AI Sovereignty in Europe

Open-weight models are about 4 months behind the best closed models; the best EU-developed model measured is about 16 months behind.

No independent test yet shows whether open models are good enough for European public services.

Europe allocates public computing at scale (about 55 million EuroHPC node-hours in 2025), but actual use is poorly reported.

Much of the programme is still to come: most of the 19 AI Factories are being set up, and AI Gigafactories won't run before 2028.

The strongest national assets are shared ones, such as LUMI-AI (six countries) and Alice Recoque (three).

The AI Act's high-risk rules are delayed to December 2027 and August 2028, and the relief for open-source models is narrow.

Licence terms matter as much as model quality; some licences withhold rights from EU-based companies.

The hardware will remain non-European for years: the EU makes about 10% of the world's semiconductors.

Customers now have a legal right to leave their cloud provider, but moving an AI system is still hard.

Make the Cloud and AI Development Act's sovereignty test evidence-based. Providers should demonstrate where services run, who can access data and keys, what falls outside their commitments, and how customers can leave.

Require regular usage reporting for EU-funded computing. Each EuroHPC facility should publish quarterly figures on time awarded, time used and waiting times.

Fund independent tests of open and closed models on public-sector work, through the Apply AI Strategy and the AI Office, with published methods and results.