AMD announced its acquisition of World Labs in a deal valued at $8.2 billion, bringing into its plans a company specializing in developing deep-learning models capable of building a more realistic understanding of the physical world. World Labs founder and Stanford University computer scientist Fei-Fei Li is expected to join AMD as deputy chief executive officer and chief scientist.
The deal comes as chipmakers seek to expand competition from selling processors to building integrated ecosystems for running and training artificial intelligence models. World Labs said that developing artificial intelligence requires close cooperation among model research, systems, and computing capabilities, while AMD believes that understanding the advanced workloads the company is developing will affect its chip roadmap.
From Language Models to Simulating Reality
Li founded World Labs in 2024 with the aim of developing models that understand the physical world more deeply, based on her view that achieving true artificial general intelligence requires grounding in physics and the ability to understand and analyze data beyond text. Li had gained widespread recognition in computer vision through the creation of the ImageNet database and the competitions it inspired.
World Labs introduced its first products, Marble, as a tool for creating entertainment experiences and simulation environments that can also be used to train robots. The term “world models” remains broad and not precisely defined; it may refer to models that understand visual inputs, or to systems that create and maintain high-fidelity simulations of reality.
What Changes in Practice?
AMD believes that integrating World Labs’ research expertise into its hardware roadmap could help it design chips and systems better suited to workloads associated with these models. The two companies entered into a partnership last year to improve inference and training processes, and Li appeared as a guest in AMD’s presentation at CES earlier this year.
These models are particularly important in the development of robotics, including autonomous vehicles, industrial robots, and versatile humanoid robots. The shortage of real-world data needed to train robots makes the synthetic data generated by world models a potential component in building broader training environments.
Direct Competition with the Nvidia Ecosystem
The deal could give AMD a stronger presence in the software and model layer surrounding artificial intelligence chips, an area where Nvidia competes through open-weight world models such as Cosmos. By contrast, AMD had made available general-purpose models based on text and video, without an equivalent announced offering for world models.
Li wrote that the acquisition decision aims to expand the scope of World Labs’ technological achievements and bring its work closer to hardware. However, the deal is still awaiting regulatory approvals and is expected to close before the end of 2026. According to the available information, the announcement does not clarify the details of the plan to integrate Marble or the agreement’s other financial terms.