AI Policy Wiki
Dashboard

World Labs

medium confidence · updated 2026-06-06

Frontier AI startup founded by Fei-Fei Li to build large-scale World Models — generative-pixel-space simulators of 3D environments. Launched Marble (Q4 2025), a system that generates explorable 3D worlds from text, images, video, or 3D layouts.

World Labs is a US AI startup founded in 2024 by Fei-Fei Li to build large-scale World Models — generative simulators of 3D environments. Its first product, Marble, launched in late 2025 and generates explorable 3D worlds from text, images, video, or 3D layouts. The company is associated with the generative-pixel-space approach to world modeling and has raised more than $1 billion in capital.

FieldValue
TypeWorld-Model AI startup (generative-pixel-space)
HQSan Francisco / Palo Alto
Founders[[entities/fei-fei-liFei-Fei Li]], Justin Johnson, Christoph Lassner, Ben Mildenhall
Capital raised$1B+ (per World Models coverage)
Flagship productMarble — 3D-world generator from text / images / video / 3D layouts (launched late 2025)

Snapshot

Funding

DateCapital raisedSource
2026 (reported)$1B+per World Models coverage; [[sources/world-models-mccormick-dewitteMcCormick & De Witte 2026]]

Positioning

World Labs is one of three competing schools in the 2026 World Models field, alongside AMI Labs (latent / JEPA) and General Intuition (action-conditioned). World Labs sits in the generative-pixel-space camp, alongside Google DeepMind's Genie line, NVIDIA Cosmos, and Sora-derived OpenAI work. According to McCormick & De Witte 2026, the approach bets that observable, watchable 3D worlds will find immediate market in AR/VR, entertainment, and digital twins, while incrementally compounding toward more general world-modeling capability.

The company's pitch to investors draws on Fei-Fei Li's framing that "intelligence is the AI of the future, but the world is what intelligence acts in," casting embodied AI and robotics as the successor to the LLM era.

Marble

Marble is World Labs' first product release, launched in late 2025. It generates fully explorable 3D worlds from any of four input modalities: natural-language descriptions, single or multiple still images, short videos, or coarse 3D scene layouts. The output is a Gaussian-splat scene a user can navigate.

Marble's reception marked "World Models" as a fundable investment category in late 2025. By early 2026 the term was sufficiently marketed that AMI Labs CEO Alexandre LeBrun joked, "in six months, every company will call itself a World Model to raise funding" (McCormick & De Witte 2026).

Relationships