AMD Buys World Labs for $8.2 Billion as Fei Fei Li Becomes Chief Scientist
AMD is making one of its largest AI acquisitions yet, agreeing to buy Fei Fei Li’s World Labs for approximately $8.2 billion in stock as it pushes beyond accelerators and deeper into the models that will shape future AI hardware. The deal brings World Labs’ spatial intelligence research, models, and engineering team directly into AMD, while Li will become Executive Vice President and Chief Scientist reporting to CEO Lisa Su. AMD expects the transaction to close before the end of 2026, subject to regulatory approval and other customary conditions.
The acquisition is not simply about adding another AI product to AMD’s portfolio. In its official announcement, AMD says World Labs will give its engineers deeper insight into how emerging AI workloads are evolving, allowing future GPUs, CPUs, software, and complete systems to be designed around the needs of new model architectures. World Labs will continue focusing on AI model research after the transaction closes, which makes the deal as much about bringing model development closer to AMD’s hardware roadmap as acquiring an existing business.
World Labs specializes in spatial intelligence, building AI systems that can understand, generate, reconstruct, and simulate 3D environments rather than working primarily with text or flat images. Its first commercial product, Marble, can create persistent 3D worlds from text, images, video, or existing 3D layouts. The company recently introduced Atlas, a multimodal world model trained to operate across text, images, video, camera positions, and 3D depth information. Atlas can generate controlled camera sequences, reconstruct physical environments, produce point clouds and Gaussian splats, and support simulation workflows for robotics.
$AMD is buying @theworldlabs for ~$8.2B in stock, and Fei-Fei Li (@drfeifei) will join as EVP and chief scientist reporting to @LisaSu.
— Patrick Moorhead (@PatrickMoorhead) September 28, 2026
I called Fei-Fei a great hire for Google Cloud back in 2017. Same here. This wasn't out of the blue, either. Lisa had her on stage at CES in… pic.twitter.com/ZXNpGiCiXz
That direction puts AMD into a part of the AI market where NVIDIA has already invested heavily through Cosmos. NVIDIA’s current Cosmos 3 platform combines reasoning, world generation, and action prediction for robotics, autonomous vehicles, and other physical AI applications. World Labs gives AMD its own major research organization working on similar spatial and physical intelligence problems, but AMD has not announced that Atlas, Marble, or another future model will become a direct Cosmos replacement. Describing the acquisition as AMD building a Cosmos competitor is therefore an industry interpretation rather than an official product roadmap.
The relationship between the 2 companies was already substantial before the acquisition. World Labs and AMD began working together on model training and inference optimization using AMD GPUs last year, while AMD participated in World Labs’ $1 billion funding round in February alongside NVIDIA, Autodesk, Fidelity, Sea, and other investors. World Labs did not disclose a valuation for that financing. Bloomberg had previously reported that the company was discussing funding at around a $5 billion valuation, so comparisons suggesting AMD is paying a precise 62% premium should be treated carefully because the $5 billion figure was never confirmed as the completed round valuation.
Industry analyst Patrick Moorhead described the acquisition as part of AMD moving into the model layer of AI, noting that Li and her team bring expertise AMD did not previously possess internally at this scale. That characterization fits AMD’s own explanation that the acquisition brings model researchers directly into the company, but calling the entire $8.2 billion transaction a talent grab understates the assets involved. AMD is also gaining World Labs’ models, research, products, software, intellectual property, and ongoing work in robotics and simulation.
For AMD, the strategic value could extend directly into future Instinct hardware. World Labs has repeatedly emphasized that world models are extremely compute intensive. Its RTFM research noted that interactive 4K world generation at 60 FPS could require processing more than 100,000 tokens per second, while persistent environments can create extremely large context requirements. Having the researchers building those workloads inside AMD could give the chip designer earlier visibility into memory capacity, bandwidth, networking, inference, and software requirements for future physical AI systems.
That fits AMD’s broader effort to challenge NVIDIA with more than individual accelerators. We have already seen AMD expand open AI infrastructure around Instinct MI355X, while ROCm 10 is targeting some of the software barriers that have historically made NVIDIA’s ecosystem difficult to challenge. World Labs adds another layer by bringing frontier model research directly inside AMD instead of relying only on external developers to define future workloads.
The $8.2 billion price makes more sense when the acquisition is viewed as an attempt to influence AMD’s entire future AI stack rather than simply buying a startup with a 3D generation product. Fei Fei Li and the World Labs team can tell AMD what the next generation of spatial and physical AI actually needs before those workloads become mainstream, potentially feeding that knowledge directly into future GPU architecture, memory systems, software, and rack scale infrastructure.
The NVIDIA Cosmos comparison is useful, but it should not become the story by itself. World Labs already has its own research direction through Marble, RTFM, and Atlas, while AMD says the team will continue advancing its models. The bigger shift is that AMD is no longer positioning itself only as the company providing compute underneath other AI labs. It now wants part of the frontier model research happening inside the company that designs the hardware.
Does AMD need its own frontier world model research team to compete more effectively with NVIDIA in physical AI, or should it remain focused primarily on open hardware and software infrastructure?
