Reports on artificial intelligence throw up spectacular figures about the energy use of data centres every day. Accounts range from comparisons between a single ChatGPT query and leaving a light bulb burning, to alarming statements about the share of computing clusters in the national power grid. Anyone following these publications closely soon discovers that the data presented vary widely. In the foundational overview of the debate about data centres and AI computing capacity in the Netherlands we analysed the physical space and public discussion around data centres; in this deeper treatment we focus specifically on the methodology behind the energy consumption figures and on how to interpret those figures responsibly.
Quantifying the power hunger of artificial intelligence requires a clear distinction between measured physical values and theoretical models. Many of the statistics doing the rounds are based on extrapolations using outdated assumptions about hardware efficiency or cooling. To understand how much electricity an AI model actually consumes, you need to know exactly what has and has not been included in the calculation. From the individual microchip to the complete high-voltage substation of grid operator TenneT: every layer in the chain has its own losses and measurement errors.
The basic formula: PUE and the limits of efficiency metrics
In the data centre sector, Power Usage Effectiveness (PUE) has for decades been the standard measure for expressing the energy efficiency of the infrastructure. The PUE value is calculated by dividing the total energy use of the entire facility by the energy use of the IT equipment alone. A theoretically perfect data centre has a PUE value of 1.0, meaning that 100 percent of the electricity supplied goes to the servers and accelerators. In practice the average PUE of modern hyperscale data centres is around 1.1 to 1.2, with the remaining 10 to 20 percent going to cooling, transformer losses and uninterruptible power supplies (UPS).
Even so, the PUE metric falls short when you look specifically at the impact of AI workloads. A low PUE does not guarantee that a facility runs sustainably; it indicates only how efficiently the supporting infrastructure works relative to the IT power delivered. If the IT equipment itself runs inefficient code or operates on outdated hardware, the PUE can score superbly while total power consumption remains unnecessarily high. For deeper insight into the broader relationship between hardware efficiency and sustainability questions, you can consult the broader analysis of AI and energy on the subdomain.
A second limitation of PUE is that it leaves out the cooling medium. Facilities using evaporative (adiabatic) cooling can artificially lower their electrical PUE because water takes over the cooling work from electrical compressors. This lowers the electrical PUE score but drastically raises water consumption. Without accompanying statistics such as Water Usage Effectiveness (WUE), a published PUE value therefore gives an incomplete picture of the total ecological footprint.
Training versus inference: how energy demand shifts
When analysing energy reports it is essential to check whether the figures relate to the training phase or the inference phase (the usage phase) of language models. Training an advanced frontier model is a short-lived but extremely intensive process. For months on end, tens of thousands of graphics processors (GPUs) run continuously at maximum power capacity to process billions of parameters and tokens. This peak load requires an enormous concentration of electrical power at a single location.
The inference phase has a completely different consumption profile. Generating a single answer from a trained model takes a fraction of the energy of a training run. Because the model is consulted by millions of users simultaneously, however, the cumulative energy use of inference over the longer term almost always exceeds the initial training consumption. Reports claiming that "training a model costs as much electricity as a small village uses in a year" often fail to mention that this peak is one-off, while continuous use represents the real structural load.
The shift from training to continuous inference also affects the geographical spread of energy demand. Where training clusters can be built in locations with surplus or cheap renewable power available, inference hardware often has to sit close to the end user to minimise network latency. This brings power consumption directly to urban areas where the electricity grid is already heavily loaded.
Hardware specifications: the power profile of modern AI accelerators
To validate power calculations, you have to look at the physical components in the server racks. The power hunger of an AI data centre is driven primarily by specialised hardware such as GPUs and TPUs (tensor processing units). Where a traditional enterprise server draws on average between 300 and 800 watts, a single modern AI server with eight high-end GPUs quickly consumes between 10,000 and 14,000 watts. In the industry, the power requirement of accelerators is expressed as thermal design power (TDP).
To see which technological shifts these enormous power densities cause in silicon design and market supply, we recommend reading the global race for AI chips and computing power . A crucial measurement error in many external reports is simply multiplying a chip's maximum TDP by 24 hours and 365 days. In reality, actual power consumption fluctuates constantly according to the specific computing operation and the memory bandwidth being addressed.
During inference tasks in which a model waits for user input or mainly performs memory-intensive actions, the chip's energy consumption drops to a fraction of the maximum TDP. Only during heavy matrix multiplications does consumption peak. Calculations that assume a continuous load at 100 percent of TDP therefore give a distorted, exaggerated picture of actual kilowatt-hour consumption.
Cooling technologies and their influence on the power and water balance
The extreme energy density of AI hardware pushes traditional air cooling to physical limits. When a server rack consumes more than 40 to 50 kilowatts of power, moving chilled air is no longer sufficient to remove the heat generated efficiently. Data centres are therefore switching to liquid cooling on a large scale. Here a coolant circulates directly past the cold plates on the chips (direct-to-chip cooling), or the servers are fully submerged in a dielectric fluid (immersion cooling).
The switch to liquid cooling has a direct effect on the data centre's energy measurements:
- Higher coolant temperatures: Liquid can absorb heat far more efficiently than air. This allows the cooling water to be led outside at a higher temperature (30 to 45 degrees Celsius, for example), so fewer energy-hungry chillers are needed.
