The $5 Billion Startup Betting on Spatial Intelligence

Imagine what it would be like to shrink yourself to the size of an atom or molecule, when you look around you and you can see small particles whizzing around, you can play 3D tetris taking small drug molecules and figuring out how they fit into bigger protein molecules.

You aren’t doing it alone, your team have also shrunk to the size of Ant Man and you are on a molecular mission together.

This isn’t fantasy!

This is what my company did ten years ago.

Now imagine that you did all that in VR and it actually assessed your understanding of science, no more 2 hour chemistry exams sat at a desk.

This is just one of many use cases of spatial intelligence.

This is part 3 of my deep dive into Fei Fei Li and her $5bn company World Labs.

Today we will looking at the formation of World Labs & the motivation behind it

If you’d rather watch/listen the YouTube video that this essay was adapted from is here:

You can also go back and read part 1 & 2 on my substack, but here’s a quick recap of what you missed:

1. Fei Fei Li’s academic origins, cementing herself as a world leader in AI through ImageNet and bringing together the powerful thruple of data, neural networks and GPUs.

2. How she developed a reputation for being the ethical voice of AI

Before we get stuck in - if this type of thing is your interest and especially if you have made it all the way to part 3 - make sure you subscribe to my substack if you haven’t already. This is a brand new Substack so my first goal is a teeny tiny 100 subscribers - don’t you want to say you were here at the beginning? I know you do.

Visual interrogation has always been the domain of Fei Fei Li. To me, it makes a lot more sense than a language driven AI, I am a big picture thinker afterall and I find language models frustratingly linear. As someone who loves the multidimensional worlds described by physics, and inhabited by humans, why be tied to a single dimension of language models?

While ImageNet and other related work was 2D, Fei Fei Li has been working on video content for a while, expanding to a time dimension. Then in the time period 2015-2017 she and I were both doing similar things again (albeit at different scales). I was leading a startup building 3D visualisations of dynamic atoms and molecules, as she was looking at modelling genomes.

Yes, I do feel inadequate - let’s move on.

The healthcare work that is described in Fei Fei Li’s book also includes spatial modelling, so it makes sense that Fei Fei Li and others working across AI can see a different approach for the development of AI.

On May 3 2024 Reuters exclusively reported that Fei Fei Li had started a spatial intelligence company “World Labs”and Fei Fei Li said:

“The pursuit of visual and spatial intelligence has been the North Star guiding me since I entered the field. It’s why I spent years building ImageNet… It’s why my academic lab at Stanford has spent the last decade combining computer vision with robotic learning.”

Even though it was announced in May 2024, it started in stealth mode.

World Labs was founded in January 2024 by four co-founders: Fei-Fei Li (CEO), Justin Johnson, Christoph Lassner, and Ben Mildenhall. The company was created to develop 3D environment generation with real-time simulation capability, a field Li had come to call “spatial intelligence.”

Justin Johnson is a computer vision researcher and former PhD student of Li’s, he spent time at Meta and University of Michigan, but came back together with Fei Fei Li after they realised that their research interests were converging.

Christoph Lassner brings expertise in 3D graphics from prior roles at Meta Reality Labs and Epic Games, and had previously developed Pulsar, a differentiable renderer that laid groundwork for a technique called Gaussian Splatting.

Ben Mildenhall, formerly at Google Research, was the co-creator of NeRF (Neural Radiance Fields), a landmark method for generating photorealistic 3D scenes by encoding geometry and appearance within neural networks.

The initial valuation of World Labs was $200m but by Sept 2024 it had completed a couple of funding rounds and increased to a valuation of $1bn, demonstrating a textbook example of investing in the reputation of the founding team and a big vision.

So as World Labs emerged from stealth on 13 September 2024 announcing $230 million in total funding, co-led by Andreessen Horowitz, NEA, and Radical Ventures. Other investors in this round included Marc Benioff (Salesforce CEO), Ashton Kutcher, Adobe Ventures, AMD Ventures, Databricks Ventures, and NVentures, the venture arm of Nvidia. Notable angel investors included Jeff Dean of Google and former Google CEO Eric Schmidt.

