At Rise of AI 2026 in Berlin, a panel with Dennis Geisse (HPE), Robert Friedrich (Nvidia) and Anna from deepset presents a jointly built solution for deploying agentic AI in classified and highly sensitive environments, developed together with secunet. Robert Friedrich traces the idea back to a US government white paper on agentic AI and momentum from Davos, and describes projects like an Agentic AI Hub and an initiative called Spark. Anna explains that deepset's public-sector and defense customers kept asking whether AI tools were BSI certified, where models were hosted and whether deployments could run fully on premises, which pushed the partners to build a packaged, self-hosted "private cloud for AI" combining secunet's certified security layer, HPE's data and compute infrastructure, Nvidia's GPUs and open-source Nemotron models, and deepset's Haystack orchestration layer. Anna demonstrates a multi-agent situational-awareness system called Aida, built from observation, assessment and interaction agents, that keeps a human in the loop for critical decisions. Dennis Geisse walks through the resulting platform interface. In Q&A, the panel discusses incident response for on-premises agent deployments and where automation should stop in defense contexts.
Key takeaways
- Robert Friedrich (Nvidia) says the sovereignty debate treats open source as key to managing risk, citing the Ukrainian army's requirement for non-revocable software licenses.
- Anna (deepset) says customer questions about BSI certification, model hosting and on-premises operation drove the joint HPE, Nvidia, deepset and secunet solution.
- The partners built a self-hosted "private cloud for AI" combining secunet security, HPE infrastructure, Nvidia GPUs and models, and deepset's Haystack orchestration.
- deepset's demo system Aida uses observation, assessment and interaction agents but keeps a human in the loop for all critical decisions.
- Nvidia releases its Nemotron models with open weights plus full training data, post-training data and libraries, which Robert Friedrich argues is necessary for real openness.
Detailed notes
Origins: from a government white paper to a joint solution
Robert Friedrich says his blueprint for agentic AI in government came from a US white paper on the topic, with momentum from Davos turning it into concrete plans, citing an Agentic AI Hub connecting local government and startups and a larger initiative called Spark as early large-scale projects. Anna, from deepset, says the company's customers in the public sector and defense, who use its Haystack open-source AI orchestration framework, kept asking whether solutions were BSI certified, where the underlying model was hosted, and whether it could run fully on premises. This pushed deepset to team up with HPE, Nvidia and secunet to build an out-of-the-box answer to the infrastructure layers below orchestration.
The joint stack: a private cloud for classified data
Dennis Geisse describes each partner's role: secunet provides BSI-certified security, domain separation and hardware-level encryption for disconnected environments; HPE supplies data management and AI compute infrastructure; Nvidia contributes GPUs, DPUs and software for optimized local inference, including Nvidia Inference Microservices running large language models locally; and deepset's Haystack forms the orchestration and governance control plane. Combined, the partners present this as a "private cloud for AI" delivered as a pre-installed box to the customer's own premises, where users get an app-store-like interface and can connect their existing data sources without copying data, all enclosed within secunet's encrypted, separated security domain.
Open source and sovereignty
Robert Friedrich frames sovereignty as a scalable, adjustable way of managing risk and argues open source is a key component of it, citing a conversation with the Ukrainian army's IT department, which required non-revocable licenses from providers as part of its strictest sovereignty requirements. He says Nvidia's Nemotron model family is released with open weights plus the full training data, including post-training reinforcement-learning and security-trace data, along with open libraries, training scripts and reference blueprints for use cases like video search and retrieval-augmented generation. He cites an independent "openness index" from Artificial Analysis on which Nvidia's models score highly, and argues open source gives users full control over model behavior and a more cost-efficient path to implementation.
Building agents with Haystack: the Aida demo
Anna recounts that deepset, founded in 2018, initially trained models for customers, including the first non-English BERT model, and built its first customer pipeline with Airbus in Toulouse before releasing Haystack at the end of 2019 and adding an AI platform for building, monitoring and evaluating applications in 2021. She demonstrates Aida, a multi-agent system for classified-domain use: an observation agent continuously collects field information such as weather and sensor data, an assessment agent compares that information against a plan and flags deviations, and an interaction agent surfaces relevant information to a human operator through a chat interface. She stresses that a human should always stay in the loop for such decisions and that giving agents enough context is critical to prevent them from missing relevant information.
Q&A highlights
Asked by a CISO from an energy company how incident response works for isolated on-premises agent deployments outside business hours, the panel responds that on-premises does not mean disconnected: a secured network still provides access, and a trained operations team, plus governance and visibility into agent workloads, remains necessary from management, operational and regulatory perspectives, including monitoring for accuracy drift over time. Asked whether human-in-the-loop oversight will remain permanent, Anna argues for a risk-based approach similar to the EU AI Act, keeping a human in the loop for critical and defense applications while automating lower-stakes workflows, adding she is not sure defense use cases should ever be fully automated.
Quotes
The best infrastructure is nothing without load, right? So you need applications, and only if we build applications can we bring AI into action.
Anna (deepset)
I think we need a risk-based approach like in the EU AI Act, and for critical applications we have to have a human in the loop.
Anna (deepset)
A key component of this sovereignty is open source.
Robert Friedrich
Questions answered in this talk
How can agentic AI be deployed in classified or highly sensitive environments?
According to the panel at Rise of AI 2026, HPE, Nvidia, deepset and secunet built a self-hosted "private cloud for AI" that combines certified security, dedicated compute infrastructure, open-source Nvidia models and deepset's Haystack orchestration, deployed on the customer's own premises so classified data never leaves a secured, encrypted environment.
Should AI agents in defense or classified settings be fully automated?
Anna from deepset argues at this talk for a risk-based approach similar to the EU AI Act, keeping a human in the loop for critical and defense-related decisions while allowing automation for lower-risk workflows.
What does Nvidia mean by open source AI models?
Robert Friedrich explains at this talk that Nvidia's Nemotron models are released not just with open weights but also with the full training data, post-training and security-trace data, and open libraries and training scripts, which he says is necessary for true transparency and control over model behavior.