World Labs

World Labs is an artificial intelligence company working on spatial intelligence — systems that understand and generate three-dimensional space rather than only text or images. Founded in 2024 by Stanford professor Fei-Fei Li, the creator of ImageNet, together with Justin Johnson, Christoph Lassner and Ben Mildenhall, it builds world models that turn a photograph, a sketch or a line of text into a persistent, navigable 3D environment. On 28 September 2026, AMD announced an agreement to acquire the company in an all-stock deal valued at approximately US$8.2 billion, expected to close by the end of that year subject to regulatory approval.^[1]

History

World Labs was founded in early 2024 and emerged from stealth that April with US$230 million in funding at a valuation of around US$1 billion, backed by investors including Andreessen Horowitz, NVIDIA's venture arm and Radical Ventures. Its first product, Marble, was previewed during 2025 and released to the public on 12 November 2025, making it one of the first commercial products in the emerging world model category.^[2]

In February 2026 the company raised US$1 billion in a round that included a US$200 million investment from Autodesk, along with NVIDIA, AMD, Fidelity, Emerson Collective and Sea, with reports placing the valuation near US$5 billion. Autodesk also took an advisory role, signalling interest in applying world models to design and engineering workflows. Seven months later AMD — already an investor in that round — agreed to buy the company outright, with Li joining AMD as executive vice president and chief scientist reporting to chief executive Lisa Su.^[3]

Technology

A world model differs from a video generator in three ways that define the product category:

• Persistence. The environment is generated once and stays consistent as the camera moves through it, rather than re-generating each frame.
• Editability. Scenes can be expanded and modified region by region instead of being re-rolled from scratch.
• Export. Worlds can be downloaded as Gaussian splats or triangle meshes, or rendered to video, so they can be used in game engines, design tools and web viewers.

Marble accepts text prompts, single images, short videos, panoramas and rough layouts as input, and outputs geometry, materials and lighting for the resulting scene. The stated goal is applicability beyond entertainment — robotics training, simulation, and scientific visualisation are all cited as target uses. Fei-Fei Li summarised the thesis behind the company: if AI is to be truly useful, it must understand worlds, not just words.^[4]

Strategic significance

The AMD acquisition illustrates why spatial models matter to hardware makers. Chip vendors compete on accelerators and software ecosystems, and owning a team that builds frontier world models gives a processor company early insight into how 3D and physical AI workloads will stress future hardware, allowing its roadmaps to be shaped around those demands rather than around customer requests that arrive after the fact. It is AMD's largest acquisition since Xilinx in 2022.

>See Also

• World Models
• Physical AI
• Gaussian Splatting
• Neural Radiance Fields
• AMD