AMD to Acquire Fei-Fei Li’s World Labs for $8.2 Billion
AMD will buy Fei-Fei Li's World Labs for $8.2 billion, adding Marble and world model research to challenge Nvidia in AI chips and robotics by year end.
Summary
AMD agreed on September 28, 2026, to acquire World Labs for $8.2 billion, subject to regulatory approval, with closing expected before year end. Founder Fei-Fei Li will become AMD executive vice president and chief scientist. The companies began an inference optimization and training partnership in 2025, and Li appeared at AMD’s CES presentation in 2026. AMD expects World Labs’ frontier workloads to shape its chip roadmap, while Li says closer access to hardware can scale research beyond the laboratory.
Li, a Stanford computer science professor who pioneered computer vision through ImageNet and its competitions, founded World Labs in 2024 to develop deep learning models grounded in physics and capable of reasoning beyond text. Its first product, Marble, creates entertainment experiences and simulated environments for robot training. The deal strengthens AMD against Nvidia, whose open weight Cosmos models already target world modeling, while AMD has publicly released only text and video models. Synthetic environments could address scarce real world training data for autonomous vehicles, industrial robots and general purpose humanoids pursued by companies including Tesla and Figure, although “world model” still spans loosely defined technologies from visual language models to persistent simulations of reality.
Positives
- $8.2 billion gives World Labs the resources and hardware access to scale its physical intelligence research.
- Fei-Fei Li will become AMD executive vice president and chief scientist, adding an ImageNet pioneer to its leadership.
- Marble gives AMD technology for entertainment creation and simulated robot training environments.
- World Labs’ frontier workloads will directly inform AMD’s future chip development roadmap.
- Synthetic world data could reduce dependence on scarce real world training data for robots and autonomous vehicles.
Risks & concerns
- Regulatory approval remains outstanding before the acquisition can close by the end of 2026.
- Nvidia already offers open weight Cosmos world models, while AMD has released only text and video models publicly.
- “World model” remains loosely defined, spanning visual language systems through high fidelity, persistent reality simulations.
- General purpose robot developers still face a shortage of useful real world training data.
- World Labs says progress requires close coordination across research, systems and compute, creating a significant integration challenge.