From decade-long grid queues to facilities running at half capacity, a new study exposes the energy crisis at the heart of Europe's push to boost its AI capabilities.
Every time you ask an artificial intelligence chatbot with a question, somewhere, possibly a continent away, a warehouse full of computers is working hard to answer it and a mind-numbing amount of energy is churned to give you a quick reply.
ADVERTISEMENT
ADVERTISEMENT
Data centres, the physical locations that house the supercomputers and associated components that underpin the dramatic increase in AI, are critical in our age of advanced data processing.
But their appetite for electricity is becoming a problem in its own right. These facilities are growing bigger, more numerous and dramatically more power-hungry, and the energy required to run them is scaling just as fast.
The United States currently dominates the global scene with roughly 5,400 facilities compared to around 3,400 across Europe, according to Cloudscene data, and Europe is desperate to close that gap.
The problem is that closing it comes at an enormous energy cost — and the continent's electricity grid is already struggling to cope with existing demand.
A major new study authored by Maria Nowicka at Interface, a European energy and digital policy think tank, highlights just how severe that tension has become.
They warn that without urgent reform, Europe's AI ambitions could end up as costly stranded assets, hoovering up power and public money while being ignored for better options elsewhere.
"Constructing multi-hundred-megawatt facilities that fail to use their contracted capacity effectively would be unsustainable not only economically but also from an energy- and climate-system perspective," the report said.
Electric mega-absorbers
A typical European household uses around 3,600 kilowatt-hours of electricity a year, or roughly 10 kilowatt-hours a day.
The data centre behind your AI assistant can burn through the daily equivalent of tens of thousands of those homes before breakfast.
"The power capacity of top AI clusters is increasing from around 13 MW in 2019 to an estimated 280–300 MW for xAI's Colossus in 2025 — comparable to the demand of roughly 250,000 European households," the report explained.
All this energy has to travel through something, and that something is already under serious strain.
Europe's electricity grid, the vast network of power lines, substations and transmission infrastructure that moves electricity from where it is generated to where it is needed, was not built with AI in mind.
When a single new facility demands hundreds of megawatts at once, it does not suffice to just plug it in. It strains and saps the entire system around it, potentially forcing costly upgrades and crowding out other users competing for the same capacity.
"ChatGPT-4 training reportedly consumed around 46 GWh in total energy — equivalent to a sustained 20 MW draw over three months, and enough to power the entire Brussels Capital Region for over four days," the report continued.
The most advanced models being built now are estimated to consume far more. The International Energy Agency projects that global data centre electricity use will "more than double by 2030, largely due to AI workloads".
Traditional server farms were built around modest, flexible power loads. AI clusters pack specialised chips running at near-maximum intensity for days or weeks at a stretch, behaving, as the report puts it, like "electro-intensive industrial plants connected to constrained grids".
"Grid connection capacity, connection lead times, local congestion, and most recently energy prices, have already become binding constraints, delaying or redirecting large deployments despite initial investment interest," according to Interface.
Will the grid keep up?
Nowhere is this more visible than in Europe's most sought-after data centre markets or what the industry calls the FLAP-D cities, or Frankfurt, London, Amsterdam, Paris and Dublin.
The queues for grid connections have grown so long that they have effectively become a ban on development.
"In the FLAP-D markets... new facilities wait on average 7 to 10 years for a grid connection, rising to 13 years in the most congested primary markets," the report explained.
Ireland has imposed a de facto moratorium on new data centres in Dublin until 2028, while the Netherlands and Frankfurt have effectively banned new connections until at least 2030.
The report noted that OpenAI has been "putting their UK and Norway investments on hold due to high electricity prices," a possible signal that even the world's best-capitalised AI companies are being stopped in their tracks by Europe's energy constraints.
What needs to change
Europe's electricity grid is already contending with the demands of electrifying transport and heating, the uneven rollout of renewables, and what the report calls the risks of "tight gas and power markets," further strained by Russia's invasion of Ukraine and ongoing conflict in the Middle East.
The report recommends that European facilities be integrated into national and EU grid planning from the outset, with siting decisions tied to renewable energy availability.
Piling on hundreds of megawatts of AI infrastructure risks making all of that harder and more expensive.
"The long-term value and acceptability of large AI compute clusters will depend on whether they are conceived, regulated and operated as critical energy infrastructure distinct from traditional data centres," the report concluded.