Radical ventures are particularly interesting due to the responsible AI framework that is written into their investment terms. I’ll write more on responsible AI investing frameworks, where I will delve into this further.

In the early days of World Labs, Google became the main compute partner, which is not surprising based on the relationship that Fei Fei Li has with google. Initially, in December 2024, World Labs released its first public demonstration: early models capable of generating interactive 3D scenes from a single image. This wasn’t yet wildly groundbreaking.

Despite the decades spent in research Fei Fei Li, and the team are new at the tech startup game, and when you listen to them talk about the startup journey it is clear that they are doing what all early stage start up founders have to do. They are figuring out product market fit and trying to show some market traction. The only difference here is that they had hundreds of millions of dollars to do it - it makes the first funding of £80k in my business, look like a very measly amount of money!

In November 2025, World Labs publicly launched Marble, its first commercial product. Marble is described as a “large world model” that can reconstruct, generate, and simulate 3D worlds from text, images, videos, or coarse 3D layouts, allowing users to move through, edit, and inhabit the results.

The strategy with the launch of Marble is two fold:

1. To put a product out there - in this case a 3D image generator, which can create 3D “worlds” from text or image inputs.

2. This first product could also help with the direction of travel towards the vision of full spatial intelligence.

Marble was made available across four pricing tiers, from a free plan (four world generations) to a professional plan at $95 per month with commercial rights and export to Unreal Engine, Unity, Blender, and Houdini - engines that are used to 3D models.

I’ve tried playing with Marble, the UI is not great, the UX is worse and Justin and Fei Fei openly admit that they need some product people on the team. It is not immediately clear how this is different from the games engines that have always existed, Justin’s blog on the architecture begins to explain it.

The emphasis here is on machine learning. The question that they are trying to answer:

Can a machine predict what will happen next - even without humans giving the rules, just by learning from training material?

An even deeper thread of curiosity expressed by Fei Fei Li and Justin Johnson:

Can it learn the physics rules not by being taught them explicitly but through observation?

After all that is what humans do - we inherently learn all the basic rules of physics when we are toddlers, learning to walk (gravity), playing with cars (force = mass x acceleration) and conservation of momentum when we play conkers.

We might not be able to write those equations or explain the concepts with physics language, but we inherently know what is going on. We know if we push something off a high surface it will fall to the floor, it happens every time, it doesn’t just occasionally float in mid-air before falling like the Wile E. Coyote.

Even more fascinating, if an AI was asked to observe then come up with a set of rules, would it even come up with Newtonian physics equations?

This is a question that World Labs would love to be able to answer, but it also sounds like a very academic question.

But academic questions alone do not attract $1bn dollars - so what is the pitch for World Labs?

“Vision has long been a cornerstone of human intelligence, but its power emerged from something even more fundamental. Long before animals could nest, care for their young, communicate with language, or build civilizations, the simple act of sensing quietly sparked an evolutionary journey toward intelligence.” Fei Fei Li - Substack 2025

Fei Fei Li can see the value of AIs having Spatial intelligence, which is the ability to visualise, understand, and mentally manipulate objects, dimensions, and physical spaces. She has previously said that AGI is not possible without it. Can anything really be deemed intelligent if the only understanding of the world around us is fed secondhand (albeit at scale), and delivered in language, even if that language writes code which can “do” things?

The Cambrian explosion for robots sounds like Fei Fei Li’s North Star, it also sounds terrifying. And it is what World Labs is here to do.

Fei Fei Li is not the only one that thinks this, the rest of her founding team are equally passionate about Spatial Intelligence, by the looks of their limited online presence.

Other researchers are also building world models, the models that generate 3D spaces and predict the next “state” of a physical system. Estia Ryan at Eka Ventures has written a great Substack on this.

