From Silicon to Steel: Powering Physical AI and Robotics with AI Factories

Dennis Hoppe, head of Converged Computing at HLRS, opens this session on physical AI and robotics with a short framing talk ahead of two invited expert presentations. He observes that the field's emphasis has shifted from ever-larger LLM training toward inference, and argues the next frontier is moving AI from the digital world, where the worst failure is a hallucinated line of text, into the physical world of robotics and autonomous driving, where consequences are tangible. This transition imposes new constraints: edge devices demand smaller models, low latency, and low energy consumption, creating a gap between large-scale training on massive GPU installations and fast inference on specialized embedded hardware. Hoppe outlines a physical-digital data loop with four repeated steps: simulation, AI model training, deployment from the digital domain into physical systems, and collection of metrics from deployed robots and vehicles to refine simulations and retraining, forming a continuous feedback cycle. He notes that while the community has expertise in each step individually, orchestrating the full workflow seamlessly remains an open challenge. As a potential IT backbone for physical AI in Europe, he points to the EuroHPC AI factories, with 19 established across EU member states, including Hammerhai at HLRS in Stuttgart, an AI-optimized supercomputer entering operation at the end of the year that is intended to support these orchestrated workflows bridging AI and the physical world.