Cooling can rival servers for data centre electricity. Why fan power scales with the cube of fan speed, and why PUE cannot see the fans inside a chassis.
The Congressional Research Service puts computing and server systems at roughly 40 per cent of a data centre's electricity and cooling systems at another 38 to 40 per cent [1]. For every watt spent on computation, close to another watt is spent moving the resulting heat somewhere else.
Data center cooling energy consumption is the electricity spent moving heat away from IT equipment, and it spans two separate budgets: the fans and pumps inside the chassis, which are billed as IT load, and the air handlers, chillers and cooling towers outside it, which are billed as facility overhead. The split matters more than it sounds, because the industry's standard efficiency metric only sees one of them.
| End use | Approximate share of facility electricity [1] |
|---|---|
| Computing and server systems | ~40% |
| Cooling systems | 38 to 40% |
| Network and storage equipment | ~10% |
| Power conversion, UPS, lighting | the balance |
These shares move with climate, economiser hours and cooling architecture. The structural point survives: cooling is a co-equal line item alongside the compute, not a rounding error. Our overview of data centre thermal management covers how those architectures differ.
The most useful arithmetic in this field is one of the fan laws. Airflow varies in proportion to fan speed, pressure with the square of speed, and shaft power with the cube:
``` Power_final = Power_initial × (RPM_final / RPM_initial)³ ```
The US Department of Energy's fan sourcebook puts the consequence plainly: "Reducing fan rotational speed by 20 percent decreases fan power by 50 percent" [2].
| Fan speed vs baseline | Airflow | Static pressure | Shaft power |
|---|---|---|---|
| 100% | 100% | 100% | 100% |
| 90% | 90% | 81% | 73% |
| 80% | 80% | 64% | 51% |
| 70% | 70% | 49% | 34% |
| 50% | 50% | 25% | 13% |
Two things follow. First, the last increment of airflow is extraordinarily expensive: the step from 90 to 100 per cent speed costs more power than the whole span from rest to half speed. Second, you do not need a large airflow reduction to get a large energy reduction. An intervention that lets a fan curve back by a fifth has already halved its power draw, without replacing the fan. That is a more forgiving argument than the usual one about ionic wind versus fans as air movers.
This matters because fans are rarely run fast for the average thermal load. They run fast because one location in the box demands it. The control loop watches the worst sensor, and the whole chassis pays cube-law prices for a square centimetre of silicon, a pattern examined in our piece on semiconductor cooling and local hot spots.
Power usage effectiveness is total facility energy divided by IT equipment energy. Uptime Institute's 2025 global survey reports a weighted average annual PUE of 1.54, "marking the sixth consecutive year that this headline figure has virtually stood still", against 2.50 in 2007 and 1.65 in 2014 [3]. The easy facility-side wins were taken a decade ago.
A PUE of 1.54 means about 35 per cent of the electricity entering the building never reaches the IT equipment. That overhead is what the metric was built to expose, and it does so well.
What the ratio cannot do is look inside its own denominator. Server fans, chassis blowers and any pump drawing from the server's own supply count as IT load, sitting on the bottom of the fraction. The consequences are awkward:
None of this makes PUE a bad metric. It is a facility metric doing a facility job, but it is the wrong instrument for anything that lives inside the chassis, and a claimed data center pue improvement attached to an IT-side thermal device deserves a second look. The honest measure is total facility energy per unit of delivered compute, or at the bench, kelvin of junction temperature drop per watt at the wall.
Electrical-to-fluid energy conversion for corona-driven ionic wind sits at around 1 to 2 per cent. As a figure of merit for a bulk air mover that is poor, and should be stated as such. When the job is volumetric flow through a duct, a fan remains the correct answer.
The fan laws reframe the comparison. Go and colleagues superimposed an ionic wind on an existing 0.3 m/s bulk airflow and measured roughly a factor-of-two increase in the average heat transfer coefficient and about 20 K of additional cooling, from a device drawing under 100 mW, specifically 67 mW at 15 µA [4]. The question is not whether those 67 mW move air efficiently, but what they let you stop doing.
If doubling the local heat transfer coefficient at the hot spot that sets the fan curve lets that fan run 20 per cent slower, the cube law says its power has halved. The sum is therefore a ratio: the actuator's wall-plug draw, divided by the fan power you actually avoid. Use measured fan power at both operating points, not nameplate ratings. The efficiency of the air-moving step barely enters the calculation, because the device is not being asked to move the air. It is being asked to disrupt a boundary layer so that air already in motion carries more heat away. That is the case for energy efficient cooling at the component scale rather than the room scale, and why the Y-Flow thermal work targets the boundary layer rather than the duct. Air the IT equipment does not demand is also air the facility need not condition, though that second-order effect should not be claimed without measuring it.
For the physics underneath the device, see how DBD plasma actuators work. The energy case does not rest on the actuator being efficient. It rests on the machine it lets you throttle being priced by a cube law.
What share of data centre energy is used for cooling? The Congressional Research Service puts cooling systems at 38 to 40 per cent of a data centre's electricity consumption, roughly comparable to the 40 per cent attributed to computing and server systems [1]. The real figure varies widely with climate, cooling architecture and how many hours a year the site can run on free cooling.
Why does a small airflow reduction save so much energy? Because of the fan laws. Fan power varies with the cube of rotational speed, so a 20 per cent speed reduction cuts fan power by about half [2]. Small reductions in demanded airflow produce disproportionately large reductions in fan energy.
Does a lower PUE always mean lower total energy? No. PUE is total facility energy divided by IT energy, and server fans count as IT energy. Anything that makes server fans work harder inflates the denominator and improves PUE while increasing the building's total draw. PUE is a facility overhead metric, not a total efficiency metric.
How can a device drawing under 100 mW be worth using if it is only 1 to 2 per cent efficient? Because the comparison that matters is not joules of air moved per joule consumed. Published work measured roughly a doubling of the local heat transfer coefficient and about 20 K of additional cooling from 67 mW [4]. If that lets a fan drawing far more than 67 mW run slower, the cube law does the rest. The actuator is disrupting a boundary layer, not moving bulk air.
What is the current average data centre PUE? Uptime Institute's 2025 global survey reports a weighted average annual PUE of 1.54, essentially unchanged for six consecutive years, compared with 2.50 in 2007 and 1.65 in 2014 [3].