NVIDIA is promoting open AI models and locally controlled infrastructure for European public-sector projects, highlighting applications involving German tax systems, French employment services and Polish healthcare.
At GTC Berlin on 22 October 2026, the company is positioning government-controlled AI as an alternative to relying exclusively on externally hosted, proprietary APIs. The approach gives public institutions more options for deciding where sensitive information is processed, how models are adapted and which suppliers operate the underlying systems.
For government technology buyers, the development raises practical questions about procurement, data sovereignty, compliance and long-term supplier dependence. Open models provide additional control, but running an AI system locally does not automatically make it secure, compliant or independent.
What NVIDIA Is Showing European Governments
NVIDIA’s GTC Berlin public-sector programme describes AI applications operating on models and infrastructure controlled by the institutions using them. Its examples span tax administration in Germany, employment services in France and healthcare in Poland.
The programme also covers inspectable models used with the MONAI medical imaging framework, confidential computing for sensitive workloads and reference architectures that systems integrators can adapt for other government services.
According to NVIDIA’s official GTC Berlin public-sector programme, these environments require deployment location, auditability, and adaptation to local regulations.
NVIDIA’s wider open-model portfolio includes Nemotron models and tools for language processing, retrieval and specialised AI applications. These provide building blocks that developers can adapt to institutional requirements rather than depending entirely on a fixed remote service.
However, the programme does not provide complete public technical specifications, independent security assessments or procurement details for every deployment it references. Those claims should not be treated as evidence that each application has passed an external compliance audit.
On-Premises AI and Sovereign Cloud Are Different Procurement Choices
The term “sovereign AI” can describe several infrastructure arrangements. Each provides a different degree of control over data, models and operations.
| Deployment approach | What the institution controls | Main limitation |
|---|---|---|
| On-premises open models | Local inference, model storage, infrastructure access and internal data processing | Hardware maintenance, security updates and specialist engineering remain the institution’s responsibility |
| European sovereign cloud | Regional processing requirements, contractual data controls and potentially dedicated infrastructure | The provider may still control platform operations, access mechanisms and infrastructure dependencies |
| Confidential computing | Additional workload isolation and protections for sensitive processing | Protecting data during execution does not establish full visibility into a proprietary model or its behaviour |
One useful detail in NVIDIA’s programme is that its public-sector approach is not limited to open-weight models. It also describes confidential computing that can support proprietary models operating on local infrastructure while protecting both patient data and model intellectual property.
That creates another possible procurement route: an institution may retain control of the processing environment without receiving permission to inspect or modify the underlying model weights.
The trade-off needs to be explicit. A model can run in a government data centre while remaining dependent on an external supplier for licensing, updates, specialised software, or replacement hardware.
The Biggest Governance Risk May Be Outside the AI Model
Keeping inference local is useful, but government AI systems rarely consist of a model alone. They often include document extraction, embedding services, vector databases, monitoring software, identity management and integrations with existing case-management systems.
Each component creates another potential route for sensitive information to leave the approved environment.
For example, a department might install an open model on its own servers while sending document embeddings to an external service. Although the model remains local, part of the information-processing pipeline still runs elsewhere.
Another common problem arises when application logs record prompts, retrieved documents or model responses. A remote monitoring platform could then receive information that the organisation intended to keep entirely within its infrastructure.
Infrastructure practitioners repeatedly highlight the tension between the operational convenience of managed enterprise AI services and the greater administrative responsibility of self-hosting. Data protection agreements can address certain risks, but they don’t replace technical control over every processing component.
Public-sector teams should therefore map the complete information flow, including backups, telemetry, error reporting, software updates and administrative access, before describing any deployment as sovereign.
European AI Compliance Still Depends on the Application
NVIDIA’s positioning aligns with the requirements European institutions face when handling sensitive information. However, infrastructure ownership and regulatory compliance address different questions.
Under the GDPR, organisations still need a lawful basis for processing personal data, appropriate security measures, defined retention policies, and required assessments. Hosting an open model within Europe does not remove those obligations.
The EU AI Act introduces additional responsibilities based on an AI system’s purpose and classification. Certain applications affecting access to public benefits, employment or other sensitive decisions may fall within high-risk categories.
Where applicable, those requirements can include human oversight, operational monitoring, record-keeping, registration and fundamental rights impact assessments.
