A single AI training rack can now draw more power than an entire row of servers did a few years ago. That shift has quietly rewritten the rules of data center design, and nowhere is the pressure sharper than in cooling. The accelerators driving today’s models throw off heat in concentrations that air simply cannot carry away fast enough, and the gap widens with every new chip generation.
For the teams building and operating these facilities, the question is no longer whether to move to liquid. It is how to do it without trading one set of problems for another.
Why density broke air cooling
Air cooling scales by moving more air faster, but there is a ceiling. Once a rack passes roughly 40 to 50 kilowatts, the fan power, airflow, and containment required to keep pace stop making economic sense. Hot spots appear, throttling kicks in, and the expensive silicon that justified the build cannot run at full tilt.
AI racks blow straight through that ceiling. Densities of 100 kilowatts and beyond are already in planning, and the roadmap points higher. At those levels, the heat has to be intercepted much closer to the source than a stream of chilled air can manage.
The efficiency case for a phase change
Liquid carries heat far better than air, but not all liquid cooling is equal. Single-phase systems circulate a coolant that stays liquid and warms as it absorbs heat. That works, but the amount of heat a warming liquid can carry is limited by how much its temperature is allowed to rise.
Two-phase cooling changes the math. Instead of just warming, the coolant boils. The energy absorbed in that liquid-to-vapor phase change is large, which means a two-phase system moves more heat with less coolant flow and lower pumping energy. For the densest racks, that headroom is the difference between keeping accelerators at full performance and leaving capacity on the table.
Serviceability is the hidden decider
Raw cooling capacity gets the headlines, but operators know the real test comes at 2 a.m. when a component fails. A cooling approach that requires draining a tank or dismantling a pod to reach a single card carries an operational cost that never shows up in a spec sheet.
This is where cooling that delivers coolant directly to a cold plate on the chip has a practical advantage. Technicians can service a rack much as they always have, and the facility keeps a form factor its staff already understand. The result is high density without a maintenance model that fights the people who have to run it.
Retrofit, not rebuild
Most operators are not building greenfield. They are trying to raise the density of facilities that already exist, on timelines that do not allow for a ground-up reconstruction. A self-contained, waterless loop that fits standard racks can be introduced incrementally, rack by rack, while the rest of the room keeps running. That path lowers both the risk and the capital shock of the transition.
Vendors in this space are converging on exactly that model. Accelsius, for example, builds a two-phase direct-to-chip cooling platform designed to hit very high rack densities while staying waterless and serviceable, so operators can scale from air to liquid without rearchitecting the building.
What to evaluate before you commit
If you are assessing cooling for a dense AI deployment, three questions cut through most of the noise. First, what rack density does the approach actually support at sustained load, not peak. Second, how does a technician service the hardware, and how long does the facility stay degraded during that work. Third, what does adoption look like in your existing rooms, and can it be phased.
The takeaway
The industry has already accepted that AI density and air cooling cannot coexist. The more useful decision now is choosing a liquid approach that delivers the thermal headroom of a phase change while keeping the facility serviceable and the migration incremental. Map your real densities and your maintenance model first, and let those numbers, rather than the spec sheet alone, point you to the right method.
