NVIDIA spotlights open AI models powering European government projects, from agentic labor offices to sovereign classified-cloud plans

Governments across Europe are increasingly turning to open AI models, NVIDIA says, to gain the benefits of modern machine learning while retaining control over where models run and how sensitive data is handled. In a new blog post, the company described a growing set of projects in multiple countries that use open models, locally controlled infrastructure, and governance features to support public administration and public services.

At the center of NVIDIA’s European message is the idea that “open models” can be evaluated and adapted before deployment. NVIDIA said governments can test models, run them on infrastructure they control, apply their own safeguards, and keep sensitive information inside controlled environments. The company also pointed to an additional verification step where some open models publish training datasets, allowing agencies to assess whether training data was properly licensed and whether it may carry bias or compliance concerns.

NVIDIA highlighted work tied to Europe’s public-sector AI platform, AI@EC, the European Commission’s shared environment for building and running AI applications on commission-managed infrastructure. Berlin-based deepset is expanding its Haystack Enterprise Platform, which NVIDIA described as the foundation of AI@EC. The update, NVIDIA said, connects the development, execution, optimization, and governance of agents, while adding ways for customers to compare operating costs, evaluate answer quality, and trace problems when outputs are wrong or unsafe.

NVIDIA also described the infrastructure and safeguards used in AI@EC, including Triton Inference Server, which supports shared services for finding and ranking relevant information. In NVIDIA’s account, the platform’s reusable components, role-based access, audit trails, and guardrails are intended to support applications such as document summarization, knowledge management, and multilingual policy analysis. The emphasis is on operational control and accountability rather than simply deploying models.

Beyond the Commission’s shared platform, NVIDIA pointed to national and partner-led efforts that use open-model-based “agentic” systems, including specialized work for healthcare, employment services, and public records. In Poland, Deloitte is helping build agentic systems for healthcare and employment services using open models. NVIDIA said the country’s eHealth Center is developing an Intelligent Services Platform for medical imaging using NVIDIA DGX BasePOD and the MONAI framework, aiming to help physicians analyze X-rays, CT scans, and MRI images by detecting abnormalities and prioritizing urgent cases to reduce turnaround time.

For employment-related work, NVIDIA said Poland’s Ministry of Family, Labour and Social Policy is developing an agentic factory on NVIDIA accelerated computing infrastructure designed to support more than 400 labor offices. The company said the system is intended to run models including PLLuM, GPT OSS, and Inkling for capabilities such as document search, summarization, and digital assistants, with the goal of helping staff assist job seekers more efficiently. NVIDIA also said it is working with SpeakLeash in Poland and the University of Ljubljana in Slovenia on country-specific additions to synthetic personas, with localized datasets planned to follow.

NVIDIA’s post also emphasized open datasets designed to reduce privacy risk when evaluating models for public-sector use. Locai Labs in London released Nemotron-Personas-UK, a dataset NVIDIA said contains 1 million synthetic personas reflecting the UK’s demographic and geographic diversity. NVIDIA described use cases such as healthcare evaluation and bias testing across age groups, education levels, and rural versus urban populations, with the broader purpose of testing whether public-service models behave consistently across communities. In Germany, NVIDIA described Soofi, the Sovereign Open Source Foundation Models research initiative, which is developing Nemotron-Personas-Germany, an open dataset of 1 million records grounded in the 2022 census across Germany’s 16 federal states and 400 districts, created without using real people’s private data.

For organizations dealing with classified information, NVIDIA said a set of partners are moving toward a sovereign cloud offering intended to be authorized for classified workloads. NVIDIA, secunet, Red Hat, and deepset are developing a stack based on secunet’s SINA Cloud foundation. NVIDIA said the planned offering combines NVIDIA accelerated computing plus NVIDIA Nemotron open models, Red Hat OpenShift AI for deploying and managing AI workloads, deepset’s Haystack AI for agent orchestration, and infrastructure operated by secunet in German data centers. The company positioned this as a step beyond standard public clouds for sensitive government environments.

NVIDIA also described a broader ecosystem push to help local teams adapt open models. It pointed to CANUT, France’s public purchasing body for digital services, selecting LINAGORA and Docaposte to deliver a sovereign collaborative workspace built, with NVIDIA technology supporting speech transcription and French-hosted infrastructure. Separately, the company said it is bringing startups, software providers, and labs together through an Open Models Codefest focused on turning open models into locally controlled solutions for public-sector challenges such as medical reasoning, AI tutoring, and model governance, with Deutsche Telekom and T-Systems providing compute infrastructure.

What NVIDIA did not specify in detail is the commercial timetable for many of these initiatives, the performance benchmarks achieved, or the procurement outcomes for specific agencies. The post also does not provide verification results for model accuracy, bias testing outcomes, or security assurance documents, even though it references governance and traceability features. For readers, the practical takeaway is that NVIDIA is positioning open models and companion tooling as a policy- and operations-friendly route for governments, but the final effectiveness will depend on local deployments and validation.

Looking ahead, the projects NVIDIA named suggest that Europe’s public sector may increasingly measure AI not only by model quality but by controllability: where the model runs, who can access it, how outputs are audited, and how training data and evaluation are documented. For NVIDIA and its partners, the next milestone to watch will be how quickly these “agentic factories,” sovereign stacks, and dataset efforts translate into sustained, operational services in government workflows, along with any published outcomes from pilots and deployments.