Physical AI Was Nvidia's to Lose. AMD Is Making It a Fight With Its $8.2 Billion World Labs Acquisition.

AMD said on Monday it will acquire World Labs in an all-stock deal worth about $8.2 billion. Fei-Fei Li, who co-founded the company in 2024, will join AMD as executive vice president and chief scientist and report to CEO Lisa Su. Co-founders Justin Johnson and Ben Mildenhall will keep running the World Labs team inside AMD. The companies expect the deal to close by the end of the year if regulators approve it.

AMD's announcement talks about roadmaps and workloads. Read more simply, AMD had no real position in physical AI, the market for robots, autonomous vehicles and the simulated environments used to train them. Nvidia has had one for years.

What World Labs makes

World Labs builds world models, which are AI systems that generate and simulate 3D environments from text, images or video. Its first product, Marble, came out in November 2025. You give it a few photos or a prompt and it produces a 3D scene you can move around in.

Game studios and film teams were among the first users. Most of the long-term interest in the category comes from robotics, though. Teaching a robot to handle objects requires a lot of data about how things behave physically, and gathering that data in the real world takes time and money. A realistic simulation can produce it much faster. Li made this argument when she appeared with Su at CES in January, saying robots and vehicles could train in virtual worlds that follow the laws of physics before going out into the real one.

How far ahead Nvidia is

Nvidia started building for this market well before most of its competitors. In addition to GPUs, it sells robotics simulation software and releases its own open-weight world models, called Cosmos. A robotics company can buy most of what it needs from Nvidia alone.

AMD sells competitive chips, but that was about where its offering stopped. TechCrunch noted that the only models AMD had released publicly worked with text and video. It had no equivalent to Cosmos and no research group known for work in this area. Buying World Labs gets it a model team and Li herself, whose ImageNet dataset played a major part in the rise of deep learning.

The price

When World Labs came out of stealth in September 2024, it had raised $230 million at a $1 billion valuation. In February it raised another $1 billion in a round led by Autodesk, which contributed $200 million. AMD, Nvidia, a16z and Fidelity also invested. Reports put that round at roughly $5 billion, although World Labs never confirmed the figure.

AMD is paying around 60% more than that, seven months later. Spread across a staff of about 100, the price works out to more than $80 million per person. Marble costs between $20 and $95 a month, and World Labs hasn't said how much revenue it brings in. AMD is mostly paying for the researchers and what they're working on.

Because the deal is paid in AMD stock, AMD doesn't have to spend much cash on a bet that could take years to pay off. The stock structure has one odd side effect, since Nvidia and Intel are both World Labs investors. If their stakes are converted like everyone else's, both companies will end up owning AMD shares.

How the deal came together

The companies began working together in 2025, when World Labs started optimizing its training and inference on AMD GPUs. AMD Ventures later invested in the company and said it was helping World Labs run more of its compute on AMD Instinct chips. Li and Su then appeared together at CES in January.

Li wrote in a post on Monday that growing World Labs' research required being closer to the hardware. She and her co-founders concluded the lab would get further as part of a chip company than as a customer of one, and AMD paid a significant premium to make that happen.

Other infrastructure companies are doing the same thing

This is the third deal of its kind since August. Stripe bought OpenRouter, a service that routes businesses' AI requests to hundreds of different models, for about $7.5 billion. Nvidia then acquired Hugging Face, where much of the open model community shares its work, for $13 billion. Now AMD is buying World Labs.

All three buyers run infrastructure that AI companies depend on. Each has decided it needs direct ownership of part of the model business instead of only selling to it. Owning a lab means seeing changes in demand sooner, and it can help steer where that demand goes.

That matters a great deal to a chipmaker. Chips take years to design, so AMD's engineers are making choices now about hardware that customers will use around 2029 or 2030. The researchers who best understood future robotics and simulation needs used to be Nvidia customers, and Nvidia tended to hear from them first. From now on AMD's chief scientist will be one of them.

What could go wrong

AMD says the deal will strengthen an open AI ecosystem, and Li's post also talked about working in the open. That positioning makes sense for a company competing with Nvidia's closed, integrated stack. It will be tested after the deal closes. If new World Labs models run much better on AMD hardware than on anything else, AMD benefits and customers have fewer choices.

AMD also has to keep the team. Research groups acquired by big public companies often lose key people within a couple of years as priorities shift. AMD is counting on Li's leadership and access to a lot more compute to prevent that.

Existing customers carry some risk as well. Autodesk put $200 million into World Labs in February to build its models into Autodesk's design tools. That product now belongs to a chipmaker. CNBC reported that AMD will keep World Labs separate from its chip business until the deal closes, and AMD hasn't said what happens after that.

AMD shareholders, meanwhile, are paying for a company with no reported revenue whose value to AMD may not show up until a future generation of chips. Finimize wrote that the strategy is easier to see than the financial return.

What this means for enterprise buyers

Companies looking at world models for robotics, simulation or digital twins should expect the main options to be tied to chipmakers, with Cosmos at Nvidia and World Labs at AMD. It's worth asking vendors how their models perform on different hardware and getting portability commitments into contracts now, while the two are still competing for customers.

Current Marble and World Labs API customers shouldn't see changes before the deal closes. After that, they should keep an eye on pricing and on any new hardware requirements.

Teams negotiating long-term GPU contracts now have a better argument for pressing Nvidia on price, since AMD can offer a physical AI story of its own.

Founders of model companies have more potential buyers than they did a year ago. Chipmakers and infrastructure providers have become some of the most active acquirers in AI, particularly for startups working on world models, video and robotics.

Nvidia spent years assembling hardware, software and models for physical AI while AMD concentrated on chips. The World Labs deal gives AMD its own research lab in the field. Whether that closes the gap will depend on what the lab produces over the next couple of years and how well it runs on AMD hardware.