As AI and high-performance computing push power densities to new levels, the traditional language of megawatts and annual electricity consumption is becoming less useful on its own. The more relevant question for operators, customers and policymakers is increasingly this: how much useful computing output does a data centre deliver for every kilowatt-hour it consumes?
The data-centre debate has become dominated by scale. New facilities are measured in megawatts, power demand is discussed in gigawatt- and terawatt-hours, and the expansion of AI infrastructure is raising questions about grid capacity, water consumption and long-term energy availability. Yet those figures reveal only part of the picture.
A data centre that consumes more electricity is not automatically less efficient than one with a lower power draw. If more advanced hardware completes a computational task several times faster, the total energy required for the result may be lower. This is becoming particularly relevant for AI workloads, where high-density GPU clusters can consume substantial power over short periods while producing far more output than older systems. The distinction matters because data-centre efficiency is still too often judged primarily by how much energy a facility consumes, rather than by what it achieves with that energy.
PUE Measures Infrastructure, Not Computing Productivity
Power Usage Effectiveness remains one of the sector’s most established performance indicators. PUE compares the total energy consumed by a data centre with the share used directly by IT equipment. The closer the figure is to 1.0, the less energy is being lost to functions such as cooling, power conversion, lighting or other supporting infrastructure. The metric remains useful, but it has an important limitation: it says little about the efficiency of the IT hardware itself. Two facilities can record almost identical PUE figures while delivering very different levels of computing output from the same amount of energy. One may use newer processors and accelerators capable of completing workloads significantly faster, while the other may require much longer operating periods for the same task. For AI and high-performance computing, infrastructure efficiency and computational efficiency therefore need to be assessed together.
The Industry Is Moving Towards Output-Based Metrics
High-performance computing already offers an established example. FLOPS per watt measures how many floating-point operations a system performs for each watt of power consumed, placing computational performance directly in relation to energy use. AI introduces similar but more workload-specific measurements. Model-training systems can be assessed through samples processed per joule, while the efficiency of large language model inference can be expressed in tokens generated per joule. These metrics change the way infrastructure is evaluated. A system with a higher instantaneous power draw may still be more efficient if it generates substantially more usable output per unit of energy. For cloud providers, that means efficiency improvements do not necessarily translate into falling total electricity consumption. They may instead mean significantly greater computational output from the same amount of energy, or a relatively modest increase in energy consumption compared with a much larger increase in processing capacity. The shift is particularly important as AI growth makes absolute energy demand a political and economic issue across Europe. Efficiency can no longer be treated solely as an exercise in reducing auxiliary consumption. It increasingly concerns the productivity of the entire computing system.
Energy Is Only One Part of the Resource Equation
That broader perspective is already reflected in additional data-centre metrics. Water Usage Effectiveness measures the amount of water used relative to IT energy consumption, while Carbon Usage Effectiveness incorporates the carbon intensity associated with the electricity supplying the facility. Taken together, PUE, WUE and CUE provide a broader picture of infrastructure performance. But as computational demand rises, the next step is to connect those figures with actual digital output. The question for the sector is therefore becoming more complex: how much computing performance is being generated for a given combination of electricity, water and carbon impact? This also means that facilities serving different workloads may require different efficiency benchmarks. An AI training cluster, an HPC environment and an inference platform handling millions of short requests do not operate in the same way and should not necessarily be measured by the same performance assumptions.
Cooling Becomes a Strategic Infrastructure Issue
The move towards high-density computing is making cooling one of the decisive factors in data-centre design. More powerful accelerators concentrate greater amounts of heat within smaller physical spaces, reducing the effectiveness of conventional air-based cooling approaches. OVHcloud says it has been developing its own water-cooling architecture for around two decades and now holds more than 100 patents related to the technology. Its approach uses a closed water loop connected directly to CPUs and GPUs. Direct-to-chip water blocks remove heat at source before transferring it through the facility to dry coolers. According to the company, its data centres achieve an average PUE of approximately 1.24, while newer facilities operate below 1.15. OVHcloud compares this with an average of around 1.52 for German data centres. The company also reports a WUE below 0.3 litres per kilowatt-hour, compared with a cited industry average of 1.8 litres per kilowatt-hour. These figures underline a wider industry trend: as compute density increases, cooling architecture is no longer simply a supporting technical function. It becomes directly relevant to operating costs, available rack density and the amount of useful computing capacity that can be delivered within existing power constraints.