A healthcare document assistant and an automated eligibility assessment system could use the same underlying model yet face substantially different regulatory obligations.
Government suppliers must evaluate the finished application, its intended use, and how responsibilities are allocated between provider and deployer. UK public-sector projects also need to consider applicable domestic legal and regulatory requirements rather than assuming EU rules apply the same way.
Localisation Requires More Than Translating an AI Interface
European public-sector systems must often handle national terminology, regional languages, specialist vocabulary and administrative procedures that general-purpose models may not recognise reliably.
Open models offer opportunities for retrieval-augmented generation, domain-specific fine-tuning and controlled evaluation using locally approved material.
However, translating an interface or supplying local documents does not prove that the underlying model understands national rules.
Consider an employment-support application that retrieves guidance in multiple languages. A model might produce fluent explanations while confusing eligibility criteria, outdated rules or different benefit categories.
A practical evaluation process should test whether the system retrieves the correct policy version, identifies exceptions and refuses to make unsupported claims. Testing should include regional terminology, minority languages where relevant and documents containing conflicting information.
For public institutions, accuracy against approved local procedures matters more than performance on a general AI benchmark.
What Public-Sector Buyers Should Put Into Their AI Procurement Requirements
NVIDIA’s announcement gives suppliers a stronger argument for offering open-model infrastructure. Buyers should translate that proposition into verifiable contractual and technical requirements.
- Model rights: Establish whether the institution can download, inspect, modify, operate and retain the model under its actual licence.
- Data location: Identify where prompts, retrieved documents, embeddings, telemetry and backups are processed and stored.
- Operational control: Specify who can administer systems, access logs, deploy updates and approve changes to the model or infrastructure.
- Evaluation evidence: Require testing against local languages, accessibility requirements, security threats and the actual government workflow.
- Exit provisions: Confirm whether model weights, configurations, evaluation records and application data can be transferred to another provider without prohibitive costs.
- Ongoing assurance: Define responsibilities for security patches, incident reporting, model changes, human oversight and periodic performance reviews.
One useful procurement test is to ask a shortlisted supplier to demonstrate the same application operating without its managed cloud endpoint. If essential functions immediately fail, the institution has identified a dependency that should be reflected in its contract and risk assessment.
For teams considering their own infrastructure, DIY AI’s private AI server and self-hosted LLM guide explains the practical differences between desktop, server and managed deployment approaches.
Open Models Could Reduce API Dependence but Increase Operating Costs
Self-hosting changes the economics of AI services but doesn’t automatically make them cheaper.
Public institutions running workloads continuously may find dedicated infrastructure easier to budget for than variable API charges. They can also gain greater control over model versions, upgrade schedules and service availability.
However, those advantages come with costs for GPUs, electricity, cooling, engineering, maintenance, security operations and spare capacity.
A relatively small document-processing service with unpredictable demand might remain cheaper on a contractually approved managed platform. A consistently busy application processing sensitive records could make dedicated infrastructure more attractive.
The correct comparison is cost per successfully completed, compliant task over the expected contract period. Token pricing or hardware purchase costs alone cannot establish the better option.
DIY AI’s AI model cost comparison provides a framework for understanding inference costs and how workload characteristics affect model economics.
What NVIDIA’s Public-Sector Push Has Yet to Prove
NVIDIA has set a clear direction for its European public-sector offering: adaptable models, locally controlled infrastructure, and security technologies designed to support sensitive institutional workloads.
Its programme identifies relevant government applications and describes deployment approaches that suppliers can incorporate into production systems.
Less clear is how easily those architectures can transfer between institutions, cloud providers, and hardware environments. Independent evaluations, long-term operating costs, contract terms and documented exit procedures will determine how much control buyers actually retain.
A broader sovereignty question also remains. European institutions may control where their applications run while still depending on an American supplier for accelerated computing hardware and parts of the software stack.
For governments evaluating NVIDIA-based solutions, the immediate priority should be a limited, independently assessed pilot that tests data boundaries, local-language accuracy, operational resilience and supplier portability before committing to a larger deployment.
Open models give public bodies more choices about how AI is delivered. The procurement challenge is ensuring those choices remain available after the initial contract has been signed.
Reporting basis
Sources
Primary source: NVIDIA Official
