Europe’s attempt to build more sovereign artificial intelligence infrastructure risks replacing dependence on foreign computing capacity with greater exposure to imported energy, drawing the continent’s digital and electricity strategies into the same policy problem.

The Shift Project, a French energy transition think tank, has warned that rapid growth in AI oriented datacentres could increase fossil fuel demand unless construction is coordinated with grids, electricity generation, and controls on consumption. Its intervention follows the European Union’s programme to expand domestic cloud capacity, AI factories, and much larger computing facilities.

Ireland, the Netherlands, Germany, and several Nordic markets have already confronted the effects of concentrated datacentre demand on electricity networks, planning decisions, and local infrastructure. The warning therefore reinforces an established constraint rather than identifying a sudden change, while placing it more directly against Europe’s technology sovereignty objectives.

AI facilities intensify the pressure because they can require large, continuous supplies of electricity and are often planned more quickly than new transmission lines or generating assets can be delivered. A permitted industrial site does not guarantee a timely grid connection, while additional renewable capacity does not automatically provide firm electricity throughout every hour of operation.

Compute capacity now sits inside energy policy

The European Commission wants to expand the continent’s computing base through AI factories and gigafactories intended to support advanced model development. Its proposed Cloud and AI Development Act would increase datacentre capacity, improve access to sites, and strengthen European control over critical infrastructure.

Those objectives remain constrained by the physical pace of the electricity system. Transmission projects frequently take longer to approve and build than datacentres, while shortages of transformers, switchgear, and specialist engineering capacity can delay both generation and large new loads.

When demand arrives before low carbon supply and grid reinforcement, gas fired generation or electricity imports can fill the gap. European companies may then gain access to locally hosted compute while remaining exposed to prices driven by international fuel markets and geopolitical disruption.

Governments also face an allocation problem. Datacentres compete for network capacity with housing, transport electrification, heat pumps, manufacturing, and public infrastructure, all of which feature in European decarbonisation and growth plans. A megawatt assigned to a computing campus cannot simultaneously serve an industrial plant or a new residential development.

Forecasts for future AI demand vary widely because model design, hardware efficiency, utilisation, and deployment patterns remain uncertain. The broad direction is clearer: training and inference will require additional electricity, while efficiency gains do not necessarily reduce total demand when cheaper computing encourages more use.

Location will determine the value of new capacity

European datacentres have traditionally clustered around Dublin, Frankfurt, London, Amsterdam, and Paris, where operators can access fibre, customers, specialist workers, and established commercial networks. Several of those locations now face constrained grids or tighter planning rules.

Regions with abundant low carbon electricity may attract more development, although compute cannot be distributed solely according to energy price. Latency, network routes, water, cooling, resilience, land, and proximity to customers continue to shape investment, while regulated data and public services may require specific jurisdictions.

Coordination between operators, energy companies, grid owners, and public authorities could align connection dates with new generation and transmission. Heat reuse and flexible workloads may improve local energy economics, provided nearby customers can use the recovered heat and contracts support the required infrastructure.

Flexibility also has technical limits. Model training can sometimes be scheduled around electricity availability, whereas inference supporting healthcare, communications, industrial control, or public administration must respond when users require it. Customers will not accept critical systems that disappear whenever the grid tightens.

Credible policy will need to distinguish between workloads that can move and services that require continuous operation. It will also require transparent reporting of electricity consumption, carbon intensity, water use, and local network effects rather than company wide renewable claims that may not describe the power serving an individual facility.

Europe’s compute programme addresses a genuine strategic weakness because businesses, researchers, and public bodies cannot develop advanced systems without sufficient infrastructure. However, the programme will merely relocate dependence unless generation, networks, planning, and computing investment are treated as parts of the same industrial system.