The Incredibly Inspiring Startup Story of Dr. Fei-Fei Li and World Labs

Fei-Fei Li, co-founder of World Labs, has built a company that reached a $1 billion valuation within roughly twelve weeks of emerging from stealth in September 2024, after raising $230 million initially and $1.23 billion total from investors including Nvidia, AMD, and Autodesk. The startup, which shipped its first model Marble in 2025 and Atlas on September 1, 2026, aims to achieve 'spatial intelligence' by creating AI that understands physical 3D space, building on Li's earlier work creating ImageNet, the dataset that powered the 2012 AlexNet breakthrough in deep learning.

Fei-Fei Li landed in New Jersey as a teenager with less than $20 and a family that would soon be running a dry-cleaning store to survive. Forty years later, she is running a company that hit a billion-dollar valuation in about twelve weeks. Neither part of that sentence explains the other on its own, and that's exactly the point. - Fei-Fei Li immigrated from Beijing to New Jersey as a teenager with under $20 to her name, and her family ran a dry-cleaning store while she studied her way to a full scholarship at Princeton and a PhD from Caltech - She built ImageNet, a dataset originally projected to take 19 years to complete, and in 2012 it powered the AlexNet breakthrough that kicked off the modern deep learning era - In October 2023 she co-founded World Labs to pursue "spatial intelligence," the idea that AI cannot be truly general until it understands physical, three-dimensional space the way humans do - World Labs raised $230 million to emerge from stealth in September 2024, reached a $1 billion valuation in roughly twelve weeks, and has since raised $1.23 billion total from Nvidia, AMD, and Autodesk among others - The company shipped Marble in 2025 and Atlas on September 1, 2026, models that build persistent, explorable 3D worlds from a single photo, turning the spatial intelligence thesis into a shipped product Start with the arrival, because everything else about Fei-Fei Li reads differently once you know it. She was born in Beijing in 1976 and emigrated to the United States as a teenager, leaving behind what her family describes as a comparatively prosperous life in China for something much harder in New Jersey. The family borrowed money to open a dry-cleaning store. Li studied five days a week and worked the store on weekends, learning English largely on the job, at the register, with customers who had no idea the teenager taking their shirts would eventually be called the godmother of AI. That detail matters for reasons beyond biography. It is the exact opposite of the founder narrative Silicon Valley usually tells about itself, the one where conviction comes from confidence and confidence comes from a head start. Li's conviction came from somewhere else entirely, from having watched her family rebuild an entire life from almost nothing and succeeding anyway. People who have done that once tend not to be intimidated by doing something difficult a second time. The unglamorous years that made ImageNet possible Li earned a full scholarship to Princeton, where she studied physics, and went on to a PhD in electrical engineering at Caltech. Neither of those is a computer science credential, which is worth sitting with, because the field she would go on to define didn't fully exist yet in the form she'd eventually build it. In 2007, working with Princeton professor Kai Li, she started ImageNet. The goal: a dataset of images labeled across tens of thousands of categories, structured the way WordNet had structured language. At the outset, the project was estimated to take nineteen years to complete by conventional methods. Nineteen years is not a research timeline. It's a career, or most of one. She and her team found a way to compress it dramatically through crowdsourced labeling, which was itself a fairly radical bet on an unproven method at the time. Fei-Fei Li's World Labs Launches Atlas to Build Walkable AI Worlds https://startupfortune.com/fei-fei-lis-world-labs-launches-atlas-to-build-walkable-ai-worlds/ World Labs, the spatial AI startup co-founded by Fei-Fei Li, launched Atlas on September 1, 2026, an AI model that builds persistent, explorable 3D worlds from a single photo. The release skips a research paper, pricing, and named partners even as World Labs races Google DeepMind's Genie 3 and Nvidia's Cosmos for the emerging spatial AI category. - how to build 3D worlds from single photographs https://startupfortune.com/fei-fei-lis-world-labs-launches-atlas-to-build-walkable-ai-worlds/ - AI system that generates explorable walkable 3D environments https://startupfortune.com/fei-fei-lis-world-labs-launches-atlas-to-build-walkable-ai-worlds/ Here's the part that made it matter beyond academia. In 2012, a convolutional neural network called AlexNet, trained on ImageNet, entered the annual ImageNet competition and beat the next-best entrant by ten percentage points. That gap is enormous by the standards of incremental research progress. It is generally treated as the moment deep learning stopped being a niche academic pursuit. From there it became the thing that would eventually power almost every AI product built since. Li had spent five years building the dataset nobody else thought was worth the trouble. The field spent the next decade proving her right. She joined Stanford as an assistant professor in 2009, before ImageNet's payoff was obvious, and stayed through the moment it became one of the most consequential datasets in the history of computing. That timing is worth noticing on its own. Most researchers chase the result first and the faculty position second. She did it backwards. She took the position while the result was still an open question, betting years of an academic career on a dataset that hadn't yet proven anything to anyone outside her own lab. In 2019 she co-founded