- Lower fan power: In traditional systems, server fans consume a considerable share of the IT power. With liquid cooling these fans disappear or run at minimum speed, which raises internal server efficiency.
- Use of waste heat: Because the heated cooling water has a higher and more constant temperature, it is better suited to connection to local heat networks for heating homes or greenhouses.
When assessing reports, it is important to check how the power for the pumping installations and external cooling towers has been allocated. Is the consumption of the coolant pump counted as IT consumption or as infrastructure consumption? Depending on the definition used, reported PUE figures can shift by several percentage points without the physical reality changing.
Grid congestion and the Dutch context of energy infrastructure
In the Netherlands the discussion about the energy use of data centres plays out against the background of a heavily loaded electricity grid. With AI data centres the problem is often not only total annual consumption in gigawatt hours (GWh), but above all the contracted capacity in megawatts (MW). A data centre reserves a fixed peak capacity with the grid operator so that it can be guaranteed to keep running. Because this capacity has to remain reserved continuously on the high-voltage grid, the facility occupies capacity at the distribution substations that could otherwise have been used by homes or other industries.
This phenomenon is known as grid congestion. Statistics from regulators and grid operators show that in large data centre clusters, such as the Schiphol-Rijk region and Agriport A7 in Middenmeer, the physical limits of the power grid have been reached. When you read figures about the energy use of data centres in the Dutch media, the following concepts are unfortunately confused on a regular basis:
- Connected capacity (MW): The maximum capacity a data centre may draw at any given moment under its contract with the grid operator.
- Actual capacity drawn (MW): The physical peak load in a specific hour of the day, which is often considerably lower than the contracted capacity.
- Annual electricity consumption (GWh): The total amount of energy consumed over a full year. This is the sum of the fluctuating hourly consumption.
Equating contracted capacity with actual power consumption leads to serious overestimates in national energy analyses. A data centre connected at 100 MW does reserve that capacity on the cable, but over a year may on average draw only 60 to 70 MW of continuous capacity.
Methodological pitfalls in estimates and reporting
Many widely quoted figures on the energy use of AI come from calculation models drawn up by consultancies or academic researchers. Because large AI companies and data centre operators rarely publish their exact consumption data for competitive reasons, researchers have to work with estimates. In doing so, they regularly fall into a series of persistent methodological traps.
The first trap is the use of static efficiency assumptions. Technological development in both hardware and software is extremely fast. An algorithm running today on a given GPU can, through software optimisations (such as quantisation, in which computing precision is reduced from 16-bit to 8-bit or 4-bit), consume half as much energy for exactly the same task within a few months. Studies that calculate with the software efficiency of two years ago drastically overestimate current power consumption.
The second trap is the so-called Jevons paradox. When a technology becomes more efficient, power consumption per individual transaction falls. This often means, however, that costs fall and usage increases so spectacularly that total energy consumption rises all the same. Analysts who look only at the rise in the number of tokens processed and multiply this by historical consumption figures ignore the efficiency gains. Those who look only at the efficiency gain per token ignore the growth in volume. A correct analysis must always weigh both factors at once.
For organisations or developers who want to investigate for themselves how their own computing set-up performs locally, the practical step-by-step guide explains exactly how to apply measuring the energy use of your local LLM set-up using physical kWh meters and software profilers.
Transparency requirements and supervision of data centres in Europe
To put an end to the lack of clarity and of verifiable data, the European Union has introduced new regulation. Under the revised Energy Efficiency Directive (EED) , operators of data centres within the EU with an IT capacity from 500 kilowatts upwards are required to report detailed performance indicators annually to a European database.
This reporting obligation covers, among others, the following mandatory data items:
- Total energy consumption, broken down into electricity and any other fuels for emergency generators.
- The share of that energy consumption originating from renewable sources, split into own generation and purchased green power (PPAs).
- Total water consumption and the source of the water used (drinking water vs. surface water).
- The amount of usefully reused waste heat delivered to external heat networks or buildings.
- The total volume of data processed in terabytes, so that energy intensity per data volume can be determined.
In the Netherlands, the Netherlands Enterprise Agency (RVO), working with the regional environmental agencies, oversees compliance with these obligations. Thanks to this standardised reporting, analysts will in future be able to draw on audited primary sources instead of having to rely on theoretical simulations or press releases from market parties.
Conclusion and guidelines for responsible quantification
The energy use of AI data centres is a complex, layered subject that cannot be captured in simple headlines or one-sided comparisons. Anyone wanting to assess figures on the power hunger of artificial intelligence professionally and factually would do well to work through the following checkpoints for every publication:
Checklist for reading AI energy figures:
- Has a distinction been made between the one-off training phase and continuous inference?
- Is the reported value based on directly measured kWh meters or on theoretical extrapolations?
- Which PUE and WUE values were used, and are cooling losses fully included?
- Is the calculation based on contracted grid capacity (MW) or on actual measured power consumption?
- Have recent software optimisations and the latest generation of hardware been taken into account?
By asking these questions consistently, a realistic and well-founded picture of the physical impact of AI emerges. Artificial intelligence requires considerable amounts of energy and poses major challenges for the power grid, but only with clean measurement methods and transparent data can policymakers, developers and society work towards sustainable and responsible solutions.
