CapitalUnited States
AMD Buys World Labs For $8.2 Billion To Get AI Model Cred
Useful
CapitalUnited States
As you might expect, the Hugging Face open weight AI model distribution hub, which has over 3 million AI models, over 1 million datasets, and over 1.4 million interactive spaces that it hosts, was something that AMD had to focus on as it was trying to increase adoption of its Instinct GPU accelerators. But now Hugging Face is under control of Nvidia after a $12.9 billion acquisition by Big Green earlier this month.
And so, AMD has to do something dramatic to change the narrative as well as to become a model builder in its own right as Nvidia has been since November 2023 with the launch of its Nemotron3 8B family of models. That dramatic thing is to acquire World Labs, a maker of so-called world models that are designed to understand the physics of the world rather than the generate text and images based on the corpus of human knowledge.
AMD has agreed to pay an incredulous $8.2 billion to acquire World Labs, an AI startup that was founded in January 2024 by Fei-Fei Li, Justin Johnson, Ben Mildenhall, and Christoph Lassner in April 2024 and that has raised $1.23 billion in two rounds of funding – including angle investors Jeff Dean (Google), Geoff Hinton (Google), Marc Benioff (Salesforce), Reed Hoffman (LinkedIn), and Eric Schmidt (Sun Microsystems and Google). Nvidia and AMD were part of the $1 billion Series B funding in February of this year. At that time, World Labs was valued at around $5 billion, so AMD is paying a premium of around 75 percent to get a team that can build world models.
AMD, which has broken through a $1 trillion market capitalization recently, is paying the $8.2 billion to the owners of World Labs with stock – what we lovingly call funny money – rather than real cash, which AMD cannot afford to blow right now given the investments it must make and the chip capacity and ingredients it must pay for ahead of time to build its CPUs and GPUs.
While World Labs focuses on world models, AMD’s aspirations might be – and I would argue, should be – larger than the niche that World Labs has been attacking alongside Nvidia with its Cosmos world models.
But World Labs needs to mind the store, too, with Tencent has HunyanWorld and Google has Genie 3 in the game, and AI upstarts Runway, Odyssey, and Decart are also building world models, and World Labs did a public launch of its Marble world model in November 2025 and followed up with a World API cloud access version of Marble in January of this year. The company’s follow-on Atlas world model engine – an “omni world model” that is technically a multimodal autoregressive diffusion transformer – was launched with early access to a select number of customers in early September, which will be plugged into future Marble versions apparently. Yann LeCun, who used to head up AI research at Meta Platforms, is also up to something in the world model space with his AMI Labs startup.
However, this deal is as much about buying a team as it was buying a product, just like the $40 billion acquisition of Xilinx and the
Fei-Fei Li, who was chief executive officer at World Labs and is now chief scientist at AMD, has a bachelors degree in physics from Princeton University and a PhD in electrical and electronics engineering from Caltech and is one of the two dozen AI luminaries on Earth. Significantly, Li was chief scientist of AI/ML at Google back in January 2017, and left the company in October 2018 as Google’s “Project Maven” collaboration with the US Department of Defense caused an uproar inside of the search engine and cloud giant. Since 2014, Li has been the director of the Stanford AI Lab and was co-director of the Stanford Institute for Human-Centered Artificial Intelligence since leaving Google. Li has, of course, been a professor of computer science at Stanford University for nearly two decades after stints at Princeton and the University of Illinois.
Justin Johnson got his bachelors in computer science at Caltech and his PhD at Stanford, where he met Li. Johnson was then an AI research scientist at Meta Platforms from 2016 through December 2023 and is currently an adjunct professor of computer science and engineering at the University of Michigan. Ben Mildenhall got his bachelors of computer science at Stanford, where he met Li, and then followed up with a PhD in computer science from the University of California at Berkeley. Mildenhall spent several years as a research scientist at Google, focusing on computer vision and generative 3D models. Christoph Lassner got his informatik degrees at the University of Tübingen in Germany and then became an applied scientist at Amazon in September 2017. He moved to Meta Platforms as a research scientist and manager after that and did a year at Epic Games building 3D worlds for video games.
From the announcement put out by AMD, it looks like Johnson and Mildenhall will continue running World Labs inside of AMD, while Li will report directly to AMD chief executive officer Lisa Su and will presumably have a much broader role than helping AMD create world models for robots. At the very least, AMD now has aspirations to move its compute engines into robots, as rival Nvidia is already doing. And by the way, Nvidia co-founder and chief executive officer Jensen Huang has put his son, Spencer, in charge of robotics at Big Green, and his daughter Madison is in charge of product marketing and operations for 3D simulation and Omniverse software.
No one is talking about the headcount at World Labs or the revenue stream and customer count for the acquired company, but presumably the whole team is moving over and is considerably richer today than they were as last week ended. (Interestingly, co-founder Lassner is not mentioned in the AMD announcement, so maybe they are minus one with the World Labs staff.)
Here is the thing. If Nvidia can compete with its customers when it comes to model building, then AMD can, too. And probably has to so it can ensure there are open source frontier models to show AMD is more open than Nvidia with its open weight Nemotron GenAI models and with its free and open source Cosmos world model. (The 20 million hours of video used to train Cosmos is not open, and very likely because it was scraped from the Internet and Nvidia does not own it.)
The question I have is will Li and Su decide to create an open source GenAI model, and open it all the way up – source code, weights, and training data – to prove it is much more open than Nvidia. That’s what I would do if I was Su, and the World Labs team can do this as well as maintain and expand its world models.