How NVIDIA Builds Open Data for AI

The quality and accessibility of training data are critical for AI progress, yet building high-quality datasets remains a major bottleneck. To address this, NVIDIA releases permissively licensed open datasets, tools, and training recipes to accelerate development and improve evaluation across the ecosystem. The company has shared over 2 petabytes of data spanning domains like robotics, autonomous systems, and biology. Specific examples include the Physical AI Collection for robotics, the Nemotron Personas collection of synthetic demographic data, and the SPEED-Bench benchmark for evaluating model performance.