World-Model Startups Are Raising Billions While Saying Little

World-model companies are raising large rounds while saying little about what they are building, according to a September 18, 2026 feature from TechCrunch.

The report followed a panel on world models at the All In conference that week, moderated by the publication's author. It named Yann LeCun's AMI Labs and Fei-Fei Li's World Labs as the big players. Both have drawn buzz and funding while ranking low on commercialization, or turning research into products that earn money.

Michael Rabbatt, co-founder of AMI Labs and VP of World Models, declined to say exactly what the company was working on. "We’ll talk about it when we’re ready to talk about it." AMI Labs said it was still in research and building, with no public product plans or timeline. That silence is intentional. As of September 18, 2026, the company was less than a year old.

AMI Labs was described as working on AI that learns from reality, not just from language, like learning by watching the world rather than only reading about it. LeCun confirmed in December 2025 he had launched the startup, then reported to be seeking a $5 billion valuation TechCrunch. It plans to build AI systems using world models that can understand the physical world. It later raised $1.03 billion to build world models TechCrunch.

World Labs, founded by Fei-Fei Li, has moved further toward product. It launched Marble as its first commercial world-model product TechCrunch. TechCrunch described Marble as the most fully developed product in the world-model space, with demos for media creation, explorable environments for video games, and CGI effects.

World Labs is working to advance "spatial intelligence," the ability to understand 3D space, shapes and movement. It raised $1 billion in a round reported in February 2026 Reuters, after an earlier $230 million raise to launch the startup, which Li founded with three colleagues. The company states it builds general-purpose world models in pursuit of spatial intelligence and has introduced Atlas as its next-generation model. World models are also known as world simulators.

The lack of detail is felt upstream. Alex de Vigan, CEO of Physicl, a data supplier for world-model companies, said he remains in the dark about what they are building with his company's data. He said more disclosure would let Physicl build more useful data.

Odyssey, a Palo Alto-based AI lab, secured a $310 million Series B at a $1.45 billion valuation.

The broader context here is a research phase funded like a product phase. For enterprise buyers and infrastructure providers, the practical questions are what data, sensors, physics limits, testing methods, and computing budgets these systems will need. Without that, suppliers and customers cannot plan pipelines or capacity.

In my view, the reticence is worth flagging but not surprising. Labs working on video prediction, 3D reconstruction, action-conditioned simulation, and policy learning have reason to keep designs and data mixes closed while work moves fast. I have watched my own children use generative tools without asking how they worked, a reminder that adoption turns on useful results. If Marble points to near-term use in media, games, and CGI, and Atlas and AMI Labs point to longer-term systems that learn from reality, the disclosure to watch is less about model weights and more about interfaces, controllability, and cost.