AMD’s $8.2bn bet on virtual worlds and the woman behind ImageNet

Advanced Micro Devices has decided that the next battle in artificial intelligence is not only about bigger models, but about the environments in which those models learn to act. Its $8.2 billion purchase of World Labs is a clear sign that the chipmaker wants a front-row seat in a field that could shape the next generation of hardware, software and robotics. For a company chasing Nvidia, that is hardly a trivial hobby.

The deal also brings in Fei-Fei Li, one of the best-known names in AI research. According to a statement from Li, she will become AMD’s executive vice president and chief scientist once the acquisition closes. Her reputation rests in part on ImageNet, the large labelled image collection that helped accelerate computer vision, and on her work in what she calls spatial intelligence, the ability of AI to understand and work with three-dimensional spaces.

World Labs, which she co-founded in 2024, builds world models. These systems are designed to understand how objects and spaces behave, then generate interactive 3D environments from text, images and video. One obvious use is training robots in virtual settings before they are sent into the physical world. The less glamorous part is that accurately simulating reality is still hard, which tends to complicate even the most enthusiastic pitch.

AMD and World Labs have already been working together since last year to train and run the company’s models on AMD chips. That makes the acquisition look less like a leap into the unknown than a decision to bring a promising customer closer to the factory floor. AMD says the point is to understand what emerging AI systems need and to shape future hardware and software around those demands. Chief executive Lisa Su put it plainly in the company’s announcement: building the compute platforms for the next generation of AI requires a deep understanding of how models are evolving.

The field is already crowded, which is usually what happens when investors discover a new phrase. Google DeepMind’s Genie 3 creates virtual environments users can explore. Meta’s V-JEPA 2 focuses on understanding video, predicting outcomes and planning actions. Nvidia has its Cosmos models for robotics and related uses. In other words, everyone wants to own the map before the territory even exists.

For AMD, the attraction is obvious enough. World models require serious computing power, and the company is betting that owning more of the stack will help it anticipate where demand goes next. Li said in a blog post that there is now tangible proof of the possibilities and that the task is to accelerate what comes next. She also argued that doing so requires scaling efforts, widening reach and getting closer to hardware. In the semiconductor business, that is usually where the money lives.

Written by Thorben Thiede thorben.thiede@alpineweekly.com