Summary
Fei-Fei Li says World Labs is joining AMD, where she will become executive vice president and chief scientist. The announcement links the startup’s spatial AI research with AMD’s hardware and software work.
Fei-Fei Li announced on September 28, 2026, that World Labs is joining AMD. Li, who co-founded the spatial-intelligence startup, said she will become AMD’s executive vice president and chief scientist, working with CEO Lisa Su and her team.
The announcement describes the move as a way to bring World Labs’ AI research closer to the hardware used to train and run models. Li said the companies had already begun a technical partnership in 2025, focused on training and inference optimisation on AMD GPUs. The post describes World Labs as joining AMD but does not give a transaction price or legal structure.
Why AMD and World Labs are combining their work
World Labs was founded in 2024 by Li and researchers Ben Mildenhall and Justin Johnson. Its stated focus is spatial intelligence: AI systems that can represent and reason about the structure and behaviour of the physical world, rather than working only with language or other digital inputs.
That focus has applications in areas such as robotics, design and scientific work, where a system may need to interpret a scene or predict how it would look from another viewpoint. Li’s explanation for joining AMD is that advancing such systems requires scaling research and connecting models more closely to computing hardware. The strategic idea is to develop AI across hardware, software, models and data platforms; the announcement presents this as a shared goal, not a report of completed products built through the combination.
What World Labs says it has built
Li highlighted Atlas, which she described as an “omni” model architecture trained from scratch to predict a new camera view from input 2D images. The company says Atlas combines generative models with multiview geometry and performs better than results from specialised models on the task.
The technical problem is related to sparse reconstruction: recovering a scene’s 3D structure when only a limited set of images or camera views is available. Multiview geometry uses information from different views to infer spatial relationships; a model that can also generate plausible new views could be useful in applications that need a representation of a scene, rather than just a description of it. Li said World Labs has seen interest in using Atlas for robotics reinforcement-learning environments, scene generation and reconstruction for real estate, design and construction.
Li also said World Labs is developing robotics-simulation capabilities following its acquisition of SceniX. The announcement describes these projects as part of a wider effort to build models for physical-world tasks. It does not provide technical evaluation details for Atlas or specify how World Labs’ models will be integrated into AMD products. For now, the stated change is organisational and strategic: World Labs’ team and research are moving under AMD, with Li taking a senior research leadership role.