The two companies said on Monday that AMD will pay $8.2B for World Labs, the physical AI lab Fei-Fei Li started in 2024.
One of computer vision’s best-known researchers ends up inside a chipmaker. Li built the ImageNet database and teaches at Stanford; on completion of the deal she takes the titles of executive vice president and chief scientist at AMD.
Models that reason about physical reality, rather than text alone, are what World Labs makes. Marble, its first product, turns out simulated environments for entertainment and for training robots. An inference optimization and training partnership struck last year already had the two working side by side.
Supplying robot hardware is a race AMD is still losing to Nvidia. Open-weight world models already arrive from Nvidia through its Cosmos line, while AMD’s public output has gone no further than text and video. A world model lab of its own gives AMD a story to tell around humanoid platforms, industrial machines and autonomous vehicles.
Robots need training data that the real world cannot supply in quantity, which is the case for world models. Synthetic environments generate that experience without putting prototypes at risk on factory floors.
Li cast the sale as a scaling call: pushing the lab’s work beyond research, she argued, requires sitting closer to the hardware.
Regulators have yet to clear it, and the companies expect the purchase to wrap up before the year ends.
