When to switch to liquid cooling, decided by constraint rather than fashion. Compare air, solid-state augmentation, rear-door, direct-to-chip, immersion.
The Uptime Institute's 2025 global survey puts average modal rack density at almost 9 kW, up from 8.3 kW a year earlier, and finds more than 80 per cent of operators have no rack above 30 kW anywhere in their facility [3]. Its cooling survey puts perimeter air cooling at 75 per cent against 22 per cent using direct liquid cooling of any kind, with lack of standardisation (39 per cent) and cost (38 per cent) the leading barriers [4]. Most rooms are not yet liquid cooling problems. A meaningful minority already are, and more will be within a refresh cycle. The useful question is not which technology is best in the abstract, but which constraint in this room is binding and what most cheaply relieves it.
When to switch to liquid cooling is decided not by rack density alone but by the first hard constraint the room actually hits: the air path can no longer carry the heat at an acceptable fan power, or the floor area, noise budget or containment needed to make air work is no longer available. Density is a proxy for that moment, not the moment itself, and treating it as such is how facilities end up plumbing racks that a containment retrofit would have handled.
Improved air cooling. Aisle containment, blanking panels, raised supply temperature, variable speed fans, in-row or overhead units. Cheap per kilowatt and still the default. Where air stops working is covered in data centre thermal management.
Solid-state augmentation. Thin dielectric barrier discharge actuators that move air by ionic wind with no moving parts. They do not replace a rack cooling system. They attack the boundary layer at a surface where bulk air is present but stagnant. See how DBD plasma actuators work.
Rear-door heat exchangers. A water or refrigerant coil on the back of the cabinet. Liquid in the room, none inside the server. Often the lowest-disruption first step.
Direct-to-chip cold plates. Liquid taken to the package through a manifold and quick disconnects. High capacity, and it changes server, rack and service procedure at once.
Immersion. The chassis in dielectric fluid, single or two phase. Highest capacity, largest change to floor loading, warranty and staff practice.
| Constraint | Improved air | Solid-state augmentation | Rear-door HX | Direct-to-chip | Immersion |
|---|---|---|---|---|---|
| Density band it suits | Low tens of kW, most of the surveyed estate [3] | Any band: component level, not rack level | Up to the 30 to 59 kW band one in eight facilities reports [3] | Tens to hundreds of kW, the range ASHRAE calls "10s to 100s kW" [5] | Highest density and mixed loads cold plates cannot reach |
| Retrofit into a live room | Easiest. No pipework, no server change | Easy. Board or enclosure level, no facility change | Moderate. Facility loop and penetrations, servers untouched | Hard. Loop, CDUs, leak detection, new servers and procedures | Hardest. Tanks, floor loading, fluid logistics, retraining |
| Water availability | None beyond existing plant | None | Facility loop needed, though refrigerant variants exist | Facility loop needed | Little in the tank, but heat still leaves the building somehow |
| Acoustics | Worsens as density rises | Silent. No blades, no membrane | Better if server fans slow, though doors often carry fans | Substantially quieter at the rack | Quietest in the white space |
| Maintenance and failure mode | Filters and fan bearings. Fails by bearing wear and dust ingestion | No bearings or membranes. Fails by dielectric ageing or drive supply | Coil cleaning and water quality. Fails by leak at door or connection | Fluid chemistry and disconnect discipline. Fails by leak at plate or manifold | Fluid inventory and drain procedures. Fails by fluid loss or contamination |
| Where it stops working | When required air volume exceeds the aisle or the noise budget | When the problem needs high volumetric flow, not local mixing | When load exceeds what a cabinet-back coil absorbs | When components off the cold plate still overheat | When change cost outweighs thermal benefit |
Those bands follow the reporting bands in the Uptime Institute 2025 survey [3] and the qualitative range in the ASHRAE white paper [5]. They are not certified thresholds. The crossover in any room depends on supply air temperature, containment quality and floor area, so treat them as a hypothesis to test.
This needs saying plainly. If a rack dissipates tens of kilowatts concentrated in accelerator packages, liquid is correct and nothing solid-state changes that. Ionic wind produces metres per second, not tens of metres per second, and its electrical-to-fluid conversion efficiency for corona-driven flow sits at roughly 1 to 2 per cent [1]. As a bulk air mover that is poor. A direct-to-chip loop moves heat out of the white space; a plasma actuator moves it a few millimetres off a surface into air that something else still has to remove.
So: high volumetric flow, liquid or fans. Rack-level heat rejection, liquid. Hundreds of watts per socket, cold plates. Anyone selling solid-state cooling as a replacement for a facility water loop at those densities is selling something else.
The case is narrow and real. In published work, an ionic wind superimposed on an existing 0.3 m/s bulk airflow produced roughly a factor-of-two increase in the average heat transfer coefficient and about 20 K of additional cooling, drawing under 100 mW (67 mW at 15 µA) [1]. The discharge is non-thermal, so the gas stays near ambient and the actuator adds no meaningful heat of its own [2]. The device is around 200 micrometres thick with no moving parts.
That profile fits four situations a decision table should flag:
Among solid-state options, piezoelectric coolers vibrate a membrane and fail by fatigue, and thermoelectric coolers have no moving parts but poor efficiency and solder fatigue under cycling. The Y-Flow thermal line is in development for these enclosure-level cases.
DBD actuators need a kilovolt-class AC drive supply, a real integration cost and an EMC and safety conversation. They generate some ozone in air, dependent on waveform, geometry and duty cycle, manageable by design but not to be assumed away; we treat that in ionic wind reliability and ozone. Dielectric ageing under continuous electrical stress is the long-run wear mechanism to qualify. And the bulk flow ceiling is not negotiable: if the answer needs cubic metres per second, this is the wrong device. Nor does local augmentation change the facility energy picture by itself; that depends on what it lets you switch off, a separate calculation covered in energy efficient cooling.
The honest summary of data centre cooling solutions is that these five are not competitors on one axis. They sit at different points on density, disruption and water. High density rack cooling above the low tens of kilowatts is a liquid problem. Below it, the binding constraint is usually a hot spot, a noise limit or a containment gap, and those have cheaper answers.
At what rack density should I switch to liquid cooling? There is no certified threshold. As a working rule, well-contained air handles the low tens of kilowatts, which covers most surveyed estates, since more than 80 per cent of operators report no rack above 30 kW [3]. Above that, rear-door exchangers and then direct-to-chip become the practical options. Test the band against your own supply temperature and aisle geometry rather than adopting it.
Can plasma actuators replace liquid cooling in a high density rack? No. Ionic wind produces metres per second of air movement at roughly 1 to 2 per cent electrical-to-fluid conversion efficiency [1]. It disrupts boundary layers locally; it does not reject rack-level heat from the white space. At tens of kilowatts per rack, liquid is the correct answer.
What if my site has no water available? Rear-door and direct-to-chip both need a facility loop, so they leave the list first. That narrows you to improved air cooling, air-cooled or refrigerant-based rear-door variants, and component-level augmentation. Immersion reduces white space water use but the heat still has to leave the building.
Is a rear-door heat exchanger a liquid cooling system? Yes, but only in the room. It brings water to the cabinet without bringing it inside the server, which is why it is often the lowest-disruption first step for an operator who cannot change tenant hardware.
Does solid-state cooling help acoustically constrained sites? It can. DBD actuators have no blades and no vibrating membrane, so they add no acoustic output. Where noise rather than thermal capacity sets the ceiling, replacing or slowing a fan with a silent boundary-layer device changes the binding constraint.