AMD has agreed to pay $8.2 billion — entirely in stock, not cash — to acquire World Labs, the world-model startup founded by Fei-Fei Li, Justin Johnson, Ben Mildenhall, and Christoph Lassner. The price represents a premium of roughly 75 percent over the company's $5 billion valuation at its February Series B, when Nvidia and AMD both participated in a $1 billion round.

The move follows Nvidia's $12.9 billion acquisition of Hugging Face earlier this month, which removed from AMD's reach the open-weight model hub hosting over 3 million models, 1 million datasets, and 1.4 million interactive spaces. With Hugging Face under Nvidia's control, AMD needed another route to model-building credibility — and a way to counter Nvidia, which has built its own models since launching the Nemotron3 8B family in November 2023.

What does AMD actually get?

World Labs builds so-called world models — systems designed to understand the physics of the world rather than generate text and images from a corpus of human knowledge. The company publicly launched its Marble world model in November 2025 and followed with a World API cloud-access version of Marble in January. Its follow-on Atlas engine, an "omni world model" that is technically a multimodal autoregressive diffusion transformer, entered early access with selected customers in early September and will feed into future Marble versions.

As much as product, the deal buys a team — a parallel, the source notes, to AMD's $40 billion Xilinx acquisition. Li, World Labs' chief executive, becomes chief scientist at AMD and will report directly to CEO Lisa Su, presumably with a mandate broader than world models for robots. Johnson and Mildenhall will continue running World Labs inside AMD. Notably, co-founder Lassner does not appear in AMD's announcement, so the team may arrive minus one. Neither company has disclosed World Labs' headcount, revenue, or customer count.

Who is Fei-Fei Li?

Li holds a physics bachelor's from Princeton and a PhD in electrical and electronics engineering from Caltech, and ranks among the two dozen top AI figures. She served as Google's chief scientist of AI/ML starting in January 2017 and left in October 2018 amid internal uproar over Google's "Project Maven" collaboration with the US Department of Defense. She has directed the Stanford AI Lab since 2014, co-directed the Stanford Institute for Human-Centered Artificial Intelligence, and taught computer science at Stanford for nearly two decades.

Her co-founders bring comparable depth:

• Justin Johnson: Caltech BS in computer science, Stanford PhD, AI research scientist at Meta Platforms from 2016 through December 2023, now adjunct professor at the University of Michigan.
• Ben Mildenhall: Stanford BS, Berkeley PhD, several years as a research scientist at Google working on computer vision and generative 3D models.
• Christoph Lassner: informatik degrees from the University of Tübingen, applied scientist at Amazon from September 2017, then research scientist and manager at Meta, and a year at Epic Games building 3D game worlds.

World Labs raised $1.23 billion across two funding rounds, with angel investors including Jeff Dean and Geoff Hinton of Google, Marc Benioff of Salesforce, Reid Hoffman of LinkedIn, and Eric Schmidt of Sun Microsystems and Google.

Why stock instead of cash?

AMD, which recently broke through a $1 trillion market capitalization, cannot afford to spend $8.2 billion in real cash given the investments it must make and the chip capacity and ingredients it must pay for in advance to build CPUs and GPUs. Paying in shares keeps the balance sheet intact.

What comes next?

Competition in world models is crowded. Tencent fields HunyuanWorld, Google has Genie 3, startups Runway, Odyssey, and Decart are building their own, Nvidia offers its Cosmos world models, and Yann LeCun — former head of AI research at Meta — is at work in the space with his AMI Labs startup.

AMD now has aspirations to push its compute engines into robotics, as Nvidia already does. The strategic logic is straightforward: if Nvidia can compete with its own customers in model building, AMD can too — and arguably must, to ensure open-source frontier models exist and to position itself as more open than Nvidia. Nvidia's Nemotron GenAI models are open-weight and Cosmos is free and open source, but the 20 million hours of video used to train Cosmos remain closed, most likely because the footage was scraped from the Internet and Nvidia does not own it.

The open question is whether Li and Su will go further: build an open-source GenAI model and open it fully — source code, weights, and training data — to prove AMD's openness claim. The World Labs team can do exactly that while maintaining and expanding its world models.