Strategy

Fei-Fei Li's World Labs Has $1.23 Billion and a Contrarian Bet on Spatial AI

While rivals race toward full automation, World Labs is selling a copilot. That strategic choice will either look prescient or costly within the next 18 months.

Written by Vivian Cole · Bureau Business · Jun 18, 2026

The Contrarian Position That Became a Billion-Dollar Company

In the summer of 2023, while Silicon Valley was still genuflecting at the altar of large language models, Fei-Fei Li leaned over to Andreessen Horowitz partner Martin Casado at a lunch convened by Marc Andreessen and whispered a heresy: somebody needed to build a world model. Casado had reached the same conclusion. Within months, World Labs was incorporated. By September 2024 it emerged from stealth with $230 million and a billion-dollar valuation—and no product.

That product, Marble, arrived in November 2024. It uses a class of AI called a world model to generate interactive, navigable 3D environments from photos or text prompts. The technical differentiator is what World Labs calls Gaussian splats: overlapping mathematical blobs that encode how a scene looks from any angle. Once generated, they persist. The view can change; the splats don't move. A simulated piano stays where you put it.

That permanence matters commercially. Traditional 3D scene creation costs $30,000 to $40,000 per scene and takes days. Marble does it for under a dollar in minutes. Lightwheel, which builds virtual environments to train robots, went from recycling the same handful of rooms to generating a thousand distinct rooms per day via World Labs' API.

The B2B Play Underneath the Consumer App

Li is direct about where the money comes from: Marble's tiered consumer subscriptions (up to $95 per month) are a discovery layer, not the business. The real model is enterprise API access—piping Marble's spatial intelligence into other companies' products.

OpenArt, a generative-art platform with 8 million monthly active users, now offers 3D world generation through World Labs' API. Preview.io, backed by Sequoia Capital, uses Marble's spatial controls for AI-powered film production tools serving Fortune 100 automotive clients. Autodesk, which wrote the $200 million check, is planning interoperability between Marble and its Flow Studio 3D modeling suite.

Enterprise pricing is undisclosed, but the ceiling is visible: Anthropic and Salesforce charge upward of $400 per month for high-end enterprise plans. World Labs has room to run.

Where the Strategic Fault Line Runs

The world-model field is crowded and moving fast. Nvidia announced Cosmos 3 and GR00T N2 in March 2026. Google DeepMind's Genie 3 feeds Waymo's self-driving training. Runway hit a $5 billion valuation by pivoting to world models. Yann LeCun left Meta to launch AMI Labs, which has also raised $1 billion.

The competitive divide isn't just technical—it's philosophical. World Labs is building a copilot. General Intuition, Tesla, and others are building autopilots. Nicole Fraenkel of Khosla Ventures, which backs General Intuition, calls Marble 'just an expensive camera' and 'ultimately economically useless,' arguing that agents and robots need to understand what happens in a world when they act in it—not a prettier picture of it.

That critique has teeth. Gaussian splats are a strong solution for visual consistency, but the higher-value robotics and autonomous-agent markets may ultimately reward companies that prioritize physical reasoning over spatial fidelity. Even Casado acknowledges the architecture risk: 'We don't even know if the architecture is right. Literally, nobody's built one before.'

Li's counter is that the creative and enterprise markets are real, paying, and forgiving of imperfect models—and that spatial consistency is the foundation everything else gets built on. Autodesk CEO Andrew Anagnost is buying that argument, literally.

What the Losing Side Looks Like

If the autopilot camp is right, World Labs risks being a well-funded tool company in a market that rewards platform builders. If Li is right, the companies chasing full autonomy will burn capital on compute—Sora reportedly cost OpenAI $15 million per day before the company shut it down in March 2026—while World Labs compounds on a defensible B2B base.

The PwC estimate puts the physical AI market at $503 billion by 2030. The question isn't whether world models matter. It's whether the copilot framing captures enough of that market to justify a $1.23 billion bet—or whether it cedes the highest-margin segments to competitors with fewer scruples about replacing the humans in the loop.

Key takeaways

World Labs raised $230 million at launch in September 2024, then added $1 billion more in February 2026—including $200 million from Autodesk alone—for a total of $1.23 billion raised.
Its Marble app uses 'Gaussian splats' rather than video frames to generate persistent, navigable 3D environments, cutting scene-creation costs from $30,000–$40,000 to under a dollar.
Li's explicit B2B strategy targets film production, architectural design, and robotics training; paying customers include OpenArt (8 million monthly active users), Preview.io, and Lightwheel.
Competitors include Nvidia (Cosmos 3, GR00T N2), Google DeepMind (Genie 3), Runway ($5 billion valuation), and Yann LeCun's AMI Labs—all pursuing world models with varying degrees of automation ambition.
Khosla Ventures, which backs rival General Intuition, publicly dismisses Marble as 'ultimately economically useless'—a sign that the strategic fault line between copilot and autopilot is already hardening into a competitive war.

FAQ

What is a world model, and how does it differ from a large language model?

A large language model predicts and generates text based on patterns in written data. A world model is trained on visual and multimodal data—video, images, sensors, simulations—to understand how physical space works: distance, movement, cause and effect, and 3D consistency. The goal is AI that can reason about the real world, not just describe it.

What does Marble actually do, and who is paying for it?

Marble generates persistent, interactive 3D environments from photos or text prompts using Gaussian splat technology. Paying customers include creative studios, film production companies, and robotics training firms. World Labs offers consumer subscriptions up to $95 per month and undisclosed enterprise API pricing. Its primary business model is B2B.

Why did Autodesk invest $200 million in World Labs?

Autodesk sees world models as the next evolution of design tools—analogous to what ChatGPT did for text, but applied to 3D space and built environments. CEO Andrew Anagnost has cited both the technology's potential and Li's human-centered AI philosophy as reasons for the investment. Autodesk plans interoperability between Marble and its Flow Studio product.

What is the main competitive risk World Labs faces?

The core risk is that the highest-value applications of world models—autonomous robots, self-driving vehicles, AI agents—may favor competitors focused on physical reasoning and full automation over World Labs' copilot-first, visual-consistency approach. Rivals include Nvidia, Google DeepMind, Runway, and Yann LeCun's AMI Labs, all with significant capital and different architectural bets.

How does World Labs' Gaussian splat approach compare to video-based world models?

Video-based world models like Google DeepMind's Genie 3 generate new frames sequentially, which is computationally expensive and loses spatial consistency over time. Marble's Gaussian splats encode a full scene mathematically so it can be viewed from any angle without regenerating content. This makes Marble's environments persistent and cheap to maintain, but critics argue the approach is better suited to visual applications than to the physical reasoning needed for robotics and autonomous agents.

Citations

Inside Fei-Fei Li's $1 billion new AI company, World LabsWorld Labs raised $230 million at launch in September 2024 and an additional $1 billion in February 2026, including $200 million from Autodesk.
Inside Fei-Fei Li's $1 billion new AI company, World LabsTraditional 3D scene creation costs $30,000–$40,000 per scene; Marble can produce a scene for under a dollar in minutes.
Inside Fei-Fei Li's $1 billion new AI company, World LabsRunway hit a $5 billion valuation in 2026 by pivoting to world models; Yann LeCun's AMI Labs has raised $1 billion; OpenAI's Sora was reportedly burning $15 million per day before being shut down in March 2026.
Inside Fei-Fei Li's $1 billion new AI company, World LabsPwC estimated the total market for physical AI at $503 billion by 2030; Nvidia VP Rev Lebaredian told the Financial Times the potential market for world models could be $100 trillion.