What Are We Giving Up for a Better Data Centre PUE?

Image credit: Canva

A data centre is no longer judged only by what happens inside the building.

Its power demand affects a grid that also serves homes, industries, and other essential services. Cooling draws on supplies with competing uses. As AI pushes facilities towards higher computing densities, decisions about where and how data centres are built have effects well beyond the site itself.

Regulators are responding in different ways. Some are making access to power as a pre-requisite. Others are allocating new data centre capacity selectively, introducing enforceable efficiency requirements, while some are simply halting builds until they have a holistic way to manage them. The tension is not only on the site location, but what alternative uses those same resources can support.

That opportunity cost makes efficiency more than an operating objective for DC owners. It becomes part of the case for committing scarce shared resources to additional compute capacity.

Power Usage Effectiveness (PUE) has become one of the most widely used indicators in this discussion. Yet a lower data centre PUE alone can't tell us whether more useful compute was produced, the trade-offs made to achieve it, or who ultimately has the ability to improve the outcome.

The closer data centres come to competing directly for shared resources, the less comfortable we should be with using one efficiency metric speak for the whole system.

An Efficient Facility Can Support Inefficient Computing

PUE compares the energy used by the entire facility with the energy consumed by the hosted IT equipment within. It helps reveal the overheads associated with cooling, electrical distribution and other supporting infrastructure. It was not designed to measure how much useful work is produced for each unit of energy consumed.

Consider two racks drawing similar amounts of power. One may be running valuable workloads at high utilisation. The other may contain equipment whose capacity is largely idle. Their contribution to the facility’s PUE could look similar, even though the amount of useful work delivered is different.

This distinction is particularly important in co-lo setups. Operator provides the building, power and cooling, while customers own the servers, and the decisions with regards to hardware refresh cycles, workload placement and utilisation. While DC owners can improve their facilities, they might not have direct influence over how their customers use the energy delivered to their equipment.

There is also a further complication. PUE can improve simply because server utilisation increases and a larger proportion of the facility's fixed overheads are spread across a higher IT load, even when nothing is on the facility side of things. The metric therefore needs to be interpreted alongside operating conditions and workloads rather than treated as a standalone verdict.

Cooling Does Not Eliminate the Trade-Off

AI’s higher-density equipment is also changing the cooling conversations. Moving heat directly from chips into a liquid loop can reduce the work required to cool the surrounding air, and could potentially improve PUE while enabling greater computing density.

However, liquid cooling is just one variable and doesn't totally eliminate the trade-offs. The other key aspect is the way heat is ultimately rejected from the building.

A lower PUE may be achieved through design choices that shift power consumption towards other resources.

The cooling medium introduces additional considerations. While direct-to-chip systems commonly use water-based mixtures, immersion cooling uses a wider range of fluids, some of which have attracted environmental scrutiny.

Therefore, a low PUE doesn't necessarily mean the facility is superior, if it is accompanied by materially higher water demand, increased operating complexity, or other environmental impact. The performance of a data centre needs to be assessed in its entirety rather than through any single measure.

We Can Meter Consumption More Easily Than Efficiency

Power usage is easily measured and commonly used in co-locations for billing, planning and operational management. Modern rack power distribution units provide visibility down to individual pieces of equipment when power feeds are properly accounted for.

But while power consumption and efficiency are related, they are not the same thing.

A high-density AI server may consume significantly more power than a conventional server while delivering relatively higher computational output. Conversely, a device with a relatively lower power draw may spend much of its life underutilised. In both cases, the power meter accurately records energy consumption, but it does not reveal how effectively that energy was converted into useful work.

The challenge is that efficiency depends on context. Equipment performing different functions, workloads varying over time, and utilisation levels fluctuates. A GPU cluster, storage array and database server consume electricity differently and difficult to compare just by power consumption alone. Understanding efficiency requires both energy data and an assessment of the workload.

In a colocation environment, the distinction between consumption and efficiency is particularly important. While operators can provide customers with detailed power data and identify areas of unusually high consumption, the decisions surrounding the hosted equipment remain under the customer's control.

What Does the Number Allow Us to Decide?

If power and water were unlimited resources, these distinctions might be less significant. But as data centres increasingly compete with other uses of grid and water capacity, efficiency measures facilitates informed decisions on new growth, allocations and expansions.

PUE tells us something valuable about facility overheads. Water-use measures tell us something different. Equipment-level readings tell us where power is consumed. Each contributes useful information, but none on its own answers whether scarce resources are being converted to useful compute, or whether the party being assessed has the authority to improve the overall outcome.

This doesn't mean we should discard PUE or wait for a perfect metric. It just mean we need to be precise about what each measure can establish, before using it to support decisions with wider consequences.

Once data-centre growth carries a genuine opportunity cost, that distinction stops being academic.