Europe’s AI ambitions are running up against the limits of its power grids

Data centers consumed approximately 3.1% of EU electricity in 2024, while the expansion of artificial intelligence could account for around 10% of the growth in electricity demand over the next decade, according to an EUISS analysis. The authors propose that new computing capacity be located where energy and grid capacity are available, and that some data centers temporarily reduce their consumption during peak periods.

Europe risks turning a lack of capacity in electricity grids into a constraint on the development of artificial intelligence if new data centers are built without regard to energy availability and local infrastructure, warns an analysis by the European Union Institute for Security Studies. Clotilde Bômont and Caspar Hobhouse argue that European AI policy must be linked more directly to energy policy, so that computing power is located in areas where electricity is available, grid congestion is lower, and renewable energy that would otherwise be curtailed can be used.

In brief

Data centers accounted for approximately 3.1% of total EU electricity demand in 2024, according to data cited by the EUISS authors.

Data center development could account for around 10% of the growth in electricity demand over the next decade. The figure does not mean that data centers will reach 10% of total EU consumption.

In 2024, approximately 72 TWh of electricity, mainly from renewable sources, would have been curtailed or unused in the EU because of system constraints.

The authors propose locating flexible AI workloads, such as model training, in areas and periods with surplus renewable electricity and lower pressure on the grid.

One scenario cited by the analysis estimates that making data center consumption flexible for only 120 hours per year could avoid the need for approximately 4 GW of backup fossil-fuel capacity. This is a modeled result, not capacity already saved in the EU.

Energy is becoming an increasingly important component of global competition for artificial intelligence because advanced models require computing infrastructure that operates at large scale and consumes significant amounts of electricity. The EUISS authors compare the Union's situation with that of the United States, China, and the Gulf states, where access to abundant energy, capital, and the ability to build infrastructure quickly can facilitate data center expansion.

Worldwide, data centers consumed approximately 415 TWh in 2024, equivalent to 1.5% of global electricity demand, according to the figures used in the analysis. The United States accounted for approximately 45% of this consumption, China 25%, and Europe 15%. Global data center demand is expected to grow sharply by 2030, and AI-dedicated infrastructure is considered one of the main sources of this growth.

In the European Union, the analysis estimates that data centers accounted for approximately 3.1% of total electricity consumption in 2024. The authors state that their development could generate around 10% of the growth in electricity demand over the next decade, an important distinction because the percentage refers to demand growth, not the share of data centers in the EU's future total consumption.

Europe's problem is not only the total amount of electricity produced. Grid capacity, the geographical location of generation, the speed of connection, and the availability of energy at the time and place where a data center needs it may become equally important factors. The analysis indicates connection periods ranging from approximately three years in some Nordic countries to ten years in Germany or the Netherlands.

A significant part of Europe's electricity infrastructure is more than 40 years old, while the electrification of industry, transport, and heating is simultaneously generating additional demand for grids. Artificial intelligence is therefore competing for the same infrastructure with other components of the energy transition and the economy.

The authors also warn about dependence on imported energy. Almost half of the electricity produced in the EU in 2025 came from renewable sources, according to data cited in the document, but imported gas continues to play an important role in balancing the system and ensuring generation when renewable energy is unavailable. Uncoordinated data center expansion could increase the need for backup capacity and dependence on fossil fuels during certain periods.

However, this same energy transformation also creates an opportunity. Solar and wind generation are increasingly producing periods in which electricity supply exceeds local demand or the grid's capacity to transport energy to other areas. Operators are then forced to reduce generation, even though the electricity could be produced.

The analysis cites an estimate of approximately 72 TWh of energy, predominantly renewable, that would have been curtailed in the EU in 2024. In Germany, generation reductions would have reached nearly 10 TWh in 2025. The frequency of negative electricity prices is another signal of this mismatch between generation, grid capacity, storage, and consumption.

In seven member states, negative prices reportedly occurred during approximately 5% of the hours in 2025, while in Germany the price fell to minus 499 EUR/MWh during a period in May 2026. A negative price occurs when supply is so high, and the ability to reduce generation, store, or export energy so limited, that participants are willing to pay for electricity to be consumed.

Bômont and Hobhouse argue that certain AI workloads could become part of the solution if infrastructure is designed to respond to these conditions. Not all artificial intelligence operations have the same technical requirements, and the difference between training a model and using it daily may allow consumption to be located and scheduled differently.

Training large-scale models requires powerful processor clusters that can operate almost continuously for days or weeks, but some of these tasks can be shifted over time or between locations. The authors believe that such activities could more frequently be carried out in areas with abundant renewable generation, less congested grids, or quantities of electricity that would otherwise be curtailed.

Inference, meaning the actual use of a model after training, has different constraints. A single response may consume far less energy than the training process, but millions of continuous requests can generate significant cumulative consumption. Applications requiring very low latency often need to be located closer to users, industry, or other digital infrastructure and cannot be moved as easily to a distant region simply because energy is cheaper.

The authors therefore propose a European “geography of computing” adapted to energy realities. Areas for new data centers could be selected based on grid capacity, clean energy resources, cooling conditions, water availability, and the possibility of reusing heat generated by servers.

At present, data center locations are determined primarily by private companies' decisions. Land prices, taxes, tax incentives, permitting speed, connectivity, and energy supply contracts may weigh as heavily as grid conditions. EUISS warns that the geography resulting from these commercial incentives does not necessarily coincide with the optimal geography of the European energy system.

One solution under discussion is stronger use of local price signals, so that building infrastructure that consumes very large amounts of energy in a congested area becomes less attractive. Another option is coordinated planning through acceleration zones, in which new data centers would be encouraged where the grid and local generation can support additional demand.

Consumption flexibility is the other component. Data centers could receive grid connections in exchange for accepting temporary consumption reductions when the grid is under very high pressure. These agreements would allow operators to connect large consumers more quickly without sizing all infrastructure for rare peak situations.

The analysis cites a study by Agora Energiewende and Deloitte according to which data center flexibility for approximately 120 hours per year could avoid the need to install approximately 4 GW of backup fossil-fuel capacity. The figure comes from a modeling scenario and does not mean that the EU has already eliminated 4 GW of power plants through current data center flexibility.

EUISS also proposes a discussion on prioritizing access to computing resources and the grid. In the authors' view, publicly supported infrastructure should be directed primarily toward applications with public, economic, or security value, including cybersecurity, defense, healthcare, scientific research, grid optimization, climate adaptation, and industrial competitiveness.

High-consumption applications, those with significant redundancy, or those with limited public benefits should not automatically receive the same level of support or preferential access to scarce energy infrastructure, according to the document. This is a recommendation by the authors and not an existing rule regarding data centers' access to electricity in the EU.

The analysis does not argue that the expansion of artificial intelligence should be slowed to protect the energy system. Its thesis is that digital and energy infrastructure must be planned together. If data centers are located and operated flexibly, they can absorb some of the renewable electricity that would otherwise be lost and contribute to more efficient use of the grid.

In the opposite scenario, concentrating new capacity in already congested areas could extend connection timelines, increase the need for backup power plants, and deepen dependence on gas. Europe's ability to build competitive AI models would thus depend not only on chips, investment, and researchers, but also on how quickly the electricity system can be expanded and managed.

The document is an analysis by Clotilde Bômont and Caspar Hobhouse, published by the European Union Institute for Security Studies. EUISS specifies that the opinions belong to the authors and do not automatically represent the position of the European Union.

https://2eu.brussels/ro/analysis/ambitiile-europei-in-ai-se-lovesc-de-limitele-retelelor-electrice