Adam Grzywaczewski

Applied scientist by background, with close to twenty years in AI and machine learning. For the last decade I have been at NVIDIA, working at the intersection of research and production — both hands-on and as the leader of a ~30-person organisation that turns frontier research into production systems at scale. NVIDIA has been a unique vantage point — a front-row seat to the AI revolution and the infrastructure that made it possible. On the research side, I focus on technologies that enable the scalability of AI systems. That has meant contributing to the design and bring-up of some of the world's first AI systems (DGX-1 and DGX Station — still have a prototype in my home office) and AI supercomputers like the UK JADE system. It has also meant working on some of the earliest FP16, BF16, FP8 and now NVFP4 training regimes, on algorithms like LARC and tools like Horovod and Megatron-LM to enable large-batch training, and today on the challenges of efficient MoE training, distributed RL and Physical AI. Throughout, I have worked directly with networking and storage engineering teams — translating AI requirements into hardware and software design and contributing to NVIDIA's reference architectures. The organisation I have built covers distributed training, training resiliency, model compression, inference and — more recently — Robotics, Physical AI and world models. The team carries extensive research pedigree: contributions to reduced precision training, GPTQ, zero-bubble pipeline parallelism, Nemotron, world models (Cosmos) and AI in structural biology. We collaborate through joint applied research with organisations across the full European AI landscape: AI-native companies, frontier research labs, academic institutions and large-scale HPC centres. Beyond the technical work, I have played a broader role at NVIDIA — helping establish the Deep Learning I