Google Deepmind’s, Project Genie, Yann Le Cun’s JEPA, NVIDIA’s Cosmos are examples of world models - so the biggest names in the world are investing heavily into world models.

World Labs investors are bought in too, in February 2026 Investors including Autodesk, Nvidia, AMD, Andreessen Horowitz, Fidelity, Emerson Collective, and Sea, invested $1bn at a valuation of $5bn

The current technology is still relatively limited. So use cases are very helpful. Here are some of the World Labs case studies:

• Robotics simulation (with Stanford researchers): Researchers used Marble to generate 3D kitchen environments for robot manipulation training, imported into the MuJoCo simulator to train a robotic arm.

• Scaling robotic evaluation (with Lightwheel): Marble was used to generate diverse simulation environments spanning homes and factories to help Lightwheel build a scalable Real2Sim pipeline.

• Architectural and interior visualisation (with Fenestra and Interior AI): Fenestra integrated Marble to let architects step inside their own concepts. Interior AI became the first consumer-facing app to use Marble.

• Film production and VFX: World Labs’ own launch video was built using Marble itself. Marble’s 3D assets let artists control camera movements with frame-perfect precision.

• Education and training simulations: In December 2024, World Labs announced an integration with EON Reality to power AI-based spatial learning environments, aiming to bring world-model generation to classrooms and corporate training platforms. This is not yet deployed at scale and no outcomes data has been published yet.

I spent 10 years trying to find the business model for 3D environments into teaching and learning, I didn’t have 1bn dollars to do it though! Our team developed tools which showed very positive learning outcomes. However, the use case that I wanted to see come to life was a totally different form of assessment - able to measure much more sophisticated and nuanced skills all without having to sit at a desk under exam conditions. There are many hurdles to cross before any of this gets implemented into the classroom. Even if technically possible, its not enough to see benefits, what are the risks, or widely perceived risks, steamrolling tech into a classroom has consistently led to a massive backlash.

There are also applications discussed as future targets but not yet deployed:

• Autonomous vehicles: this is an identified use case but no published case study.

• Scientific discovery: Repeatedly cited as a long-term goal. No concrete applications exist yet. I have one - we made one - maybe I’ll keep that quiet for now, the more I make this video the more I think that we developed more in my startup than we realised - maybe I have transferable skills afterall.

• Healthcare: This is referenced repeatedly by Fei Fei Li but World Labs has published nothing on healthcare applications of Marble yet!

As of mid-2026, World Labs is an early-revenue stage company with one commercial product, approximately 60 employees, and an ambitious roadmap spanning robotics, scientific discovery, autonomous vehicles, and healthcare, most of which remain stated ambitions rather than deployed applications. It’s not clear if they are getting much public revenue from Marble, but I assume that partnerships will be building use cases as mentioned above.

I would be interested to hear from VFX and game designers, is the application of World Labs in your work a nightmare use of AI, or is it genuinely a helpful way for you to be more effective and give you useful tools for your art?

It’s been funny reading about all these models knowing that my startup actually did some of this, although in a simple way. We were able to create 3D spaces, based on physics, when someone would interact with particles, the system would change, based on the laws of physics, the system itself was not learning at that time, but it was dynamically responding to physical movements and inputs by the user.

I don’t think I fully appreciated the quality of our work (the technical side was led by CTO Phill Tew), we were a tiny start up with less than £0.5m initial investment (and 100k cloud credits from Oracle - eek) and we could create environments that could take real time feedback and respond to it - in 2016. I used to say if only we could get £10m we could do something massive with this! Imagine what we could have done with $1bn.

But the question of if the benefits would outweigh the negative consequences of the work was one that we never fully answered, and the question still remains.

In this series we’ve discussed the research foundation of World Labs, originally spearheaded by Fei Fei Li - what happens when silicon valley start up money and ethics collide - that is for part 4, when we dig into the controversies.

If you are still here, I’d love to hear what you think of Fei Fei Li’s story so far, do you think she will live up to the position of “godmother of AI” or the conscience of AI?