World Labs, the start-up founded by Fei-Fei Li, the Stanford professor often called the “godmother of AI”, is being bought by AMD. The chipmaker announced on Monday 28 September 2026 that it had agreed to acquire the San Francisco lab in an all-stock deal valued at about $8.2 billion. Li will join AMD as executive vice president and chief scientist, reporting directly to chief executive Lisa Su.
It is AMD’s second-largest acquisition on record, behind only the roughly $50 billion it paid for Xilinx in 2022, according to CNBC. It is also an unusual one. World Labs does not make chips. It builds world models, AI systems that generate, reconstruct and simulate 3D environments, and AMD is buying it to understand what the chips of the future will need to run.
This article sets out the deal terms, who Li is and what the lab builds, how its value climbed from $1 billion to $8.2 billion in two years, why a chipmaker wants a model lab, where Nvidia fits in, and what the deal means for businesses watching physical AI.
Table of contents
- What AMD Is Paying for World Labs
- Who Is Fei-Fei Li, the ‘Godmother of AI’?
- What World Labs Builds
- What Is a World Model?
- How World Labs’ Value Climbed
- Why a Chipmaker Wants World Labs
- The Nvidia Question
- The Risks for AMD
- How the World Labs Deal Compares
- What Happens Next
- What the World Labs Deal Means for Businesses
- World Labs FAQs
- References
What AMD Is Paying for World Labs
AMD’s announcement describes a “definitive agreement to acquire World Labs, an AI model and research lab led by AI pioneer Dr. Fei-Fei Li.” The table sets out the terms as the two companies described them.
| Item | Detail | Source |
|---|---|---|
| Announced | 28 September 2026, after the US market close | AMD |
| Value | About $8.2 billion | AMD |
| Payment | All AMD stock | AMD |
| Expected close | By the end of 2026, subject to regulatory approvals | AMD; company blog |
| Fei-Fei Li’s role | Executive vice president and chief scientist, reporting to Lisa Su | AMD |
| Team leadership | Co-founders Justin Johnson and Ben Mildenhall continue leading the team with Li | Company blog |
| Before close | The lab kept separate from AMD’s chip business | CNBC |
An all-stock deal
Paying in shares rather than cash lets AMD keep its cash for manufacturing and data centre commitments, and it gives the lab’s researchers a stake in AMD’s future. Fortune reported that AMD shares were roughly flat in after-hours trading on Monday.
What stays the same
World Labs says its team “will continue to focus on advancing AI model research” after the deal closes. In its own announcement, it said it would form “a world leading frontier research organization” inside AMD, committed to “an end-to-end open AI ecosystem spanning hardware, software, platforms, and widely accessible open models.”
Who Is Fei-Fei Li, the 'Godmother of AI'?
Li is one of the best-known researchers in AI. Fortune notes that she is sometimes called the “godmother of AI” for creating ImageNet two decades ago, a large visual database that helped make deep learning possible. She is a Stanford computer science professor and previously worked at Google, where she led AI research, CNBC reports.
Two leading women in AI, on one team
Fortune points out that the deal brings together two of the most prominent women in the industry. Su welcomed World Labs and Li “to the AMD family” in a post on X, adding: “I’ve always been a huge fan of Fei-Fei and her pioneering research in AI.” Li, for her part, called Su a “great friend” and, according to Fortune, an early investor in the company.
Li’s case for joining a chipmaker
Li explained the move in a long post on X. “Without having a focused hardware effort, AI is hobbled in efficiency. And scale. And for our purposes, remains trapped in the digital world,” she wrote, as quoted by Fortune. In AMD’s release she said that “joining AMD will give our team the resources and engineering depth to accelerate our research and help define the infrastructure needed for the next era of AI.”
What World Labs Builds
AMD describes World Labs as developing “spatial-intelligence models that generate, reconstruct and simulate interactive 3D environments from text, image and video inputs, as well as technology for robotic learning and simulation.” Li’s founding argument is that language models are not enough. “The universe isn’t made up of words, it’s made of real things,” she wrote on Monday.
| Date | Milestone |
|---|---|
| 2024 | Emerges from stealth with $230 million at a $1 billion valuation |
| November 2025 | Releases Marble, its first world model product, for editable, downloadable 3D environments |
| 18 February 2026 | Raises $1 billion, including $200 million from Autodesk; AMD and Nvidia among backers |
| 1 September 2026 | Introduces Atlas, an “omni world model” for text, images, video and 3D |
| Mid-September 2026 | Atlas beta access rolls out to users |
| 28 September 2026 | Agrees to join AMD |
Marble and Atlas
Marble, released in November 2025, lets users create 3D worlds they can edit and download. Atlas, introduced on 1 September 2026, is the model the company says will power future versions of Marble. It was pretrained from scratch to work natively across text, images, video and 3D, and the lab says it can generate up to a minute of 1440p video with precise camera control, and reconstruct real scenes from as few as two or three photographs.
