers are rebuilding data centers from the tank up in 2026. Full breakdown.
Why Air Cooling Runs Out of Room
Air has a hard ceiling as a heat-removal medium. Once a rack draws enough power, the volume of chilled air and the fan energy needed to move it stop scaling economically, and hot spots form faster than the room can clear them. Dense accelerator hardware pushes past that ceiling, which is why racks that once sat comfortably in the low tens of kilowatts are now targeting 100+ kW. At that density, the question stops being how to cool a room and becomes how to pull heat directly off the silicon.
Immersion cooling answers by submerging the hardware in a dielectric fluid that carries heat away through direct contact. Liquid conducts heat far more effectively than air, so the same rack can dissipate far more power without the acoustic and airflow problems that come with packing in more fans.
What a PUE of 1.03 Actually Means
Power Usage Effectiveness measures how much of a facility's total draw reaches the compute versus how much is spent on overhead like cooling, power conversion, and lighting. A PUE of 1.03 means almost every watt entering the building does useful work, with only a thin slice lost to running the facility itself. Traditional air-cooled halls spend a meaningful fraction of their energy just moving and chilling air; immersion collapses that overhead by removing the fan load and letting the fluid loop run at higher, more efficient temperatures.
The higher operating temperatures matter as much as the fluid itself. Because the coolant can reject heat without aggressive refrigeration, more sites can lean on ambient or free cooling for much of the year, which is what drives the overhead down toward that thin margin.
Rebuilding From the Tank Up
Retrofitting immersion into a room designed for airflow only goes so far, which is why the 2026 shift is a ground-up redesign rather than a bolt-on. When the tank becomes the primary unit of the facility, the assumptions underneath it change: the floor supports fluid weight instead of raised-floor airflow, plumbing replaces ducting, and heat leaves the building as warm liquid instead of exhaust air.
- Structure: floors and racks rated for the weight of fluid-filled tanks, not just hardware.
- Plumbing over ducting: coolant distribution loops and heat exchangers take the place of CRAC units and air handlers.
- Serviceability: hardware is lifted from and lowered into fluid, so maintenance workflows, connectors, and cabling are designed for submersion.
- Materials: components and cabling chosen for long-term compatibility with the dielectric fluid.
Planning the Transition
For teams weighing the move, the honest tradeoff is higher design and facility complexity in exchange for density and efficiency that air simply cannot reach. The practical path is to size the fluid loop and heat rejection for the density you actually intend to run, confirm hardware and cabling are rated for immersion, and design service procedures before the first tank goes in rather than after. Recovering warm output heat for reuse is worth evaluating early, since the loop already delivers it in a usable form.
The clearest signal to start planning is rack density itself. When target power per rack climbs toward the point where air stops keeping up, the tank stops being an experiment and becomes the cheaper long-term architecture.