World Labs

World Labs is a spatial-intelligence company building world models — systems that generate, reconstruct and simulate three-dimensional environments rather than producing text or flat video. It was founded in 2024 by Fei-Fei Li, the Stanford researcher who created ImageNet and is often credited with catalyzing the deep-learning era in computer vision, alongside Justin Johnson, Ben Mildenhall and Christoph Lassner — a founding team drawn from machine learning, generative AI, computer vision and graphics. The company has raised roughly 1.23 billion dollars, including a round in February 2026 in which Autodesk alone contributed 200 million dollars, and its investor list runs to about 20 names including a16z, NVIDIA, AMD, Adobe, Autodesk, Salesforce, Intel, Cisco, Databricks, NEA, Fidelity, Samsung and Temasek.

The company's thesis is that language models learned to reason about the world through text, and that the next capability step requires models that represent space directly. Its first product, Marble, generates spatially consistent, high-fidelity and persistent 3D worlds from multimodal inputs, supports interactive editing and world expansion, and exports to standard formats. World Labs also maintains an open-source renderer, Spark, and a developer platform.

In September 2026 the company introduced Atlas, which it describes as an omni world model for spatial intelligence. Atlas is a single multimodal autoregressive diffusion transformer that handles text, images, video and 3D data natively, grounding its inputs in a shared spatial context rather than treating each modality separately. It generates camera-controlled video running up to one minute at 1440p, reconstructs three-dimensional scenes from images, and generates images from text. World Labs reports 81 percent human preference over Gemini Omni Flash on camera-controlled generation and lower reconstruction error than specialist models across several standard datasets — figures the company measured itself rather than independently verified results.

Atlas is not broadly available. It entered early access with selected partners through a waitlist, with no open weights, no public developer interface for the model and no published pricing. That access model, rather than any judgment about the technology, is why Atlas has no directory tool page of its own; the honest read for a practitioner is that it is a research demonstration with a commercial path, not something to build on this quarter. Marble is the opposite case and does carry its own tool page: it is self-serve, with a free tier and published subscription pricing. The two products also differ in what they grant — the Spark renderer is MIT-licensed, while commercial rights to the worlds Marble generates begin at its Pro plan rather than at the free or entry tier. The customers most visible in its backing — Autodesk in design software, Adobe in creative tooling, Samsung in devices — indicate where the company expects world models to land first.

In September 2026 World Labs signed a definitive agreement to be acquired by AMD for about 8.2 billion dollars, after a year of joint work on training and running its models on AMD graphics processors. Fei-Fei Li is to join AMD as executive vice president and chief scientist, reporting to chief executive Lisa Su, with Justin Johnson and Ben Mildenhall continuing to lead the team inside a new AMD frontier research organization. The stated aim is an open stack spanning chips, software and openly available models. The deal is expected to close by the end of 2026, subject to regulatory approval, and until it does World Labs remains an independent company.