Articles

Sovereign AI in Europe is Now a Power Problem

November 20, 2026
6 minutes
‌

At a glance

  • The biggest challenge for sovereign AI in Europe is access to grid power on a timeline organizations do not control. 
  • ENTSO-E says European data center electricity demand will rise from 87 TWh in 2024 to at least 134 TWh by 2030, an increase of more than 50%.
  • 76% of European data center operators cite access to power as their top challenge for the next three years, ranking well ahead of regulatory compliance at 36% and permitting at 34%.  
  • Grid connection lead times in the most constrained European regions range from a few years to more than a decade, meaning your go-live date depends on a timeline outside your control. 
  • Xebia's work with NVIDIA focuses on deploying and operating AI across private, hybrid, sovereign, and cloud environments with the governance and operational control that regulated European organizations need. 

Europe's data centers are expected to use more than 50% more electricity in 2030 than they did in 2025.

That isn't a vendor forecast. It's ENTSO-E, the association of Europe’s transmission system operators, in its May 2026 report on data centers and the power system. In its estimates, total European data center demand rises from around 87 TWh in 2024 to at least 134 TWh by 2030.

Those figures cover all data centers, not only sovereign ones. Hyperscaler regions, colocation and sovereign providers draw on the same grids.

So here’s my short answer for anyone building an AI roadmap in Europe. Sovereign AI is no longer only a question of whose model and whose cloud. It’s also a question of whose grid, how soon, and how flexibly. The same goes for AI on a hyperscaler: the capacity you plan on is something your provider still has to build. A roadmap that ignores this will slip, however good the model is.

What GTC Berlin actually said

At GTC Berlin last month, NVIDIA described power as the biggest bottleneck in infrastructure and pitched its DSX platform as a way to plan, simulate and run the grid and the AI factory as one system. It also introduced DSX POD, a modular design for building AI factories faster. In the same blog post, NVIDIA mentioned more than $300 billion in equipment backlogged, so power and equipment supply both sit on the critical path.

The same week, NVIDIA pushed open models that organizations can customize with local data and deploy on infrastructure they control. Put those announcements side by side and the direction is clear. Sovereignty is moving from a policy slogan to an engineering plan. And engineering plans have dependencies.

The queue is the real constraint

Here’s the number that surprised me most in the ENTSO-E report. Citing the European Data Centre Association (EUDCA), it says more than three-quarters of European data center operators name access to power as their biggest challenge for the next three years, far ahead of regulatory compliance or permitting.

‌

Connection lead times, the same report notes, range from a few years to more than a decade in the most constrained regions.
For an AI roadmap, the translation is simple. Your deployment plan, sovereign or not, depends on capacity that someone else has to build, on a timeline you don’t control.

What changes for your program

None of this means slowing down. It means planning like an engineer rather than a buyer. Most organizations will keep buying capacity rather than building it, and that’s fine. It just means treating that capacity as a dependency. Four things I’d look at:

1. Treat capacity as a dependency, not a procurement detail.

Whether it’s a hyperscaler region, a sovereign provider or a colocation partner, ask the boring questions early. Is the capacity firm or flexible? Reserved or best-effort? When is it available, and in which region? What happens to your workloads during constrained periods, and what does the contract say?

ENTSO-E points to flexible connection agreements as one way to close the time-to-power gap: earlier access to the grid, in exchange for accepting limits when the system is under strain. Providers that connect this way may pass similar limits on to their customers. It’s worth knowing that before you lock a go-live date into the roadmap.

2. Size the model to the job.

Not every use case needs frontier scale. Smaller open models, adapted to your own data and run efficiently, often do the job for a fraction of the compute. Be nuchter about it, as we say in the Netherlands: start with the workload and the business outcome, not the GPU count. Less compute per use case also means less exposure to everything above.

3. Know which workloads can move.

ENTSO-E sees data centers as potential flexible grid resources, and notes that some workloads, such as AI training and batch processing, can often shift in time, while transactional and real-time workloads can’t. For most enterprises, the everyday AI workload is real-time inference: a copilot, a retrieval assistant, an agent inside a process. That won’t move. What often can: fine-tuning, evaluation runs, embedding and indexing jobs, batch scoring. Write down which is which. The flexible part is your room to maneuver when capacity is tight, and a stronger position when you negotiate terms.

4. Put the people who own capacity in the same room as the AI roadmap.

In many organizations, the AI steering committee plans use cases while the people who negotiate cloud and provider contracts sit elsewhere, and neither is asked how much capacity the roadmap needs, by when. They should be. The constraint is shared, so the plan should be too. And if you run your own or colocated capacity, add facilities and energy to that group.

Production is still the point

Plenty of AI strategies stall at exactly this point, between the steering committee and production. The goal is still the one I care about most: getting AI out of the pilot phase and into the business, running reliably, under your control.

That’s the focus of Xebia’s work with NVIDIA, which is about deploying and operating AI across private, hybrid, sovereign and cloud environments with the governance and operational control that regulated European organizations need.

The capacity question doesn’t change that goal. It just makes the path more honest. Better to find the constraint in your roadmap now than in your go-live date later.

So, a question for European data and AI leaders: does your AI roadmap say who commits to the capacity it needs, and what happens if that capacity arrives six months late?

Written by

Daniel van Dijk

Daniel Van Dijk

As the Sales Lead for Xebia's AI practice, Daniel focuses on building meaningful partnerships with top 250 Dutch companies and several international enterprises to help them navigate their AI journey.

Contact

Let’s discuss how we can support your journey.

‌
‌
‌
‌
‌
‌
‌
‌
‌