Nvidia exec says robotics lacks an internet-wide dataset for physical AI

At TechCrunch Disrupt 2026, Nvidia’s Les Karpas highlighted a core limitation facing the robotics industry: the absence of an internet-wide dataset for physical AI. This gap mirrors the early days of large language models, which lacked the massive, diverse datasets that later enabled breakthroughs like ChatGPT. Karpas, Nvidia Inception’s Global Head of Physical AI, noted that robotics has yet to experience its own ‘ChatGPT moment’ due to this data scarcity.

Karpas pointed to the role of self-driving fleets, such as Waymo’s, which accumulated years of road data to build foundational datasets. Robotics, by contrast, lacks a similarly scalable and shareable dataset that spans diverse environments and tasks. He emphasized that bridging this gap—between digital simulation and physical reality—requires coordinated efforts across startups, researchers, and industry players.

The challenge is not just data collection but also standardization and sharing. Karpas described how startups and companies are attempting to manufacture scale through synthetic data, simulation, and foundation models trained across multiple robot forms. These efforts aim to create the equivalent of an ‘internet for robotics,’ enabling generalization beyond narrow, task-specific applications.

Karpas’s remarks were part of a session on the Real World AI Stage at Disrupt 2026, where he was joined by founders from Shield AI, Colossal Biosciences, FieldAI, and Foxglove. The discussion underscored the interdisciplinary nature of the problem, requiring expertise in robotics, AI, manufacturing, and venture ecosystems.

Nvidia’s involvement reflects its broader push into physical AI, as the company coordinates with startups across robotics, automotive, manufacturing, and smart cities through its Inception program. Karpas’s career—spanning roles at Stanley Black & Decker, iRobot, and other firms—underscores the cross-disciplinary skills needed to tackle this bottleneck.