Why world models matter for robots
At an earlier demonstration with Su, Li said: “Intelligent agents, whether it’s robots or vehicles or even tools, can learn inside very rich physics-aware digital worlds before they even need to be deployed into the real one, making them much safer.” That is the commercial bet behind World Labs: training robots and autonomous machines in simulation is cheaper and safer than training them in the world. Our look at how video game data is feeding world models covers the wider race.
What Is a World Model?
Most of the AI in the news is built on language models, which learn from text and predict the next word. A world model learns from images, video and 3D data instead, and predicts what a scene looks like from a new angle, or what happens next in it. Fortune puts the goal simply: AI that understands how objects move in the physical world, “for instance, the way that a ball bounces.”
Worlds, not just words
Li has argued for years that this is the missing half of AI. “If AI is to be truly useful, it must understand worlds, not just words,” she said in February. “Worlds are governed by geometry, physics, and dynamics, and reconciling the semantic, spatial, and physical is the next great frontier of AI.” Her company’s own definition is that world models “generate, reconstruct, and simulate any possible world.”
How Atlas builds a scene
Atlas works a little like a language model, the company explains: it encodes its inputs into a context, then generates outputs from it. The difference is that each input image is pinned to a position in 3D space, forming what it calls a spatial context. From a handful of photographs it can produce new camera views, depth maps, point clouds and 3D Gaussian splats, a format that renders on an ordinary device.
A decades-old problem
Rebuilding a real place from a few photographs, known as novel view synthesis from sparse images, is described in the Atlas announcement as “a decades-old fundamental problem in 3D computer vision.” The company says Atlas usually gives faithful reconstructions from two or three images, and can take more than a hundred. “The more it sees, the less it imagines,” it says, which is the property a robot or a digital twin needs.
How World Labs' Value Climbed
The deal price marks a steep rise. The chart uses the valuation TechCrunch reported for the 2024 launch, the $5 billion figure reports said World Labs was targeting for its February round, and AMD’s deal value.
On those figures, the price is 8.2 times World Labs’ 2024 valuation and about 1.6 times the reported February target. The company declined to confirm the February valuation, so the middle bar is a reported number, not an official one.
Roughly $1.2 billion raised
TechCrunch’s figures add up to about $1.23 billion: $230 million at launch and $1 billion in February. Fortune describes the total as roughly $1 billion. If the February round was priced near the reported $5 billion, backers from that round, including AMD itself, Nvidia and Autodesk, stand to see a paper gain, paid in AMD stock.
Who backed the February round
The $1 billion round in February drew a mix of chip companies, investors and one strategic software partner. TechCrunch reported that Autodesk put in $200 million, alongside AMD, Emerson Collective, Fidelity and Nvidia. Autodesk became an adviser, and the two companies agreed to collaborate “at the research and model level”, starting with media and entertainment. Autodesk’s chief scientist, Daron Green, told TechCrunch that data sharing was not part of the agreement, and that customers might one day start with a world-model sketch of an office and then design the desk in Autodesk’s tools.
Why a Chipmaker Wants World Labs
AMD’s stated reason is foresight. “Building the compute platforms for the next generation of AI requires a deep understanding of how models are evolving,” Su said. The release says World Labs’ expertise “will give AMD deeper insight into how workloads are evolving and help shape its future technology roadmaps.”
Designing chips for workloads that do not exist yet
A chip takes years to go from design to data centre. CNBC notes that the lab’s research on next-generation models will let AMD plan what its AI chips need to handle years in advance. World models are a good test case: they combine video generation, 3D geometry and physics, a very different load from the text models most AI chips are tuned for today.
A partnership that became a purchase
The deal did not start cold. World Labs says the two companies began “a deep technical partnership” last year, “starting with model training and inference optimization on AMD GPUs.” As the teams worked together, it says, “we realized it would be a natural fit.” AMD had also invested in the February round.
Part of a broader AMD pattern
World Labs is AMD’s second AI acquisition in two months. On 6 August 2026 it agreed to buy Taalas, a Toronto company building specialised AI inference silicon. Before that came ZT Systems, a rack-scale systems builder, and Silo AI, a Finnish model lab. The pattern is a company assembling every layer of an AI stack, from silicon to systems to models.
The Nvidia Question
AMD is chasing Nvidia in AI processors, and World Labs sits awkwardly between them. Nvidia invested in the same February round. And on 17 September, eleven days before the deal, World Labs posted on X that its new Atlas model had been “trained on NVIDIA Blackwell GPUs.”
What that tells us
It shows how much of today’s frontier AI training still runs on Nvidia hardware, even at a lab already partnering with AMD. It also shows the prize for AMD: after the deal, the lab’s future models are likely to be built and tuned on AMD’s own chips and software, which gives AMD a showcase for physical AI workloads.
Nvidia is in the same race
Nvidia is also expanding into physical AI. Fortune notes that it has built its own world models and partnered with start-ups such as Wayve on models for self-driving cars. Other world-model efforts include Google DeepMind, Runway, and AMI Labs, launched in March by Meta’s former chief AI scientist Yann LeCun. Our report on the Isaac ROS 5 robotics release shows how Nvidia is pairing models with its robotics software.
