The deal
AMD has agreed to buy World Labs, the San Francisco research lab founded by Fei-Fei Li, in an all-stock transaction the company values at about $8.2bn. AMD expects it to close by the end of 2026, subject to regulatory approval and the usual conditions.
When it closes, Li joins AMD as executive vice president and chief scientist, reporting directly to chair and chief executive Lisa Su. The World Labs team stays on model research. It is AMD’s second-largest acquisition, behind the roughly $50bn it paid for Xilinx.
What World Labs builds
AMD describes World Labs as working on spatial-intelligence models: systems that generate, reconstruct and simulate interactive 3D environments from text, image and video inputs, plus technology for robotic learning and simulation.
That is a different problem from the one most language models solve. A model that writes a paragraph about a kitchen does not have to know where the kettle is relative to the sink. A model that generates an environment a robot can then be trained inside does. The distance between those two things is roughly the distance between the current AI market and the one every chip company is now positioning for.
Li’s standing here is not a matter of title. She built ImageNet, the labelled image dataset whose annual competition is generally credited with starting the deep-learning era in computer vision. Putting that person in an AMD reporting line, rather than an advisory seat, is the part of this announcement that is hard to buy with money alone.
Why a chip company buys a model lab
AMD’s stated reason is that it needs to see where workloads are going. “Building the compute platforms for the next generation of AI requires a deep understanding of how models are evolving,” Su said in the announcement. Li’s side of it was that AMD gives the team “the resources and engineering depth to accelerate our research”.
Read commercially, this is about Nvidia. Nvidia’s advantage has never been only the silicon; it is that the company sells simulation and robotics software alongside it, so that customers building physical systems find a full stack waiting. AMD has the accelerators and, since Xilinx, the adaptive silicon. What it has not had is an in-house research group whose output shapes the roadmap and doubles as a reason to buy AMD rather than rent Nvidia.
Buying one has an obvious cost. A lab absorbed into a chipmaker tends to lose the people who joined it because it was not one, and the retention of the World Labs team is the variable AMD cannot state in a press release.
What to watch
Three things. Whether regulators in the US, the EU and China treat an all-stock purchase of a pre-revenue research lab as routine — recent chip deals have not always cleared quietly. Whether World Labs’ models continue to be released or published in any form once they sit inside a public company’s roadmap. And whether AMD’s next data-centre roadmap update, the first one Li could plausibly influence, says anything concrete about simulation and robotics workloads rather than repeating the language in this release.
