AMD has reached an agreement to bring spatial intelligence startup World Labs directly into its hardware and software operations. Founded by computer vision pioneer Fei Fei Li, the lab specializes in training foundation models that interpret the physical world rather than just digital text. As part of the move, Li will join AMD as Executive Vice President and Chief Scientist, reporting directly to chief executive Lisa Su.
Most artificial intelligence models today focus almost entirely on predicting words. World Labs started in 2024 to break away from that limitation. Cofounders Fei Fei Li, Ben Mildenhall, and Justin Johnson built the venture on the belief that text alone cannot solve complex problems in robotics, design, or medicine. Machines need to calculate how physical objects occupy 3D space. To speed up work on physical simulation, the lab previously acquired robotics simulation specialist SceniX.
The company made a major technical breakthrough with its Atlas model architecture. Atlas solves a classic puzzle in computer vision known as sparse reconstruction. Instead of guessing the next word in a chat prompt, the model uses standard 2D pictures to predict entirely new camera viewpoints. That capability has immediate practical uses in real estate modeling, robotic navigation, and surgical therapy tools.
— Fei-Fei Li (@drfeifei) September 28, 2026
Advanced software models eventually stall without dedicated computing silicon. In her public announcement, Li explained why joining a semiconductor maker became the logical next step for the research group.
Without having a focused hardware effort, AI is hobbled in efficiency. And scale. And for our purposes, remains trapped in the digital world.
The acquisition cements a partnership that was already underway. Lisa Su backed World Labs as an early investor, and engineering teams from both groups spent the past year optimizing model training and inference on AMD graphics processors. By bringing the startup in house, AMD secures top tier research talent to help shape its future processors around spatial workloads. The combined team plans to develop open models and tools spanning everything from raw silicon to end user software platforms.
