For two decades, European tech policy has been defined more by what it restricts than by what it builds. The European Commission's AI Continent Action Plan, unveiled in April 2025 and now rolling into its operational phase, is an explicit attempt to flip that script. Through the EuroHPC Joint Undertaking, Brussels has selected consortia in France, Germany, Italy, Spain, Finland, Sweden, Poland and Greece to host the first wave of so-called AI Factories — public compute facilities anchored on new EuroHPC supercomputers and designed to give European startups, researchers and SMEs access to frontier-scale GPUs that would otherwise sit only inside US hyperscaler clouds.
A follow-on AI Gigafactories call, targeting facilities with 100,000+ accelerator chips, is now in preparation, alongside a proposed Cloud and AI Development Act intended to harmonise permitting, grid connection and water-use rules for hyperscale data centres across the single market. Taken together, this is the most ambitious industrial-policy push for digital infrastructure the EU has ever attempted. It is also, awkwardly, a tacit admission that the EU's existing regulatory environment has been chasing capital away.
Why Brussels Suddenly Wants to Build
The numbers driving this pivot are sobering. Europe hosts a small fraction of global hyperscale data-centre capacity relative to its share of GDP, and an even smaller share of installed AI-training compute. Most of the GPUs powering European AI research today are rented from US clouds. With model training costs scaling into the hundreds of millions of euros, the EU faces a stark choice: build sovereign capacity or accept permanent dependency.
The AI Factories model is a sensible response. Rather than picking national champions, EuroHPC is co-funding open-access facilities that pair supercomputers with start-up support, data lakes and model-training services. The first wave alone — roughly a dozen sites — represents billions in combined EU and member-state investment. The Gigafactories proposal goes further, contemplating public–private vehicles to underwrite multi-billion-euro facilities that no single European firm could finance alone.
The Permitting Problem Is Self-Inflicted
The harder problem is that even fully-funded data centres cannot be built where the grid will not connect them. Ireland's EirGrid has effectively frozen new large-load connections in the Dublin region since 2022, citing capacity constraints and renewable-integration targets. The Netherlands imposed a hyperscale moratorium that has pushed projects to Nordic neighbours. Germany's draft Energy Efficiency Act imposes some of the world's strictest waste-heat reuse and PUE targets on new builds. Spain and the Nordics are absorbing the diverted demand, but member-state fragmentation means a project rejected in Amsterdam may wait years for grid access in Madrid.
This is precisely the gap the proposed Cloud and AI Development Act aims to close. By creating EU-level criteria for "strategic" data-centre projects — with binding timelines for permitting, grid connection and environmental review — the Commission would do for compute infrastructure what the Net-Zero Industry Act attempted for clean-tech manufacturing and the Critical Raw Materials Act did for mining. The model is sound: harmonised rules, faster decisions, and a single point of contact for cross-border investors.
What Brussels Should Get Right
A pro-innovation reading of this agenda is straightforward: building compute is necessary, and the Commission deserves credit for treating it as infrastructure rather than as a vice to be taxed. But the Action Plan's success will depend on three design choices that are still in play.
- Don't bind AI Factories to procurement preferences for EU-only chips or models. Sovereignty goals are legitimate, but forcing public compute clusters to use sub-frontier hardware will simply mean European researchers keep renting US clouds. EuroHPC's existing procurement neutrality has worked — keep it.
- Make the Cloud and AI Development Act override, not layer on top of, national moratoria. If the Act merely creates a parallel "fast lane" while Dublin and Amsterdam keep their freezes, investors will read it as another consultative process. The strategic-project designation needs real legal force against incompatible national restrictions, subject to genuine environmental safeguards.
- Treat water and energy use as engineering problems, not moral failings. Modern liquid-cooled facilities can hit PUE below 1.2 and run on dedicated renewable PPAs. Brussels should set transparent efficiency standards and let operators meet them — not micromanage cooling technologies or impose blanket waste-heat mandates that only work in dense urban grids.
The Bigger Picture
The AI Continent Action Plan is, in effect, the EU's belated recognition that you cannot regulate your way to technological relevance. The AI Act gave Europe a comprehensive rulebook before it had a comparable industry to apply it to; the Action Plan is the overdue industrial-policy companion. If the Cloud and AI Development Act can genuinely streamline permitting — and if member states accept that strategic compute capacity is a single-market concern, not a local zoning matter — Europe could close part of the gap with the US and China over the next half-decade.
The risk is that the Act becomes another well-intentioned framework that adds reporting obligations without removing real barriers. The test will be simple: in 2028, will a hyperscaler proposing a 500 MW facility in the EU have a clear, time-bound path to grid connection? If yes, the AI Factories bet will have paid off. If not, Europe will have built some excellent supercomputers and watched the rest of the workload migrate to Virginia, Texas and the Gulf.