Rack densities outran air cooling while PUE stalled at 1.54. A technical guide to data center thermal management: air limits, direct-to-chip liquid,.
Two numbers describe the state of data center thermal management better than any roadmap deck. The first: global average PUE has sat at 1.54 for six consecutive years, according to the Uptime Institute Global Data Center Survey 2025. The second: a single NVIDIA GB200 NVL72 rack is specified at roughly 120 kW. One number has not moved in half a decade. The other is roughly ten times what a conventional enterprise rack drew a decade ago. Everything difficult about cooling infrastructure today lives in the gap between them.
Uptime's 2025 survey puts most production racks in the 10 to 30 kW band, with few facilities operating above 30 kW. AI training racks are the outlier that is becoming the plan: they routinely exceed 100 kW per cabinet. That is not an incremental step. It is a change of regime, and it invalidates the raised-floor, hot-aisle/cold-aisle assumptions that most existing halls were built around.
The thermal problem is not that chips got hotter. Junction temperature limits have barely moved. The problem is heat flux and total rack load. Move 120 kW out of a 42U cabinet and the airflow required, at any sane temperature rise, exceeds what the rack geometry can physically pass.
Air cooling is a mass flow problem. The heat a stream of air can carry is Q = m x cp x dT: mass flow rate times specific heat capacity times the temperature rise across the equipment. Air has a specific heat capacity around 1.0 kJ/kg·K and a density near 1.2 kg/m3 at typical inlet conditions. Water carries roughly four times the heat per unit mass and is about 800 times denser. That ratio is the entire argument.
ASHRAE TC 9.9 sets the recommended inlet envelope at 18 to 27 °C, with Class A4 equipment allowable up to 45 °C. Widening the envelope buys economizer hours and lowers facility power, but it does not increase the heat a given volume of air can remove. It shrinks dT headroom.
That leaves mass flow, which means fans, and fans scale badly. Fan power rises roughly with the cube of flow rate while the heat removed rises linearly. Doubling airflow to double the cooling costs about eight times the fan power. Somewhere between 30 and 50 kW per rack, depending on geometry and inlet temperature, the fan power, the acoustic penalty and the pressure drop stop being worth it. This is a physical ceiling, not an engineering backlog. The same scaling argument at board level is laid out in ionic wind vs fans for electronics cooling.
Direct-to-chip cooling, also called DLC or cold-plate cooling, brings a liquid loop to a metal plate mounted on the die package. Coolant enters at facility temperature, absorbs heat through the plate, and returns to a coolant distribution unit that transfers it to the building loop.
It works because it shortens the thermal path. Instead of die to lid to heatsink to air to room, you get die to lid to plate to liquid. The GB200 NVL72 uses exactly this: an in-rack CDU feeding cold plates across the Grace CPUs, Blackwell GPUs and NVLink switches. NVIDIA contributed the NVL72 rack and liquid-cooled tray designs to the Open Compute Project, which is the clearest signal available that direct-to-chip is now the default above 100 kW rather than an exotic option.
Two variants matter. Single-phase loops circulate water or a water-glycol mix and reject heat as sensible temperature rise; they are simpler and dominate deployment. Two-phase loops use a dielectric that boils at the plate, exploiting latent heat for much higher flux at near-isothermal plate temperature, at the cost of fluid handling and vapour management complexity.
What direct-to-chip does not do is cool the rest of the rack. Cold plates address the processors. DIMMs, voltage regulators, storage, NICs and optical modules are still in an air path. A liquid-cooled rack is a hybrid rack, and the residual air-cooled load is typically 10 to 30 percent of total rack power. That residual is where a surprising amount of operational trouble concentrates. The GPU-level version of that question is covered in fanless GPU cooling.
Immersion cooling submerges whole boards in a dielectric fluid. Single-phase immersion pumps warm fluid to a heat exchanger; two-phase immersion boils fluid at the components and condenses it on a coil above the tank.
Thermally it is excellent. Operationally it is expensive: servicing means lifting a dripping, heavy board out of a tank; fluid compatibility with labels, gaskets and optics is a real qualification burden; and regulatory pressure on certain fluorinated fluids has made some two-phase fluid supply chains a procurement risk. Immersion remains a strong fit for greenfield, homogeneous, high-density fleets and a poor fit for mixed estates.
Rear-door heat exchangers hang a liquid coil on the cabinet exhaust. Servers keep their fans and their air path; the coil captures the heat before it reaches the room. Passive RDHx units rely on server fans alone; active units add their own.
RDHx is the pragmatic retrofit in the 30 to 60 kW band. It requires no server modification and no fluid inside the chassis, and it converts an air-cooled hall into something closer to a liquid-cooled one without touching the IT. Its ceiling is set by the same server-fan mass flow limit described above, because the heat still has to leave the chassis by air first.
Every method above assumes you can move a fluid through the space. The persistent failures happen where you cannot.
Optical transceiver cages sit in dense faceplate rows with millimetre clearances and a case temperature budget that shrinks as module speed rises. Rear-of-rack power shelves and network gear live in stagnant recirculation zones. Sealed and IP-rated enclosures, which is most telecom and edge infrastructure, have no through-flow at all by design. In each case the limiting resistance is not the heatsink and not the coolant loop. It is the thermal boundary layer clinging to the surface, in a volume too small or too closed for a fan to help.
This is where solid-state airflow becomes relevant. A dielectric barrier discharge plasma actuator ionises a thin layer of air at an exposed electrode and accelerates it along the surface using an electric field. The result, ionic wind, is airflow generated by a flat device with no blades, no bearings and no moving parts. It does not compete with a cold plate on kilowatts. It attacks the boundary layer directly, in geometries where no rotating machine fits and no liquid loop is permitted.
Read more on the physics in our guide to ionic wind and solid-state cooling, the transceiver case in ionic wind cooling for data center optical transceivers, and the fan comparison in ionic wind vs fans for electronics cooling.
Use rack power density as the primary selector and retrofit constraints as the tiebreaker.
PUE has been flat for six years because the easy efficiency was harvested a decade ago and the hard part, moving heat out of increasingly dense volumes, is bounded by fluid physics rather than by product cycles. Direct-to-chip liquid cooling has settled the high-density rack question. It has not settled the low-flow, sealed, and boundary-layer-limited parts of the same rack, and those are growing as a share of the problem.
YPlasma builds solid-state DBD plasma actuators that move air with no blades, no bearings and no moving parts — for sealed enclosures, transceiver cages and edge hardware where conventional cooling stops. Book a technical demo.