BoF: Towards a European AI Hub¶

This BoF explores what a common European AI Hub for models, embeddings, vector databases, and RAG components could look like — and what requirements EuroHPC infrastructures must meet to support it effectively.

Abstract¶

EuroHPC's compute infrastructure is world-class, but the surrounding AI/ML data ecosystem remains fragmented. This BoF explores what a common European AI Hub for models, embeddings, vector databases, and RAG components could look like. Through interactive roundtable discussions, participants will examine storage patterns, federation models, HPC integration challenges, reproducibility requirements, and governance approaches. This non-commercial session uses minimal presentation (5 min) and extensive moderated discussions (45 min) across eight themed tables with participant rotation. Rather than proposing solutions, we aim to collect diverse perspectives, identify shared challenges, and establish design principles. Outcomes include shared understanding of community requirements, networking opportunities, and potential coordination with European initiatives.

Motivation¶

Large-scale AI/ML workflows increasingly depend on shared access to models, embeddings, retrieval-augmented generation (RAG) components, vector databases, and spatial or multimodal datasets. While EuroHPC systems provide world-class compute infrastructure, the surrounding data and model infrastructure remains fragmented across institutions, clouds, and individual projects.

Practitioners across Europe consistently raise critical questions:

• How do we store, version, and share foundation models, fine-tunes, and adapters across EuroHPC sites?
• How can vector databases, embedding stores, and RAG indices be managed efficiently, securely, and at scale?
• What role should AI application artifacts (pipelines, prompt templates, agents) play alongside raw model weights?
• How can this be accomplished in a sovereign, federated, and reproducible manner suitable for European research infrastructures?

This BoF brings together EuroHPC users, operators, system architects, and AI/ML researchers to collectively explore what a common European AI Hub could look like and what requirements EuroHPC infrastructures must meet to support it effectively.

Tentative Agenda¶

The session is structured around four thematic blocks. Each block opens with short lightning talks to prime a focused roundtable discussion with a panel of ~6 participants.

Segment	Title	Format
1. Building Blocks	What are the building blocks, how do they look, and how do we handle them? Blobs, OCI container images for frameworks (PyTorch, vLLM), model weights, spatial data (RAG, embeddings, memory), ...	Lightning talk + roundtable
2. Current State & Challenges	How do we handle them today and what are the challenges? Perspectives from the US (Project Genesis) and from EuroHPC	Lightning talks + roundtable
3. Vendor & Project Ecosystems	What do the vendor ecosystems look like? NVIDIA, AMD, Intel; CNCF OPEA; HPE; and other relevant projects	Lightning talks + roundtable
4. The Future	What do we actually want — and is it achievable? Open discussion on design principles, federation models, sovereignty, governance, and a path forward for a European AI Hub	Extended roundtable (main session)
END	BoF closes — networking continues

Session Goals¶

Rather than proposing pre-defined solutions, this session aims to:

• Collect diverse perspectives from operators, users, and researchers
• Identify shared challenges across institutions and national infrastructures
• Establish community-driven design principles for a European AI Hub
• Create networking opportunities and lay the groundwork for potential coordination with European initiatives

Organizers¶

For questions or to express interest in presenting a lightning talk, contact the organizers at hpcw (at) qnib.org.