Air cooling limits data center racks at 15 to 30 kW. The physics of rack airflow, why liquid is not automatic in that band, and where DBD plasma can fit.
A 42U cabinet drawing 8 kW and the same cabinet drawing 25 kW look identical from the cold aisle. Same doors, same perforation pattern, same tiles in front of them. The difference is entirely in how much air has to cross that 600 mm face every second. Operators talk about density in kilowatts because that is what PDU meters report. The cooling plant does not care about kilowatts. It cares about cubic metres per second.
The air cooling limits data center operators hit first are volumetric, not thermal: a rack does not run out of cooling capacity so much as it runs out of room to push enough air through the front of a cabinet at a temperature rise the IT equipment will accept.
The heat a stream of air can carry is fixed by its mass flow and its temperature rise. At sea level, with air at roughly 1.2 kg/m³ and a specific heat of about 1.005 kJ/kg·K, the volumetric flow needed is the rack load divided by about 1.21 kJ/m³·K multiplied by the rise across the server. There is nothing to optimise there. It is a conservation law.
The table applies it at an 11 K rise (a common legacy design assumption) and a 20 K rise (more typical of modern servers on aggressive fan curves). The last column converts the requirement into standard 300 CFM vented floor tiles [4].
| Rack load | Airflow at 11 K rise | Same, in CFM | Airflow at 20 K rise | 300 CFM tiles needed at 11 K |
|---|---|---|---|---|
| 5 kW | 0.38 m³/s | 800 CFM | 440 CFM | 2.7 |
| 10 kW | 0.75 m³/s | 1,600 CFM | 880 CFM | 5.3 |
| 15 kW | 1.13 m³/s | 2,390 CFM | 1,320 CFM | 8.0 |
| 20 kW | 1.51 m³/s | 3,190 CFM | 1,760 CFM | 10.6 |
| 30 kW | 2.26 m³/s | 4,790 CFM | 2,630 CFM | 16.0 |
| 50 kW | 3.77 m³/s | 7,980 CFM | 4,390 CFM | 26.6 |
You cannot place sixteen vented tiles in front of one cabinet. That is the whole problem in one number. Schneider Electric reaches the same conclusion from the room side: a conventional raised floor with one vented tile per rack cannot sustain much beyond roughly 6 kW per rack over a contiguous area, rising to about 10 kW with containment [4]. Everything above that is bought with supplementary equipment.
The obvious escape is the denominator: double the rise, halve the airflow. Two things stop that.
The first is the inlet. ASHRAE TC 9.9 puts the recommended equipment inlet envelope at 18 to 27 °C, with allowable classes from 15 to 32 °C (A1) up to 5 to 45 °C (A4) [2]. Raising the rise while holding the inlet in that envelope raises the exhaust, and the exhaust is what containment, cabling and the return path must tolerate.
The second is fan power, which rises roughly with the cube of speed. The last increments of flow are the most expensive you will ever buy, and past a point the fans spend watts that reappear as heat in the airstream they are cooling. That accounting sits in our piece on energy efficient cooling rather than here.
Rack airflow problems are usually blamed on chiller capacity. In a room that has not been recommissioned in five years, the constraint is almost always geometric: bypass air escaping through unsealed cable cutouts, recirculation over half-populated cabinets, blanking panels missing from the 8U decommissioned last quarter, and plenum pressure that falls away at the end of a long row.
Containment fixes the mixing and is the highest-return intervention in most rooms. What it does not do is raise the flow a single cabinet can accept through its own front door. Once every tile is open, every gap sealed and the CRAC units at full fan speed, the remaining limit is the pressure drop across the chassis itself. Our overview of data centre thermal management covers how room-level and rack-level decisions interact.
Uptime Institute's 2024 survey put the average typical rack density across respondents at 8 kW, with 41 per cent of deployments in the 4 to 6 kW band and 18 per cent at 10 to 14 kW. Asked about their highest-density racks, 25 per cent reported 15 to 29 kW and 20 per cent reported 30 kW and above [3]. A large share of the estate now sits in the awkward middle: too dense for a room designed around 6 kW, not obviously dense enough to justify facility water.
