On February 18, 2026, Fei-Fei Li's World Labs closed a $1 billion funding round on a thesis the rest of the AI industry isn't running: that spatial intelligence — teaching machines to understand worlds, not words — is the frontier the language-model race keeps missing. Nvidia and AMD both wrote checks. Autodesk put in $200 million and took a strategic-advisor seat on the roadmap. World Labs has shipped exactly one commercial product, and investors valued it like a frontier lab anyway.
The conversation about AI in the spring of 2026 sounds almost identical to the conversation in 2024, only louder. OpenAI is preparing to ship GPT-5.5 — the model rumored under the codename "Spud" — later this month. Anthropic spent late March managing the fallout from a Fortune scoop revealing the existence of an internal model called Mythos via a data leak. The federal lawsuits, the Pentagon supply-chain designations, the boardroom dramas — all of it orbits the same gravitational center: a handful of frontier labs racing to build progressively larger language models and arguing about who gets to govern them.
Fei-Fei Li is not in that conversation. She's in a different conversation entirely, on a different axis.
World Labs, the company Li co-founded after stepping back from her Stanford-and-Google posts, closed a $1 billion funding round in February. Nvidia and AMD both wrote checks. Autodesk put in $200 million and took a strategic-advisor seat on the company's roadmap. The valuation reporting put World Labs in the same neighborhood as the frontier LLM labs, despite the fact that the company has shipped exactly one commercial product and has not, by any conventional measure, demonstrated revenue at the scale that funding implies.
What it has demonstrated is a thesis. Li has been articulating it publicly since at least her 2024 TED talk, "With Spatial Intelligence, AI Will Understand the Real World," and she developed it further in her 2025 Substack essay "From Words to Worlds." The thesis, paraphrased: language models are extraordinary at compressing and reproducing the patterns of human text, but they have no native model of the physical world. They do not understand geometry, persistence, occlusion, gravity, or the simple fact that an object continues to exist when nothing is looking at it. For AI systems that need to act in physical or simulated environments — robots, autonomous vehicles, game engines, surgical tools, AR devices — the language-model paradigm is not just incomplete. It is the wrong shape.
Li's term for what's missing is spatial intelligence, and World Labs exists to build it.
This is the part where industry-leader profiles usually stop being accurate. So let me be precise about what World Labs has put into the world, not what it has promised.
In October 2025, World Labs released RTFM — a Real-Time Frame Model — as a research preview. RTFM is a single neural network that takes one or more 2D images of a scene as input and generates new 2D views of that scene from arbitrary camera positions, in real time, on a single Nvidia H100 GPU. It does this without building any explicit 3D representation. Each generated frame carries a pose, and the model uses those poses as a kind of spatial memory, which is how it maintains persistence — walk away from a chair, come back, and the chair is still there, in the same place, with the same shadow.
In November 2025, World Labs launched Marble, its first commercial product. Marble takes a text prompt, image, video clip, panorama, or rough 3D layout and produces a persistent, navigable, downloadable 3D environment. Output formats include Gaussian splats, meshes, and video — formats that game studios, VFX houses, and VR developers actually use. Marble worlds are compatible with Apple Vision Pro and Meta Quest 3 out of the box. There is a free tier and a paid tier. Marble also ships with an experimental editor called Chisel that lets a user block out coarse spatial layout — walls, floors, masses — and then describe the visual style separately in text. The structural blocking and the artistic style are decoupled, which Li's team has compared to the relationship between HTML and CSS.
That is the entire shipped product surface as of early April 2026. It is not large. It is also not vaporware.
The strategic case for spatial intelligence is not that LLMs are losing — they are very obviously not — but that the assumption "general intelligence will emerge from scaling language" is a bet, not a law. It is one bet, and it is the bet that the GPT-5.5 launch and the Mythos data leak are both, in different ways, expressions of. Both stories are about who controls the language frontier, how big the next model gets, and which government agency is allowed to use which lab's API.
World Labs is making a different bet. Its claim is that an AI system that can generate, edit, and reason about persistent 3D environments unlocks a category of applications that no amount of LLM scaling will reach: physically-grounded simulation for robotics training, spatial workflows for industrial design, generative environments for entertainment, and eventually, AR systems that can place persistent virtual objects into the real world without the seams falling apart the moment you look away.
Autodesk's $200 million is a tell. Autodesk is not a speculative buyer. It sells CAD and BIM software to people who design buildings, factories, and consumer products, and it is treating Marble as something its customers might integrate into a 3D design pipeline within the planning horizon of a strategic partnership. Nvidia's participation is a tell of a different kind: the company that sells the picks and shovels to every AI lab on Earth is signaling that it expects spatial AI to consume serious compute, separate from and additive to the LLM training market.
For business leaders trying to read 2026 correctly, the practical takeaway is not that LLMs are about to be eclipsed. They are not. GPT-5.5 will ship, it will be capable, and most of the productivity gains your organization captures this year will come from language models, not world models.
The takeaway is that the AI category is bifurcating, and the second branch has now passed the funding threshold where it can no longer be dismissed as a Stanford research curiosity. If your business touches anything physical — manufacturing, logistics, real estate, construction, entertainment, healthcare devices, robotics, AR, automotive — the spatial-intelligence track is the one to watch, and the relevant questions are not "which LLM is biggest" but "who is building persistent, editable, physically-consistent world models, and how do they integrate with the design and simulation tools my industry already uses?"
Li has spent her career arguing that vision came before language in evolution and that intelligence, in any meaningful form, has to be able to see, navigate, and act, not just talk. World Labs is the operational form of that argument. It is the part of the AI industry that is not, this month, fighting a federal lawsuit or leaking model names to Fortune. It is shipping Marble and RTFM, and it now has the balance sheet to keep doing so.
Whether that bet pays off is genuinely unknown. That it has been made, at this scale, by these investors, is not.
Per Li's 2024 TED talk and 2025 Substack essay "From Words to Worlds": models that understand geometry, persistence, occlusion, gravity, and the continuity of objects in space — capabilities language models do not have natively because text doesn't carry that structure. World Labs is building this layer for robots, AR/VR, autonomous vehicles, and design/simulation tools.
One commercial product (Marble — interactive editable 3D world generation) and one research demo (RTFM — real-time generative world engine). Frontier-lab funding has gone to a small commercial footprint and a thesis backed by a long Stanford-and-Google research record, not yet a revenue scale that conventionally justifies the valuation.
If your business touches anything physical — manufacturing, logistics, real estate, construction, entertainment, healthcare devices, robotics, AR, automotive — yes. The integration questions to ask are about persistent editable physics-consistent world models and their compatibility with the design and simulation tools you already use.
Discover more content:
