Why a 10 kW GPU server can't simply be cooled by more fans — air runs out of heat-carrying capacity long before GPU racks do, which is the real reason liquid cooling went from a niche HPC technique to the default architecture for AI infrastructure.
Ask why AI data centers are moving to liquid cooling and you'll often hear "because GPUs run hotter." That's true but incomplete — plenty of hot components have been air-cooled for decades. The real constraint is volumetric heat capacity: water carries roughly 3,500 times more heat per unit volume than air moving at the same velocity. A traditional CRAH/CRAC air-cooling system was designed around 5-10 kW/rack, where moving enough air through a rack is straightforward. A GPU training rack drawing 40-130+ kW needs to reject far more heat in the same physical footprint — and beyond a certain density, no amount of additional fan power and airflow can move it fast enough without impractical noise, power draw, and hot-aisle temperatures. Liquid cooling exists because water can carry that heat away in a fraction of the flow volume.
Air cooling doesn't fail because a GPU runs too hot for air to touch — it fails because moving enough air volume to carry away 40-130+ kW from a single rack footprint requires impractical fan power, noise levels, and pressure differentials, and produces hot-aisle temperatures that stress the rest of the equipment. Water absorbs roughly 3,500 times more heat per unit volume moved at the same velocity, so a liquid loop can remove the same heat load through much smaller pipes at far lower pumping energy than the equivalent air handling would cost in fan energy. That efficiency gap is also why liquid-cooled facilities tend to post better PUE — less energy spent moving the cooling medium around per watt of heat rejected.
In direct-to-chip liquid cooling — the dominant approach for AI racks today — coolant never contacts the electronics directly. It flows through sealed cold plates bolted onto the GPU/CPU die, and a coolant distribution unit (CDU) hydraulically isolates that server-side loop from the facility chilled-water loop, with leak detection and redundant containment built in. Immersion cooling is the one approach where components are submerged, but the fluid used is a dielectric (electrically non-conductive) engineered fluid specifically chosen because it won't short the electronics, not water. Both approaches are designed from the ground up to prevent the electrical hazard the misconception assumes is inherent to the whole category.
Explains why AI/GPU data centers are shifting from traditional CRAH/CRAC air cooling to direct-to-chip and immersion liquid cooling — driven by a fundamental gap in volumetric heat capacity between air and liquid, not simply because GPUs 'run hotter.' Illustrated with a traditional air-cooled data hall and a direct-to-chip liquid-cooled GPU rack.
It's easy to assume liquid cooling is just a more aggressive version of air cooling, applied because chips got hotter. But CPUs have run hot for decades and stayed air-cooled — the change with GPU clusters is density, not just per-chip temperature. An 8-GPU AI server can draw 10+ kW by itself; pack several into a rack and you can exceed 100 kW in a single rack footprint that a traditional data hall was designed to handle at 5-10 kW. Air cooling doesn't fail at the chip level, it fails at the rack/room level, because moving that much heat via air requires airflow volumes and fan power that become physically and economically impractical.
Traditional air cooling uses CRAH (computer room air handler) or CRAC (computer room air conditioner) units to push chilled air through a raised floor or overhead ducting into a cold aisle, across server intakes, absorbing heat, and returning as hot air through a hot aisle back to the cooling unit. This works well up to roughly 10-20 kW/rack; density beyond that requires supplemental measures like rear-door heat exchangers or in-row cooling before it becomes a hard wall.
Direct-to-chip liquid cooling mounts sealed cold plates directly onto the GPU and often CPU dies. Coolant circulates through the cold plates and back to a coolant distribution unit (CDU), which pumps the server-side loop and exchanges heat with the facility's chilled-water loop through a heat exchanger — the two loops never mix, and the CDU includes leak detection and containment. Immersion cooling goes further, submerging entire servers in a dielectric engineered fluid that absorbs heat directly from every component, not just the highest-power chips.
The cooling approach cascades into nearly every other design decision. Liquid-cooled racks need CDUs, secondary cooling loops, leak detection, and often a completely different rack/server mechanical design than air-cooled equipment — retrofitting an existing air-cooled facility for liquid cooling is a major capital project, not a drop-in upgrade. It also affects PUE: liquid cooling typically moves the same heat load using less pumping/fan energy than the air-cooling equivalent, and can enable higher chilled-water supply temperatures that unlock more free-cooling hours, both of which lower overall facility power overhead. Choosing air vs. liquid — and what fraction of the rack each handles — is one of the first decisions in an AI data center's mechanical design, because it determines the achievable rack density for the whole facility.
Most current direct-to-chip deployments target only the highest-power components — GPUs and often CPUs — with cold plates, while memory, storage, networking, and voltage regulators remain air-cooled by rack-level fans. This hybrid split (commonly 60-80% liquid, 20-40% residual air by heat load) is typical; full immersion cooling is one of the few approaches that cools essentially everything with liquid.
Not usually. Even a heavily liquid-cooled facility typically still needs air handling for the residual air-cooled load (the fraction of heat not captured by cold plates), general room conditioning, and any non-liquid-cooled equipment — so most real deployments run both systems in parallel rather than replacing air cooling outright.
Water has a much higher specific heat capacity and density than air, so a given volume of water moving at the same velocity as air can absorb roughly 3,500 times more heat energy. That means far smaller pipes, far less pumping power, and far less physical space are needed to move the same amount of heat compared to the ductwork and fan power air cooling would require at the same load.
Not automatically — it depends on implementation, but liquid cooling commonly improves PUE because pumping a liquid loop is more energy-efficient per watt of heat removed than moving the equivalent air volume, and higher achievable coolant temperatures can extend free-cooling hours. A poorly designed liquid system could still underperform a well-optimized air system, so the cooling approach and its execution both matter.