AMD will acquire World Labs for $8.2 billion

AMD has agreed to purchase World Labs in an $8.2 billion transaction, marking a major consolidation in the physical artificial intelligence sector. The acquisition, which is expected to close by the end of this year pending regulatory approval, will integrate the startup's deep learning expertise directly into AMD's hardware ecosystem. As part of the agreement, World Labs founder and Stanford computer science professor Fei-Fei Li will join AMD as executive vice president and chief scientist. Li, widely recognized for her pioneering work on the ImageNet database, established World Labs in 2024 to build models capable of understanding and reasoning about the physical world.

World Labs focuses on developing world models that simulate physical reality, going beyond standard text-based generative AI. The startup's first product, Marble, serves as a tool for designing entertainment experiences and generating simulated environments for training robots. These models are highly valued for their ability to generate synthetic training data, which is crucial for deploying generative AI on physical platforms like autonomous vehicles, industrial machinery, and humanoid robots where real-world data is scarce.

For AI practitioners and developers, this acquisition signals a tighter integration between hardware design and physical AI workloads. AMD and World Labs previously established an inference optimization-and-training partnership last year, and this merger will allow World Labs' complex workloads to directly shape AMD's future chip-making roadmap. By bringing these capabilities in-house, AMD aims to build a more competitive ecosystem for AI-specific silicon to challenge Nvidia, which already offers its own suite of open-weight world models under the Cosmos name.

In a public announcement, Li explained that the acquisition would help scale the startup's research by "getting closer to the hardware." This integration of model research and silicon design is expected to provide developers with hardware that is natively optimized for spatial intelligence and physical simulation workloads.