The European Union Institute for Security Studies has warned that Europe’s ambitions to expand artificial-intelligence computing are becoming inseparable from energy security, as new data centres compete for grid capacity in a region where electricity is relatively expensive and parts of the power system are already constrained.

A new EUISS brief argues that the United States, China, and Gulf states enter the AI infrastructure race with different combinations of cheaper energy, faster infrastructure development, or large domestic resources, while the EU faces higher prices, congested grids, and continuing exposure to imported fossil fuels.

Those disadvantages arrive just as European policy is trying to expand domestic computing capacity. The EU has been rolling out AI Factories around existing supercomputing sites and in July launched a procurement process for as many as seven AI Gigafactories, with up to €10 billion in public support intended to unlock at least €20 billion of additional private investment.

The compute plan cannot be separated from electricity planning because the proposed facilities are unusually concentrated loads. The International Energy Agency expects European data-centre electricity consumption to rise by more than 45 terawatt-hours between 2024 and 2030, an increase of around 70%, even though growth in the United States and China is projected to be substantially larger.

Compute capacity follows available power

The conventional technology-policy response has been to treat access to advanced chips and capital as the main constraints on European AI. Those remain important, but processors become useful computing capacity only once sites can secure enough electricity, transformers, substations, network connections, cooling infrastructure, and backup systems to operate them.

That is already changing where developers look for land. Existing grid connections have turned energy companies such as RWE into prospective data-centre developers, while regions with renewable generation and available transmission capacity have gained an advantage over traditional metropolitan clusters.

The mismatch between technology and energy development timescales makes the problem harder. Computing equipment can be ordered and a data-centre shell constructed relatively quickly, whereas transmission lines, substations, generation projects, and grid reinforcements can spend years in planning, permitting, and construction.

The IEA has noted that tightening supplies of transformers and other electrical equipment are already joining chip availability and planning approvals as constraints on data-centre growth. Europe therefore risks announcing compute capacity that cannot be connected on the timetable assumed by technology programmes unless energy infrastructure is considered when locations are selected.

EUISS proposes treating geography as part of the solution. Rather than allowing every large AI facility to compete for capacity in already congested areas, policymakers could steer projects towards regions with surplus or frequently curtailed renewable electricity, while encouraging workloads capable of changing consumption in response to conditions on the grid.

Not every AI workload needs constant power

A data centre is often described as though it were one continuous block of demand, yet workloads differ. Some online services need to respond instantly and operate with little freedom to move consumption, while training jobs, batch processing, and other computational work can offer more flexibility over when or where they run.

Using that flexibility would require operators and AI infrastructure programmes to interact with electricity markets more actively. Compute could be scheduled around periods of abundant wind or solar generation, while incentives could reward facilities capable of reducing or shifting non-urgent workloads when local grids are under stress.

The concept fits with Europe’s interest in reducing renewable curtailment, where electricity that could have been generated is wasted because the grid cannot transport or absorb it. A sufficiently flexible data-centre load situated in the right part of the network could consume some of that surplus rather than merely adding another inflexible demand centre.

There are limits. AI facilities require reliable power as well as cheap power, fibre connectivity still affects location decisions, and the processors inside expensive computing clusters are valuable enough that operators have strong financial incentives to keep them busy. Moving workloads also depends on software architecture, contracts, data residency, and latency requirements.

Europe’s current compute programme is nevertheless large enough that ignoring those constraints would simply transfer part of the AI-sovereignty problem from technology policy into electricity infrastructure. The EU has already moved from diagnosing its compute gap towards funding physical capacity; where that capacity is built now affects grid investment, industrial geography, and the eventual operating cost of European AI.

Energy dependence adds a further strategic complication. Nearly half of EU electricity generation came from renewable sources in 2025, according to the EUISS brief, but imported fossil fuels — particularly gas — still contribute to system stability. Rapid growth in electricity demand can therefore increase exposure to external energy markets if domestic clean generation and networks do not expand alongside it.

Europe’s AI strategy consequently has two infrastructure programmes moving at different speeds. Governments can subsidise accelerators and announce Gigafactories relatively quickly, while power systems respond much more slowly. Their alignment will determine whether new European compute becomes productive capacity or joins the queue of digital projects waiting for a grid connection.