Three things make that band genuinely ambiguous:
Zoom in to the chassis and the picture changes. Bulk flow through a rack is rarely uniform. Air stagnates in the wake of drive cages and cable bundles, and slows to almost nothing in the boundary layer clinging to heat sink fins and enclosure walls. More CRAC output does not help there, because the bulk air is already arriving. The heat transfer coefficient at that surface is the bottleneck, not the supply.
This is the specific gap dielectric barrier discharge plasma actuators address. A DBD device is around 200 micrometres thick, has no moving parts, and drives air by ionic wind: high-voltage AC ionises a shallow surface layer and the ions drag neutral air with them, forming a wall jet of a few metres per second. The discharge is non-thermal, so it adds no meaningful heat to what it cools. In published measurements, ionic wind superimposed on a 0.3 m/s bulk flow roughly doubled the average heat transfer coefficient and delivered about 20 K of additional cooling while drawing under 100 mW [1].
The gain came from superimposing local flow on an existing bulk flow, which is precisely the condition inside an air-cooled rack. As our note on ionic wind versus fans puts it: a fan moves the room, an actuator fixes the boundary layer a fan cannot reach. The same reasoning applies to the sealed, space-constrained modules discussed in optical transceiver cooling.
Electrical-to-fluid conversion efficiency for corona-based ionic wind sits at roughly 1 to 2 per cent [1]. As a bulk air mover that is poor. Ionic wind produces metres per second, not tens of metres per second. If your problem is the 4,790 CFM in the table above, you need fans. A DBD actuator is not a CRAC replacement, not a containment substitute, and not a route to running a 40 kW cabinet on room air.
The honest constraints are a kilovolt-class AC drive supply, some ozone generation in air that depends on waveform, geometry and duty cycle and is manageable by design, and dielectric ageing under continuous electrical stress. We treat the first two in reliability and ozone. Well above this band the argument is over: liquid is correct, and the question becomes how much residual air load remains.
Between the point where a room's air system is fully commissioned and the point where facility water is justified, there is a band in which the binding constraint is local, not global. That is where a solid-state thermal device with no bearings, no membrane and no refrigerant is worth evaluating. Y-Flow is in development, not deployed at scale, and we would rather say so.
At what rack density does air cooling stop working? There is no single number, because it depends on the room. Schneider Electric's analysis puts a conventional raised floor with one vented tile per rack at roughly 6 kW per cabinet sustained, rising to about 10 kW with aisle containment [4]. Beyond that, every additional kilowatt is bought with supplementary equipment. Well-engineered air-cooled halls run above 20 kW, but not by accident.
Why can I not just raise the temperature rise across the server and use less air? Because the inlet is bounded and the exhaust is bounded. ASHRAE's recommended inlet envelope is 18 to 27 °C [2], and pushing the rise raises the exhaust temperature that containment, cabling and the return path must handle. Server fan power also rises with roughly the cube of speed, so the last increments of flow cost the most.
Is 20 kW per rack a liquid cooling problem? Not automatically. Direct-to-chip typically captures 60 to 80 per cent of server thermal load and leaves the remainder on air [5], so you operate two systems rather than one, and the coolant distribution and piping costs do not scale down for a small number of cabinets. At 20 kW the honest first step is recommissioning the air path.
What does a DBD plasma actuator do that a larger fan does not? It works inside the boundary layer, where bulk flow has already stalled. Measurements show roughly a doubling of the average heat transfer coefficient and about 20 K of extra cooling when ionic wind is superimposed on a 0.3 m/s bulk flow, at under 100 mW [1]. A fan cannot produce that local effect at any reasonable speed.
Can plasma actuators replace CRAC units or aisle containment? No. Conversion efficiency from electrical to fluid energy is around 1 to 2 per cent and the flow produced is metres per second, not tens of metres per second [1]. Room-scale air movement remains the job of fans and air handlers.