World-Model Labs Have Billions and No Public Roadmap
Companies with large funding and public attention for world models are keeping secret what they are building.
That was the account on September 20, 2026, from TechCrunch, after its reporter spent time with founders, executives and suppliers in the sector. The silence covers product plans, timelines, and what customer data is used to train. TechCrunch
A world model, in brief, is an AI system that learns how an environment works so it can predict what happens next when something acts in it. The two teams mentioned most often are AMI Labs, linked to Yann LeCun, and World Labs, linked to Fei-Fei Li. Both have raised sums that would usually come with a public roadmap. Neither has shared one.
Michael Rabbat is a co-founder of AMI Labs and its VP of World Models. He said the company is "still in a research and building phase" and is "not talking publicly about any product plans or timeline."
AMI Labs was less than a year old as of September 20, 2026. LeCun had confirmed in December 2025 that he had launched a new world-model startup, a report at the time linked the effort to a sought $5 billion valuation. TechCrunch
In my view, that age helps explain the limited disclosure without justifying all of it. A team under a year old will have unfinished work.
What is visible
World Labs' Marble is described as the most developed product in the world-model space. Demos shown so far cover media creation, building explorable environments for video games, and CGI effects.
The broader context here is that those demos are narrow tests. They do not show whether Marble is the business itself or an early example of something larger.
Alex de Vigan is CEO of Physicl, a data supplier for the world-model business. He said Physicl does not know exactly what world-model companies are building with its data. Customers set the data specification. They do not describe the end system.
The broader context here is that engineers usually rely on close links between data collection and model testing. That link is missing here, which makes the distance stand out.
AMI Labs has disclosed work in manufacturing, biomedicine, robotics, and AI software for doctors through its Nabia partnership. What connects those areas was left unsaid.
In my view, that list looks more like research directions than a product line. Manufacturing and robotics involve physical interaction. Biomedicine and doctor software involve structured decisions with high stakes.
The reporter who wrote the September 18 account had moderated a panel on world models at the All In conference, which provided some of the on-record comments. Participants shared little in that setting.
Capital is not waiting
World Labs was launched after a $230 million raise in September 2024, founded by Li with three colleagues. By February 2026, the company had raised $1 billion in funding. Reuters
AMI raised $1.03 billion. That round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital and Bezos Expeditions. Reuters
Chinese startup ShengShu raised $293 million, saying the funding would support development of a "general world model" that processes sensory information. U.S. startup funding surged 75.6% in the first half of 2025 due to the AI boom. Named potential competitors for world-model labs include other world-model companies, the neolabs, OpenAI and Anthropic.
In my view, that crowded field helps explain the quiet. When large language models broke out, architecture details, training methods and test scores spread quickly. Startups published. Labs wrote blog posts. That openness reduced advantages quickly. World-model founders watched that cycle and seem to have chosen a different path.
The broader context here is what secrecy makes hard to judge for technical readers. Without product definitions, claims about inference latency, or response time, simulation fidelity, or how true to life the simulation is, action conditioning, or how the model reacts to actions, and testing in closed-loop control, or testing while the AI keeps interacting, are hard to assess. A demo of a game world says little about reliability in robotics. A CGI workflow says little about planning over long time spans. The field shares words, but not shared tests.
In my experience with adoption cycles, I am cautious about reading too much into early silence. I raised two children through the PC, web and smartphone shifts. In each case, useful systems looked narrow at first, then widened once developers and users could touch them. My kids did not adopt smartphones because of a keynote. They adopted them because messages, maps and cameras worked.
Well-capitalized teams are hiring, buying data, and building. They are not describing the system under construction. Suppliers do not know the end use. Timelines are undisclosed.
The broader context here is that outside developers have limited material, but not none. The stated interest areas, manufacturing, biomedicine, robotics, physician software, media, games and effects, point to settings where prediction and interaction matter more than text generation. The next disclosure worth watching will not be a funding number. It will be an interface that lets outsiders test whether a predicted world stays consistent when acted upon.
