This paper, recently submitted to the CHEP 2026 proceedings, outlines the EGI Federation’s Strategy for 2026–2030, focusing on how a massive federated e-infrastructure (serving over 160,000 researchers) is pivoting to meet the demands of the post-2023 AI landscape.
TL;DR
EGI is shifting its AI strategy from resource-heavy model training to the "last mile" of scientific research: domain-specific customization and long-term inference. By building a continuum of compute, data, and human expertise, EGI aims to bridge the gap between massive supercomputing "AI Factories" and the daily needs of diverse scientific communities.
What Problem Does This Paper Address?
The "playbook" for AI in science changed radically in 2023. Previously, applying AI meant training bespoke models on curated datasets from scratch. Today, high-quality foundation models are widely available, shifting the bottleneck from training to adaptation (e.g., fine-tuning, RAG, and prompt engineering).
However, scientific communities face significant hurdles in this new era:
- Infrastructure Gaps: Moving models and data seamlessly between high-performance computing (HPC) for training and cloud environments for inference.
- The Skills Gap: Researchers often know what they want to achieve but lack the engineering expertise to deploy AI at scale.
- The Trust Gap: As AI moves toward "agentic" behavior (autonomous, long-running tasks), ensuring reproducibility and ethical compliance becomes a critical blocker for high-risk scientific applications.
Methods
The authors propose a strategic framework based on three enablers to facilitate collective AI adoption across Research Performing Organisations (RPOs) and Research Infrastructures (RIs):
- AI-Ready Infrastructure: This involves creating a "continuum" across three layers:
- Compute: Federating HTC, Cloud, Container, and HPC resources (including GPUs across seven sites) to allow "frugal AI"—running workloads on the most energy-efficient or appropriate hardware.
- Data: Projects like RI-SCALE and EOSC Data Commons create "Data Exploitation Platforms" that stage data from long-term repositories directly to AI-ready compute environments.
- Application: Services like FedLLM (a federated inference service for Large Language Models) and AI Lab provide user-friendly environments for model customization.
- Human Skills: Establishing "advisors for adoption" (through the Fluid-AI project) and technical training pathways (via RItrain-DIGIT) to strengthen the link between domain researchers and compute centers.
- Trust in AI: Developing tools for provenance capture, model versioning, and "human checkpoints" to validate intermediate results in autonomous AI workflows.
Datasets
While this is a strategy paper rather than a benchmark study, it references several domain-specific datasets and use cases that informed the strategy:
- Physics: Radio astronomy, particle detectors, and gravitational wave data (via the interTwin project).
- Environmental Science: Flood, cyclone, and wildfire modeling.
- Aquatic Research: 17 use cases involving image segmentation for plankton taxonomy, marine litter assessment, and satellite-based oil spill tracking (via the iMagine project).
- Health: Cancer images (via the EUCAIM project).
Key Insights and Ideas
- The Division of Labour: A core insight is the emerging "specialization-based interconnectivity" in Europe. The authors argue that while EuroHPC AI Factories should handle massive model training, EGI is best suited for the customization and long-term operation of those models close to the researchers' data.
- From Training to Customization: The focus has shifted. A few GPUs can now do for domain adaptation (via parameter-efficient fine-tuning) what once required a massive cluster.
- Agentic AI: The paper highlights the rise of AI "agents" that plan and execute long-running tasks. This necessitates a shift in infrastructure to support autonomous workflows while maintaining strict trust and cost controls.
- Frugal AI: There is a growing emphasis on measuring and reducing the environmental footprint of AI by routing jobs to the most appropriate, energy-efficient resources in the federation.
Key Results
- Operational Success: The iMagine project successfully deployed an AI model for the OBSEA observatory (EMSO ERIC) that has been detecting and classifying fish species in live video streams since August 2023.
- FedLLM: EGI has already launched a federated LLM service offering six different models, including the 2.8-trillion-parameter Kimi K3, harmonized through an OpenAI-compatible API.
- Software Toolkits: The development of the
itwinaiPython library, which allows users to scale AI workloads from laptops to supercomputers with minimal overhead.
Limitations and Future Work
- Sustainability: The primary challenge is moving these initiatives from temporary, project-based funding (Horizon Europe) into the core, long-term institutional operations of EGI members.
- Business Models: There is currently no sustainable business model for cross-national cost-recovery when a researcher in one country uses expensive GPU resources in another.
- Trust Implementation: While the need for trust is identified, the specific tools for auditing non-deterministic, frequently updated models are still in the early stages of development.
- AI for Operations: The authors note that using AI to improve EGI’s own infrastructure (e.g., cybersecurity, maintenance) is a promising area but remains outside the scope of this specific paper.
Summary updated