The companies building AI that can simulate reality won't tell you how they're doing it, what data they're using, or when you'll see it work.
The Summary
- World model startups are flush with capital and hype but operating in near-total secrecy, refusing to disclose technical details, training approaches, or even what their products will actually do
- Even their data suppliers won't talk about what they're providing, suggesting NDAs thick enough to stop information at every level of the stack
- The silence points to either genuine technical breakthroughs worth protecting or vaporware sophisticated enough to attract institutional money
The Signal
World models are supposed to be the next frontier after large language models. Instead of predicting the next word, they predict the next frame of reality. Feed one a video of a ball rolling toward a wall, and it should show you what happens next. The promise: AI agents that can plan, simulate consequences, and operate in the physical world without expensive real-world training runs.
The problem is that nobody building these systems will tell you anything. Not the architecture. Not the training data sources. Not the compute requirements. Not even ballpark timelines for when something might ship.
"Everyone in the world-models space is sitting on a pile of cash and a ton of buzz, but good luck getting anyone to tell you what they're actually building."
This isn't normal startup stealth mode. Normal stealth means the founder won't demo the product but will talk about the problem space, the technical approach, maybe drop some research papers. This is stealth mode extending to the data suppliers themselves. The companies providing training data, the people who should have the loosest lips because they're one step removed from the core IP, also won't talk.
That level of coordinated silence suggests one of two things:
- These companies are onto something real and the data moats matter more than anyone outside realizes
- The whole sector is running on demo videos and vibes, and nobody wants to be the first to admit the emperor has no clothes
The money flowing in suggests investors are betting on the first scenario. World model startups have raised hundreds of millions on pitches that amount to "trust us, we're building the thing that makes AI agents actually useful." For agents to work in the real world, they need to predict consequences. To predict consequences, they need world models. To build world models, you apparently need to tell no one anything about how you're doing it.
The Implication
If world models work, they're the missing piece for autonomous agents that can operate in physical space without constant human correction. Robotics, logistics, manufacturing, anything where an agent needs to understand cause and effect in three dimensions. But the secrecy makes it impossible to gauge who's actually close and who's just riding the hype wave with better NDAs.
Watch for the first company to break ranks and ship something real, even if it's narrow. The moment one world model startup shows actual capability, the rest either have to match it or admit they were never close. Until then, assume cash-rich stealth mode means "we're either two years ahead or two years behind, and we're not telling you which."