AI Is Beginning to Manage the Infrastructure That Runs AI
The next stage is to make cooling systems themselves more adaptive. OVHcloud’s Smart Datacenter concept introduces predictive control into the cooling architecture. Its Smart Racks use a hydraulic “pull” configuration intended to provide individual servers with the water flow and pressure required at any given moment. More than 30 sensors in each cooling module monitor factors including temperature, flow rate and pressure. Predictive AI analyses this information and adjusts pumps, fans and valves in real time. Operational information from racks, cooling modules and dry coolers is also fed into a data lake and combined with local weather data. The objective is to anticipate future cooling requirements rather than responding only after temperatures or workloads change. The development illustrates a broader change within critical digital infrastructure. AI is no longer simply creating additional computing demand inside data centres; it is increasingly being used to manage the physical infrastructure required to support that demand.
Smaller Cooling Systems Can Have a Larger Impact
OVHcloud has also redesigned the external component of its cooling system. The latest generation of its Smart Dry Cooler requires around half the footprint and number of fans of the previous design, according to the company. The new architecture is intended to reduce electricity demand and noise while improving water efficiency. OVHcloud states that the system can cut cooling-related water consumption by up to 30% and electricity consumption by the cooling infrastructure by up to 50%. Such improvements become increasingly significant as AI drives higher rack densities. The more power concentrated in each rack, the greater the proportion of infrastructure design that must be dedicated to removing and redistributing heat. The efficiency of the cooling layer therefore has a direct influence on how much computing capacity operators can deploy within a given energy and physical footprint.
A Low-Power System Is Not Necessarily the Efficient Choice
The growing diversity of data-centre workloads also challenges conventional comparisons. A system designed for large-scale language-model training operates under very different conditions from one used for scientific HPC workloads or for inference services responding to high volumes of short requests. A server or cluster with a relatively low power draw can appear efficient while taking considerably longer to complete a task. A higher-performance platform may draw more power at any given moment but require significantly less time to produce the same result. The relevant figure is therefore increasingly energy per completed workload, not simply watts consumed during operation. This has consequences for investment decisions. Organisations that select infrastructure primarily on the basis of connected load or annual energy consumption risk favouring systems that appear economical but generate comparatively little output from the energy they consume.
Waste Heat Extends the Efficiency Debate Beyond the Data Centre
Efficiency can also be improved by considering what happens to the heat produced by computing systems. Water-based cooling can increase the temperature at which heat leaves the data centre, making it easier to reuse. Raising the temperature difference between the inlet and outlet of the water loop can improve the performance of the cooling system while also increasing the potential value of the recovered heat. OVHcloud says waste heat from its German data centre is already being used to heat connected office space. At greater scale, similar approaches could allow data-centre heat to feed district-heating networks, industrial processes or nearby residential areas. This adds another dimension to efficiency. A facility should not only be assessed on how little energy is lost within its own boundaries, but potentially also on whether unavoidable heat can be used elsewhere.
Europe Will Need More Than Megawatt Figures
As AI infrastructure expands, data-centre energy consumption will remain under scrutiny. Grid availability, water use, cooling demand and carbon intensity are all becoming strategic considerations for operators and governments. But simply comparing megawatts risks obscuring an equally important part of the debate. A more meaningful assessment would consider the relationship between the resources consumed and the computing value created. PUE, WUE and CUE remain important indicators, but they increasingly need to be complemented by output-based measures such as FLOPS per watt, samples per joule or tokens per joule. For the next generation of data centres, the benchmark for efficiency may therefore shift from the size of the power connection towards the productivity of every unit of energy. The question is no longer only how much electricity the infrastructure consumes. It is how much useful digital work Europe gets in return.
Editorial note: A substantial proportion of the technical assessments and company-specific figures in this article — particularly those relating to PUE, WUE, water cooling, Smart Racks, Smart Dry Coolers and stated efficiency improvements — are based on information provided by Steve Richter, Data Center Manager Germany at OVHcloud, and on technologies and operational data described by OVHcloud. [ML]