the Stanford Institute for Human-Centered Artificial Intelligence. She also did a stint as Chief Scientist of AI and Machine Learning at Google Cloud. That experience, she has said, taught her how much private industry involvement actually shapes what research becomes useful in the world, rather than staying trapped in a paper. Betting on the thing nobody was building By 2023, Fei-Fei Li was already one of the most decorated researchers in AI. She didn't need to start a company. She started one anyway. In October 2023 she co-founded World Labs alongside AI researchers Justin Johnson, Christoph Lassner, and Ben Mildenhall. It was built around a thesis that ran directly against where the rest of the industry had put its money. The thesis is called spatial intelligence, and the short version is this: language models can predict the next word extremely well, but predicting words is not the same as understanding a physical world with depth, occlusion, gravity, and consequence. Li's argument, stated plainly, is that you cannot get to real artificial general intelligence without machines that natively reason about three-dimensional space the way any child does before they can read a single sentence. That is a genuinely contrarian position to stake a company on in 2023, when the entire industry's capital and attention were pointed at scaling language models bigger and bigger. World Labs emerged from stealth in September 2024 with $230 million in financing at a $1 billion valuation. That's unicorn status in roughly twelve weeks, one of the fastest such runs in generative AI's short history. The company has since raised a total of $1.23 billion. Backers include Nvidia, AMD, and Autodesk, three companies whose core businesses all depend, in different ways, on machines getting better at understanding physical space. That is not a coincidence of investor logic. It is the thesis being validated by the exact companies with the most reason to know whether it holds up. From thesis to shipped product Conviction is common in AI right now. Product is not. World Labs shipped Marble in 2025. It generates realistic, editable 3D worlds usable inside existing game and visual effects pipelines. That's the kind of tool that turns a theory about spatial reasoning into something an artist can actually open and use on a Tuesday. On September 1, 2026, the company released Atlas, a multimodal world model that builds persistent, explorable 3D environments from a single photograph, with precise camera control over the result. Feed it one image. It constructs a navigable space around it, not a flat reconstruction but something closer to a world you can walk through. That is spatial intelligence made literal and shippable. Not a research demo gesturing at a future capability. A product doing the thing the thesis said machines needed to do. This Week in AI: Robots Take a Punch, Assistants Get Long-Term Memory, and Models Keep Getting Cheaper https://startupfortune.com/this-week-in-ai-robots-take-a-punch-assistants-get-long-term-memory/ A dense week across robotics, model releases and AI research: humanoid robots duke it out, ChatGPT gets long-term memory, and frontier-adjacent AI keeps getting cheaper and faster. - how to build robots that handle physical impacts https://startupfortune.com/this-week-in-ai-robots-take-a-punch-assistants-get-long-term-memory/ - AI models getting cheaper and more accessible features https://startupfortune.com/this-week-in-ai-robots-take-a-punch-assistants-get-long-term-memory/ The gap between Marble and Atlas, in under two years, is the real story here, more than either product on its own. It's the difference between proving a concept and building a company that can iterate on it fast enough to matter before the rest of the industry catches up. Plenty of researchers have had a correct contrarian idea. Very few have then built the organization capable of turning that idea into two shipped models and $1.23 billion in backing within three years of founding. The through-line, and why it should matter to you Li published a memoir, "The Worlds I See," in November 2023, the same year she started World Labs, and one line from it explains her career better than any funding announcement could: "I think that when you see something that's too early, it's often a different way of saying 'We haven't seen this before.' In hindsight, we bet on something we were right about." That is the whole pattern, repeated three times now. ImageNet looked like an absurd, oversized bet on labeled data when nobody wanted to fund nineteen years of annotation work, and deep learning itself barely registered as more than a fringe academic interest until AlexNet's ten-point margin made it undeniable. Spatial intelligence looks just as strange today: a company built on the bet that the industry's rush toward ever-larger language models was missing something. Each time, the field caught up to a conviction Li had already been building toward for years, quietly, without waiting for consensus to arrive first. For founders, the lesson isn't "immigrate with $20 and everything works out." That flattens an extraordinary personal story into a motivational poster. The actual lesson is narrower and more useful. The biggest wins tend to sit on the far side of a long, unglamorous stretch of work that looks unjustifiable to everyone who hasn't done the thinking you've already done. Li spent five years on a dataset people thought was oversized before it became the foundation of an entire industry. She's now three years into a company built on a thesis most of the industry still hasn't fully priced in. Being early and being wrong look identical for a long time. There's only one way to find out which one you are. Keep building until the field either catches up to you or proves you should have listened to it in the first place. So far, for Fei-Fei Li, it has been the former every single time.