AMD Acquires World Labs for $8.2 Billion in Stock

AMD announces the acquisition of World Labs for approximately $8.2 billion, in a fully stock-based transaction. The deal will see Fei-Fei Li take the helm of AMD's research and is part of a strategy to integrate spatial AI models more closely with hardware, amid competition with Nvidia.

Competition: Nvidia, Hugging Face, and Valuations

For several years, Nvidia has combined its chips with models and services, released open-weight models, and created its own solutions to foster the development of an ecosystem around its processors. This approach continued with the recent acquisition of the AI platform Hugging Face for $13 billion. In this context, AMD is valued at around $1 trillion, while Nvidia reaches approximately $5 trillion. On the target side, World Labs raised $230 million in 2024 from investors including Andreessen Horowitz, Nvidia, and AMD, followed by $1 billion in early 2025 with support from Autodesk. At that time, its value was estimated at around $5 billion. The announced price of $8.2 billion implies a multiplication of this value within a few months.

AMD's Goal: Understanding AI Workloads to Guide Hardware

AMD presents the acquisition as a means to gain a deeper understanding of how AI workloads are constructed in order to inform its future hardware roadmaps. "The more you understand from end to end, the better the system you are going to build," said Lisa Su, indicating that this logic drives the purchase of World Labs. The deal, announced for approximately $8.2 billion and fully in stock according to AMD, still requires regulatory approvals and aims for completion by the end of 2026.

Scientific Governance: Fei-Fei Li to Lead Research

Fei-Fei Li will join AMD as Executive Vice President and Chief Scientist. She will lead cutting-edge AI research and report directly to CEO Lisa Su. She argues that expanding research requires closer integration with hardware, stating that "without targeted hardware efforts, AI is hampered in efficiency."

What World Labs Does: Atlas and World Models

World Labs, founded in early 2024 in San Francisco by Fei-Fei Li, Ben Mildenhall, and Justin Johnson, has around 70 employees and focuses on spatial intelligence. The company develops world models aimed at problems that language models do not address, particularly understanding physics and spatial structures. In early September, it unveiled Atlas, a first model capable of generating, reconstructing, and simulating complete spatial scenes from just a few images. Fei-Fei Li synthesizes her vision by stating that the universe consists of real entities rather than words, and argues that multimodal language models fail at spatial tasks: they perform no better than chance at assessing distances, directions, and dimensions, and struggle to mentally rotate objects or anticipate basic physical laws. According to her, sectors such as science, entertainment, and robotics require models capable of reasoning about the structure and functioning of the physical world. In the research field, Demis Hassabis believes these models represent an essential phase toward AGI. A group led by Peking University has introduced OpenWorldLib, a standardized definition that imposes three capabilities: perceiving the environment, acting upon it, and remembering previous interactions.