Summary: On February 18, 2026, World Labs — a spatial intelligence startup founded by Stanford professor Fei-Fei Li, the computer scientist widely known as the “godmother of AI” — announced that it had raised $1 billion. The number, though, matters less than the shareholder register. Reuters put the round’s valuation at roughly $5 billion, and the company’s own blog post names its backers: AMD, Autodesk, Emerson Collective, Fidelity Management & Research Company, NVIDIA, and Sea.
On February 18, 2026, World Labs — a spatial intelligence startup founded by Stanford professor Fei-Fei Li, the computer scientist widely known as the “godmother of AI” — announced that it had raised $1 billion. Reuters put the round’s valuation at roughly $5 billion, and the company’s own blog post names its backers: AMD, Autodesk, Emerson Collective, Fidelity Management & Research Company, NVIDIA, and Sea. The $230 million round that got the company started in September 2024 was less than eighteen months behind it.
What the money is chasing is no secret: spatial intelligence. Frontier labs are pouring gigawatt-scale compute into language and image models; World Labs is building world models instead — systems designed to understand and generate three-dimensional space. Set against the 2026 model race, a billion dollars just bought a ticket on a different route.
The Round, and Who Is In It
Crunchbase’s weekly funding recap for February 20 ranked World Labs as the week’s largest deal by a wide margin, and press accounts of the round sized Autodesk’s check at roughly $200 million. The number, though, matters less than the shareholder register. NVIDIA and AMD — the two chipmakers collecting the AI data center buildout’s revenue — are at the same time buying equity in an application-layer model company. Autodesk adds a design-tools distribution channel; Fidelity adds institutional capital. Read this as a supply chain marking out positions, not as a standard venture syndicate.
Li’s biography supplies much of the narrative on its own. ImageNet — the dataset that taught machines to see, and arguably the spark that ignited the deep learning era — was built under her leadership. World Labs is her claim on the decade ahead: machines have learned to recognize images; the next lesson is modeling space.
World Models, Explained
A world model and an LLM part ways on a single question: what is being learned. An LLM absorbs the statistical structure of language; a world model absorbs the structure of physical space — geometry, persistence, which object sits where, and what changes as you move through it. World Labs positions the technology for storytelling, creativity, robotics, and scientific discovery.
Robotics needs spatial grounding, and that is exactly the capability text models cannot supply. Games and simulation need a persistent world that survives being revisited from any viewpoint, not a run of disconnected frames. The distance between “generating an image of a room” and “generating a room” is the whole product thesis.
Marble: A Product That Already Ships
Marble, the company’s first product, produces 3D worlds from images, video, or text, and it describes those worlds as “spatially coherent, high-fidelity, and persistent”. The surrounding stack includes a developer API, the Spark JavaScript library, and the experimental Marble Labs. This is the detail that should shape how you read the funding coverage: the billion is going into a product line that already ships — with an API and a web-native integration path — rather than into a slide deck. And treating the JS library as a first-class release tells you who they expect to build with this: web developers.
The Investors’ Calculations
NVIDIA cuts the most instructive figure here. It moves GPUs by the tens of thousands into buildouts like the Meta–NVIDIA Blackwell-scale expansion, and it also buys equity on the demand side of the same market. The reasoning is not complicated: world models are compute-hungry in both training and inference, and 3D generation widens what AI compute is for — past chat and search, into simulation and content production. New workload classes open new demand curves.
For AMD, an equity stake in a model company is one route to getting its accelerators threaded into this ecosystem early. For Autodesk, the math runs nearer-term: design professionals are the obvious first market for generated 3D, and a stake costs less than building the capability in-house.
Reading It as a Builder
First, treat 3D generation APIs as having reached the seriously usable stage: game prototyping, architectural visualization, and synthetic training environments for robotics are all fair tests for Marble now. Second, watch the cap table as a signal: when NVIDIA, AMD, and Autodesk converge on the same thesis, the industry is beginning to price 3D understanding as the next axis of differentiation. Third, world models complement LLMs rather than replacing them — one carries language and reasoning, the other carries space and simulation. The products worth watching in 2026 wire the two together.