The company
World Labs came out of stealth in 2024 with Fei-Fei Li, the Stanford computer scientist best known for ImageNet, as co-founder. Its framing is spatial intelligence: models that understand the three-dimensional structure of a scene rather than a sequence of flat frames. That is a deliberate contrast with the video-first world models from Google DeepMind and Runway, and it shows up in every product decision.
Marble: generate once, then own the result
Marble launched on 12 November 2025. You give it a text prompt, one or more images, a video or a coarse 3D layout, and it builds a world you can walk through, edit, expand and stitch to other worlds. The output is the point: you can export Gaussian splats, meshes (a collider mesh for basic physics and a higher-quality visual mesh) or video, and those files load into Unity, Unreal or a web viewer like any other asset.
Generations are priced in credits, with lighter draft models for exploration and heavier ones for final output. That pricing model tells you who it is for: people producing scenes, not people playing in them.
The trade against Genie is the clearest way to understand it. Genie regenerates the world every frame, so you can act in it and it reacts, but detail drifts and nothing persists after the session. Marble does not drift because nothing is regenerated, and for the same reason it does not simulate consequences. If you need a set, Marble is the better tool. If you need a sandbox, it is the wrong one.
Atlas: the model behind the next Marble
On 1 September 2026 World Labs announced Atlas, an "omni" world model pretrained from scratch and meant to power future versions of Marble. In early access it does camera-controlled image and video generation, reconstruction from a handful of photos, depth maps, point clouds, 3D Gaussian splats and multi-camera reframing, with showcase clips up to a minute long at 1440p. World Labs also points it at a robotics real-to-sim workflow: photograph a space, reconstruct it, train in the reconstruction.
Treat the benchmark claims with the usual care. World Labs reports that Atlas leads on its own camera-following preference test and on a reconstruction average across seven datasets; neither had been independently reproduced when this page was written, and pricing, parameter count and a public API were not disclosed.
The AMD deal
In late September 2026 AMD agreed to acquire World Labs for $8.2 billion in stock, its second announced acquisition of the year, with closing expected before the end of 2026 subject to regulatory approval. Reports say Fei-Fei Li will join AMD as chief scientist.
The logic is the same one running through the whole world-model market. These models are among the most compute-hungry workloads in AI, so whoever controls a leading one controls a large block of future GPU demand. Nvidia built Cosmos itself; Amazon has backed Odyssey and is reported to be negotiating for Decart; AMD has now bought rather than built. For buyers, the practical consequence is that World Labs’ roadmap will increasingly be optimised for AMD hardware.
When I would use it
- Game, film and XR pre-production: generating environments you then light, dress and ship is where Marble is strongest today.
- Digital twins and robotics sim: worth testing Atlas’s real-to-sim path, but measure how well policies trained in a reconstruction transfer back to the real site before relying on it.
- Interactive experiences: use an interactive model such as Genie or Runway instead, or export from Marble into an engine that supplies the interactivity.
Related
- World models: the explainer and the full list of world models
- Google Genie vs other world models: the interactive alternative, compared
- World model startups: Decart, Odyssey, General Intuition, Runway and the acquisition wave
- Nvidia Cosmos: the open model most robotics teams start with
Frequently asked questions
What is World Labs?
World Labs is a San Francisco AI company co-founded in 2024 by Stanford professor Fei-Fei Li to build what it calls spatial intelligence: models that understand and generate three-dimensional worlds. Its first product, Marble, launched in November 2025. In September 2026 AMD agreed to acquire the company for $8.2 billion in stock.
What does Marble do?
Marble turns text, images, video or a coarse 3D layout into a navigable 3D world that you can edit, expand and combine with other worlds. You can export the result as Gaussian splats, as meshes (including collider meshes for basic physics) or as rendered video. Generations are priced in credits rather than per seat.
How is Marble different from Google Genie?
Genie generates frames in real time as you move, so the world is interactive but exists only while the model runs and can drift. Marble generates the scene once and gives you a persistent asset that does not drift, but it does not simulate what happens when you act in it. Genie is closer to a game engine replacement; Marble is closer to a 3D asset pipeline.
What is World Labs Atlas?
Atlas is World Labs’ second-generation world model, announced on 1 September 2026 in early access. The company describes it as an omni model trained from scratch that does camera-controlled image and video generation, reconstruction from sparse photos, depth, point clouds and 3D Gaussian splats, with showcase video up to a minute at 1440p. Pricing and a public API had not been announced as of November 2026.
Why is AMD buying World Labs?
AMD has not given a detailed rationale beyond spatial intelligence, but the pattern across the industry is clear: world models consume enormous compute, and chipmakers want the demand on their hardware. Nvidia has Cosmos, Amazon has invested in Odyssey and is reported to be negotiating for Decart, and AMD now has World Labs.
Sources: World Labs (Marble announcement, November 2025), Kingy AI and Spatial Insiders (Atlas, September 2026), ITPro and CRN Asia (AMD acquisition, September 2026), TechCrunch. Capabilities are as published by World Labs; benchmark claims are company-reported.