AMD has agreed to acquire spatial-intelligence startup World Labs in an all-stock deal worth roughly $8.2 billion, its second-largest acquisition after the $50 billion Xilinx purchase that closed in 2022. World Labs co-founder Fei-Fei Li — the Stanford researcher who led the ImageNet project and is widely called the "Godmother of AI" — will join AMD as executive vice president and chief scientist, reporting directly to CEO Lisa Su. The companies expect the transaction to close by the end of the year, pending regulatory approval.

The deal gives AMD something it has lacked in its fight against Nvidia: deep, in-house expertise in frontier AI models. Su framed the acquisition in exactly those terms, saying that building compute platforms for the next generation of AI requires "a deep understanding of how models are evolving." Li, in turn, has argued that advancing AI "requires close collaboration across model research, systems and compute," and wrote in a separate post that World Labs needs to get "closer to the hardware," without which AI is "hobbled in efficiency."

What World Labs actually builds

World Labs develops what it calls "spatial-intelligence models that generate, reconstruct and simulate interactive 3D environments." Unlike large language models, which are text-centric, its world models construct and simulate 3D scenes — a capability aimed at robotics and simulation workloads rather than chatbot-style applications.

The company shipped its first commercial product, Marble, late last year. Marble creates persistent, downloadable 3D worlds from source material. World Labs says Atlas, its latest world model, will power future versions of Marble.

An existing customer relationship

AMD is not buying a stranger. The two companies had already been collaborating on optimizing training and inference for World Labs' models, and AMD participated in World Labs' $1 billion funding round in February. During Su's CES keynote in January, Li confirmed that part of World Labs' model ran on AMD Instinct accelerators, and that her team used both Instinct hardware and the ROCm software stack to squeeze out performance improvements for their applications.

That tuning work positioned World Labs as one of AMD's showcase customers for its AI accelerator stack — a stack that still trails Nvidia's CUDA ecosystem in developer adoption. AMD says the deal will "help shape its future technology roadmaps," effectively making model researchers a standing input into silicon and systems design.

Deal context

Beyond Xilinx (~$50 billion, closed 2022), AMD's largest acquisitions have been considerably smaller: ATI for $5.4 billion in 2006, ZT Systems for $4.9 billion in 2025, and Pensando for $1.9 billion in 2022. The World Labs purchase fits AMD's pattern of using acquisitions to fill gaps rather than consolidate markets — graphics with ATI, FPGAs and adaptive compute with Xilinx, AI server systems with ZT Systems, and now model research.

Notably, Nvidia was also an investor in World Labs during that February round. Nvidia, however, has been developing its own family of open-source world models, Cosmos, and earlier this month agreed to buy Hugging Face, the open-model hosting and distribution platform, for $12.93 billion. Both chipmakers are converging on the same thesis: model-layer assets increasingly determine whose hardware gets designed in.

Open models stay open

World Labs' team will continue model research after the acquisition. The companies say the deal strengthens an "open AI ecosystem," and World Labs stated in its own post that it remains committed to "widely accessible open models." Li's co-founders, Justin Johnson and Ben Mildenhall, will continue leading World Labs alongside her — including after she takes on the AMD chief scientist role at the expected close.

The open framing matters commercially. If AMD can position Instinct plus ROCm as the natural home for openly licensed world models while Li's team shapes future accelerator roadmaps, the $8.2 billion outlay becomes a bet that the next AI workload category — spatial intelligence — will be trained and inferred on AMD silicon from the start.