AMD said Monday it will acquire World Labs, the "physical AI" startup co-founded by Stanford computer scientist Fei-Fei Li, in an all-stock deal valued at $8.2 billion — the chipmaker's second-largest acquisition on record, trailing only its roughly $50 billion purchase of Xilinx in 2022.
The transaction, expected to close before the end of the year pending regulatory approval, will bring Li into AMD as executive vice president and chief scientist, reporting directly to chief executive Lisa Su. World Labs, founded in 2024, builds deep learning models that interpret and simulate three-dimensional physical space — a category researchers call "world models," distinct from the text- and image-generating systems that dominate most consumer AI products today.
Betting on the next phase of AI
World Labs' flagship product, Marble, generates simulated 3D environments used to train robots and to build interactive entertainment. The roughly two-year-old startup had raised close to $1 billion in prior funding, according to people familiar with its financing, meaning Monday's price tag represents a steep premium reflecting how central world models have become to AMD's chip strategy.
"Together, we'll combine World Labs' deep expertise in AI and world models with AMD's compute leadership to power the future of AI and strengthen the open AI ecosystem," Su said in a statement announcing the deal.
AMD and World Labs were not strangers before the acquisition. The two companies formed a partnership last year to optimize inference and training workloads on AMD's Instinct accelerators, and executives said those ties had deepened steadily since. Li framed the sale as an extension of that relationship rather than a pivot. "To do this requires scaling our efforts, widening our reach, and getting closer to the hardware," she said, explaining World Labs' rationale for joining AMD.
Chasing Nvidia's ecosystem
The purchase is widely read as an effort to close a gap with Nvidia, which has spent the past two years building a software and model ecosystem — anchored by its open-weight Cosmos world-model suite — designed to keep developers tied to its GPUs. AMD, by contrast, has so far offered the public only text- and video-based models, leaving it without a real answer to Nvidia's push into robotics, autonomous vehicles and humanoid platforms, all of which depend on systems that can reason about physical space rather than just language or pixels.
By folding Li's research group directly into its chip organization, AMD is betting that owning a frontier world-model lab will sharpen its hardware roadmap the way internal model development has sharpened Nvidia's — giving engineers direct visibility into the workloads that will strain memory bandwidth, interconnects and inference architectures over the next several years.
Li has long been one of the most recognizable figures in artificial intelligence. A Stanford computer science professor, she built the ImageNet database that helped ignite the modern deep-learning boom in the early 2010s and later led Google Cloud's AI division before returning to academia and, in 2024, co-founding World Labs with several former students and collaborators. Her arrival gives AMD a research figurehead with a public profile comparable to rivals' chief scientists.
Shares of AMD were roughly flat in after-hours trading following the announcement, according to Fortune, suggesting investors are still weighing whether the bet on spatial-intelligence research will translate into near-term chip sales. The all-stock structure limits AMD's cash outlay but dilutes existing shareholders, and analysts questioned how quickly a research-stage world-model lab can be converted into product roadmaps.
The deal still requires antitrust clearance in the United States and likely elsewhere, given AMD's size and World Labs' ties to robotics and automotive customers. If it closes as expected by year-end, it would rank among the largest AI-research acquisitions of 2026 and reinforce a broader trend of chipmakers buying their way into frontier model research rather than simply selling silicon to the labs that build it.
