AMD is spending roughly $8.2 billion of its own stock to hire a model research team. The chipmaker said on Monday it had signed a definitive agreement to buy World Labs, the San Francisco lab led by AI pioneer Fei-Fei Li, with Li stepping into a newly created role as AMD's executive vice president and chief scientist. She will report to chair and chief executive Lisa Su. AMD is targeting a close before the end of 2026, once regulators sign off.
Key takeaways
- AMD is paying roughly $8.2 billion in stock for World Labs, a lab founded in early 2024 that was valued at $1 billion within months of launching.
- Fei-Fei Li becomes AMD's executive vice president and chief scientist, reporting to Lisa Su, while the World Labs team continues its model research.
- It is AMD's second AI acquisition in a matter of weeks, following its agreement to buy inference-chip startup Taalas.
What AMD is actually buying
World Labs works on what researchers call world models. Feed one a prompt, a photograph or a video clip and it does not answer in prose β it builds a navigable 3D scene and keeps track of how the objects in it should behave. The lab has also pushed the same machinery toward robot training, where a simulated room is cheaper to iterate on than a real one.
That is a different engineering problem from the one a large language model solves. An LLM ranks candidate next tokens. A world model has to hold geometry, occlusion and physics in memory, which changes what the underlying silicon is asked to do.
The lab moved fast for its age. Founded in early 2024, it hit a $1 billion valuation within months, The Verge reported, and shipped its first commercial product in 2025: Marble, which turns a written prompt into an explorable 3D world. In her note on the deal, Li highlighted a second system, Atlas, which she credits with cracking new camera view prediction β inferring what a scene looks like from an angle the model was never shown.
Why a chip company wants model researchers
Su's public argument is that you cannot design compute for the next model generation without knowing how models are changing, and that World Labs supplies exactly that signal as AI spreads into reasoning, robotics and simulation.
The unstated half is Nvidia. The Register notes that AMD's software engineering headcount is tiny beside its rival's, and that Nvidia's moat has been usability as much as throughput β the frameworks that let a customer actually deploy the card they bought. Buying a research lab is an unusual way to close a tooling gap, but it does give AMD a demanding first-party customer for its own GPU stack, and someone senior enough to tell the hardware roadmap what it got wrong.
The second AI acquisition in weeks
This is not a one-off. AMD recently agreed to buy Taalas, whose approach bakes model weights straight into silicon for throughput above 17,000 tokens a second. Taalas was a bet on how to run a model; World Labs is a bet on which models will be worth running. Taken together they read as AMD deciding it can no longer wait for the market to tell it what to build.
What Li says changes
Li has spent years making a narrower argument than her critics assume. She is not against language models; she thinks text alone is an incomplete training signal, because the world is made of objects rather than words, and problems in science and robotics require reasoning about physical structure. Her stated reason for selling is prosaic: scale. Hardware access, she wrote, is what her research now needs most.
Outlook
An all-stock deal spreads the cost across AMD's shareholders instead of draining its balance sheet, which matters in a year when every AI buyer is also funding data centres. The bet itself is legible enough. If capital keeps flowing to text models, $8.2 billion for spatial intelligence will look early at best. If language models stall where sceptics expect them to, AMD will own the team that defines its next workload.
FAQ
How much is AMD paying for World Labs, and in what form?
Approximately $8.2 billion, structured as an all-stock transaction rather than a cash purchase. AMD expects the deal to close by the end of 2026, pending regulatory approvals and the usual closing conditions.
What happens to Fei-Fei Li and the World Labs team?
Li becomes executive vice president and chief scientist at AMD, reporting to chair and CEO Lisa Su. AMD says the World Labs team will keep working on advanced AI model research after the transaction closes rather than being folded into product engineering.
What is a world model, and how is it different from an LLM?
A world model builds and simulates 3D environments from text, image or video input, then reasons about how things inside them move and occlude one another. An LLM predicts the next token of text. The distinction matters for robotics and simulation, where a system has to model space rather than describe it.
