Semiconductor company AMD has announced an agreement to acquire World Labs for $8.2 billion. World Labs is a developer focused on deep learning models that interpret physical reality and build spatial intelligence systems.
The transaction aims to integrate World Labs research into the chip development process at AMD to better support generative artificial intelligence and robotic platforms. Both organizations stated that artificial intelligence development required close collaboration across compute, systems, and model research.
Founder Fei-Fei Li will join AMD as executive vice president and chief scientist following the completion of the deal.
Li co-founded World Labs in 2024 alongside Justin Johnson, Ben Mildenhall, and Christoph Lassner to build deep learning models with an understanding of the physical world, arguing that general intelligence requires a grounding in physics and reasoning capabilities beyond text. Li previously built the ImageNet database and the associated artificial intelligence competitions while working as a computer science professor at Stanford.
Chip Roadmap And Products
AMD stated that understanding frontier workloads from World Labs will shape its chip-making roadmap. World Labs develops what are known as world models, designed to perceive, generate, reason about, and interact with virtual and physical environments.
The first product from World Labs is called Marble, and it serves as a tool for creating entertainment experiences as well as simulated environments for training robots. Marble can create spatially consistent, high-fidelity and persistent 3D worlds from inputs including text, images, video, and 3D layouts.
The acquisition expands AMD’s push into artificial intelligence applications involving spatial intelligence and physical AI, alongside its competition with Nvidia in building an ecosystem for artificial intelligence-specific chips. Nvidia offers open-weight world models such as Cosmos, while World Labs focuses on spatial intelligence and world-model technology.
World models are increasingly being explored for deploying generative artificial intelligence on robotic platforms, ranging from industrial robots and autonomous vehicles to general-purpose humanoids. Because of a shortage of real-world data for training general-purpose robots, synthetic data from world models is viewed as key for companies pursuing robotics initiatives.
The two companies previously formed an inference optimization-and-training partnership, and Li appeared as a guest at a presentation held by AMD. The transaction is expected to conclude by the end of the year, pending standard regulatory review and approval.