The Risks for AMD
A deal this size carries real execution risk, and neither company addressed it in detail. Three questions stand out for anyone judging whether AMD has paid a sensible price.
Keeping the researchers
AMD is paying mainly for people. CNBC notes that US tech giants are “opening their wallets for top engineering and research talent”, which means the same researchers will keep receiving offers. Paying in stock ties part of their reward to AMD’s share price, and keeping Justin Johnson and Ben Mildenhall in charge of the team is clearly meant to preserve the lab’s culture. Whether that holds inside a large chip company is the first test.
Moving models onto AMD hardware
Atlas was trained on Nvidia hardware, according to the company’s own post. The partnership with AMD began with training and inference work on AMD GPUs, so some of that porting has started, but training frontier models at scale on AMD’s chips and software will be a larger job. If it goes well, AMD gains a showcase. If it goes slowly, the lab’s research pace could suffer during the switch.
A research lab inside a product company
The lab sells a product, Marble, and runs an API. AMD’s stated reason for the deal is research insight for its chip roadmap. Balancing customer-facing products against roadmap research, and honouring the promise of “widely accessible open models”, will show whether AMD bought a lab, a product line or both.
How the World Labs Deal Compares
For AMD, $8.2 billion is large. The chart compares it with AMD’s other major acquisitions, using the value each deal was announced or reported at.
The ZT Systems figure includes a contingent payment of up to $400 million, and Taalas is left out because AMD did not disclose its price.
Big Tech is paying for AI talent
CNBC places the World Labs deal in a run of large AI purchases. The table sets out the ones it cites.
| Buyer | Target | Value | What it brought |
|---|---|---|---|
| Meta | Scale AI (minority stake), 2025 | $14bn | Founder Alexandr Wang to lead a new AI division |
| Nvidia | Groq assets and Hugging Face | $33bn combined | Inference chips and the largest open model hub |
| AMD | World Labs, 2026 | $8.2bn | World models and Fei-Fei Li as chief scientist |
| OpenAI | io, 2025 | About $6.4bn | Jony Ive and an AI devices team |
Our report on Nvidia’s Hugging Face acquisition covers the open-source half of Nvidia’s spending. AMD’s move is smaller, but it is aimed at a category, physical AI, that neither of those deals targets directly.
What Happens Next
Until the deal closes, CNBC reports, AMD will keep World Labs separate from its chip business. After it closes, Li moves into AMD’s leadership and Johnson and Mildenhall continue running the research team.
Regulatory review
The deal needs regulatory approval. An all-stock purchase of a research lab with no chip products may raise fewer competition questions than a hardware merger, but AMD has only committed to closing by the end of 2026.
Open models or not
Both companies stress an “open AI ecosystem” and “widely accessible open models”. Whether World Labs releases Atlas or later models with open weights once inside AMD is the clearest test of that promise, and the one developers will watch.
What the World Labs Deal Means for Businesses
Few businesses will buy directly from World Labs. The deal still signals where AI spending is heading, and it affects anyone planning to use AI in the physical world.
Physical AI is moving into the mainstream
When Nvidia’s closest GPU rival pays $8.2 billion for a world-model lab, simulation, digital twins and robot training are no longer niche research. Manufacturers, logistics firms and property businesses exploring digital twins or robotics can expect better tools and more hardware choice over the next two years.
More competition in AI hardware
A stronger AMD model team should mean better AMD software support for demanding AI workloads, and more real alternatives to Nvidia for companies buying or renting GPUs. Our guide to local LLM hardware explains how to weigh those choices today.
Start with the problem, not the model
For most firms the practical step is modest: identify one process, such as a warehouse layout, a site survey or an inspection route, where a 3D model of the real space would save time. Our intelligent automation team can help scope where that kind of AI fits and where it does not.
World Labs FAQs
Who is buying World Labs?
AMD, the US chipmaker, announced on 28 September 2026 that it had agreed to acquire World Labs.
How much is AMD paying?
About $8.2 billion, paid entirely in AMD stock.
What will Fei-Fei Li do at AMD?
She will become executive vice president and chief scientist, reporting to chief executive Lisa Su.
What does World Labs make?
World models that generate, reconstruct and simulate 3D environments from text, images and video, including the Marble product and the Atlas model, plus technology for robot learning and simulation.
When will the deal close?
By the end of 2026, subject to regulatory approvals and other customary conditions.
References
AMD to Acquire World Labs to Advance the Future of AI Compute (AMD)
World Labs is Joining AMD (company blog)
AMD acquiring Fei-Fei Li’s World Labs AI firm in deal worth $8.2B (CNBC)
AMD acquires Fei-Fei Li’s physical AI startup World Labs for $8.2 billion (Fortune)
AMD buys firm founded by AI ‘godmother’ Fei-Fei Li (AFP via France 24)
World Labs lands $1B, with $200M from Autodesk (TechCrunch)
Atlas: A World Model for Spatial Intelligence (company blog)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
