Start with the bull case, because it is a good one. Fei-Fei Li is one of the few people in AI whose track record actually earns the hype, and she has spent the last two years, and now $1.23bn, on a bet that the next great leap in artificial intelligence will not come from language at all. It will come from space. Not outer space, but the ordinary three-dimensional world your body navigates without thinking: the geometry of a room, the way light falls on a surface, what is behind the sofa. Her company, World Labs, calls this “spatial intelligence”, and if she is right, it reframes what AI is for.
This is the opening profile in our AI Innovation section, and World Labs is the natural place to start, because it is the most credible pure-play in the fast-emerging world-models field.
The founder: the woman who taught machines to see
You cannot understand World Labs without understanding Fei-Fei Li, because the company is essentially her academic life’s work turned commercial. In 2009 she built ImageNet, a database of millions of labelled images, and the annual competition around it. That dataset is, without exaggeration, one of the reasons the modern AI boom happened: it gave the deep-learning revolution of 2012 something to prove itself on, and computer vision has never looked back. She has since run the Stanford AI Lab, served as chief scientist of AI at Google Cloud, and become known, a little grandly but not unfairly, as a “godmother of AI”.
She did not build World Labs alone. Her co-founders are a genuine dream team of 3D and vision researchers: Justin Johnson, Christoph Lassner, and Ben Mildenhall, the last of whom co-invented Neural Radiance Fields (NeRF), the technique that lets you reconstruct a photorealistic 3D scene from a handful of 2D photos. If you wanted to assemble the four people most likely to make spatial AI work, this is close to the list you would draw up.
What they are actually building
World Labs builds what it calls Large World Models, or LWMs, deliberately echoing the LLMs that made OpenAI and Anthropic famous. Where a language model predicts text, a world model builds an internal, navigable representation of a 3D environment: it knows that objects persist when you look away, that surfaces have depth, that you can walk around a thing and see its other side.
The first product, launched to the public, is Marble. Feed it an image, a video clip, or even a line of text, and it generates a spatially coherent, high-fidelity, persistent 3D world you can move through. “Persistent” is the key word. Unlike an AI video generator that hallucinates a fresh, inconsistent frame every moment, Marble’s worlds hold together: the corner of the room stays where it was when you turn back around. For anyone who has watched AI video melt into nonsense after four seconds, that consistency is the genuinely new trick.
Why the smart money is excited
On 18 February 2026, World Labs announced a $1bn funding round, bringing its total to roughly $1.23bn across two rounds since it emerged from stealth in September 2024 at a $1bn valuation. That makes it one of the most heavily capitalised spatial-AI companies in existence before it has a mass-market business.
Look at who wrote the cheques, because the investor list is the tell. Nvidia and AMD, the two companies that sell the chips this all runs on. Fidelity, the institutional money. Sea, the Southeast Asian tech giant. And, most revealingly, Autodesk, the maker of AutoCAD and the software behind much of the world’s architecture, engineering and manufacturing design, which put in $200m and signed on as an adviser. When the company that owns professional 3D design tooling invests at that scale, it is telling you where the first real revenue probably lives: not in sci-fi, but in helping architects, game studios, product designers and roboticists build and test in simulated space.
| World Labs at a glance | Detail |
|---|---|
| Founder / CEO | Fei-Fei Li (with Justin Johnson, Christoph Lassner, Ben Mildenhall) |
| Founded | Emerged from stealth September 2024 |
| Total raised | ~$1.23bn across two rounds |
| Latest round | $1bn, announced 18 Feb 2026 |
| Key backers | Nvidia, AMD, Autodesk ($200m + adviser), Fidelity, Sea, Emerson Collective |
| First product | Marble: persistent 3D worlds from image, video or text |
The friendly sceptic’s corner
Now the cold water, offered in good faith, because we like this company. First, “spatial intelligence” is a magnificent vision and a vague business. Raising $1.23bn buys you runway and talent; it does not tell you who pays, how much, or when. The honest near-term market looks a lot narrower than “machines that understand reality”, and a lot more like “a very good 3D asset and simulation tool for the design and games industries”. That is a fine business. It is not the trillion-dollar one the framing implies.
Second, World Labs is not alone, and its rivals are giants. Google DeepMind has its own world-model line (Genie), Nvidia is building simulation platforms, and every robotics lab on earth wants exactly this technology. A brilliant team with a billion in the bank is still a startup competing with the best-funded labs in history.
Third, generating a pretty 3D world and generating a useful, accurate one are different problems. For entertainment, close enough is fine. For robotics or engineering, where a simulation has to match physics precisely or the robot falls over and the bridge fails, the bar is brutal, and nobody has cleared it yet. The demos are dazzling. Dazzling is the easy part.
How to think about it: bull, bear, neutral
World Labs is private, so you cannot buy it directly. The framing still matters if you follow the sector or invest in the public names around it. (None of this is investment advice.)
Bull: the single most credible team in a field that may be the next platform shift, funded by the exact strategic partners (Nvidia, AMD, Autodesk) who both enable it and would want to acquire it. If world models become as central as LLMs, this is a foundational company.
Bear: a gorgeous research vision searching for a business, in a field crowded with better-resourced incumbents, at a valuation that already prices in success. The near-term revenue is a niche design tool, and the grand version is years and many unsolved problems away.
Neutral: watch Marble’s actual adoption by paying professionals, not the demo view counts, and watch whether Autodesk deepens the relationship or quietly steps back. Strategic investors staying in is the signal; strategic investors drifting is the warning.
What this means
World Labs is the clearest bet that AI’s next chapter is spatial, led by the person with the strongest claim to have started the last one. The technology is real, the team is elite, and the persistent-world trick is a genuine advance. The open question is not whether it is impressive. It is whether “understanding space” becomes a product people pay for at scale, or a beautiful capability that lives inside other companies’ tools. Both are possible. Only one justifies the valuation.
Related on Top Tool Stack: AMI Labs: LeCun’s world-model bet · Accelerated Understanding: physics AI
Did you know: the ImageNet dataset Fei-Fei Li built in 2009 is arguably the spark that lit the entire deep-learning era. If spatial intelligence works out, she will have kicked off two AI revolutions from the same career.