Europe’s governments buy AI on one condition: they keep the keys

Nvidia and its partners described a run of European public-sector AI projects built on open models and locally controlled infrastructure. deepset’s Haystack platform underpins AI@EC, the European Commission’s shared AI platform, on Commission-managed infrastructure. secunet, Nvidia, Red Hat and deepset plan a sovereign cloud built on secunet’s SINA Cloud, which includes components authorized for classified information. Materna offers its on-premises Sinfonie assistant in Germany. France’s CANUT purchasing body selected a Twake.ai-based workspace that more than 2,500 eligible organizations can use without running their own tender. Poland’s eHealth Centre is building a platform to help physicians read X-rays, CT scans and MRIs, and the country’s labor ministry is building an agentic system for more than 400 labor offices. Greece is developing an agentic system as well.

The common thread is control: governments want to inspect models, adapt them to local languages, run them on their own hardware and keep sensitive data in-house. The same demand runs through Nvidia’s parallel pitches to European banks (banks on open weights) and its cloud partners (European cloud partners).

Citizens’ health records and job-seeker files should not be routed through a foreign vendor’s API by accident, and open models that can be audited, in some cases with published training data so governments can check licensing and bias, are an improvement over taking a vendor’s word for it. Some deployments carry real stakes, though. Imaging software that flags abnormalities and prioritizes urgent cases can shorten waits; poorly validated triage can quietly deprioritize the wrong patient, which makes clinical validation results essential. An agentic system assisting labor-office staff is reasonable; governance rules should keep it from sliding into automated decisions about benefits without clear human review and appeal.

The synthetic persona datasets are a useful testing tool. Locai Labs released a U.K. set of 1 million synthetic personas, and Germany’s Soofi initiative built one grounded in the 2022 census across all 16 federal states and 400 districts, without real people’s private data. Testing whether a public-service model works equally well for rural pensioners and city students is what governments should do; the limit is that synthetic people are only as representative as the statistics behind them.

Many of these systems are sovereign in where data sits and who operates the servers, while still depending on one American company’s chips and software. That reflects the industry’s current structure rather than a contradiction, but local hosting and strategic independence are different things. Nvidia’s pending purchase of Hugging Face (Nvidia buys Hugging Face) would also give it a larger role in where open models are distributed. For European integrators such as deepset, secunet, Materna and LINAGORA, open models create a lane as government suppliers. More in our Nvidia coverage, regulation coverage and enterprise coverage.