Exportable 3D Worlds for Production
World Labs
This is a bet that the winning AI world tool will be the one that hands customers a reusable scene file, not just a convincing camera feed. In production and simulation, teams need something they can reopen, edit, export into Unreal or Blender, attach collision meshes to, and reuse inside Isaac Sim or MuJoCo. Marble is built around that workflow, while video native rivals mainly generate a view of a world rather than the world itself.
-
Marble outputs concrete artifacts, including .PLY splats, .GLB collider meshes, textured meshes, panoramas, and rendered video. That makes it useful for game level prototyping, VFX previs, architecture walkthroughs, and robotics simulation, where the downstream toolchain expects files and geometry, not just pixels on screen.
-
The contrast with interactive video models is practical. Google DeepMind positions Genie 3 as a real time navigable world at 720p and 24 fps, but its product surface is still centered on generated interaction. World Labs is optimizing for deterministic camera placement, editing, and export into existing pipelines, which matters more once a team is trying to ship something.
-
This also explains why Autodesk and NVIDIA matter here. Autodesk put $200M into World Labs and took an advisory role, while NVIDIA is building Cosmos and tying world models into Omniverse and Isaac Sim. That means the commercial prize is becoming infrastructure inside design and robotics stacks, not just making impressive demos.
The next step is a split market. Video first systems will keep winning on immediacy and spectacle, while exportable 3D world systems will move deeper into professional pipelines and physical AI. If World Labs keeps turning generation into editable assets that plug into incumbent software, it can become part of the production stack instead of a standalone novelty.