Why World Model Companies Won't Tell Anyone What They're Building

The Cash Keeps Coming, The Details Never Do

<cite index="19-1">World model startups have already raised a record $3.2B in 2026</cite>, yet investors and the public barely know what most of these companies are building. <cite index="2-5">The big players—Yann LeCun's AMI Labs and Fei-Fei Li's World Labs—have accumulated a lot of buzz and funding, but rank pretty low on the trying-to-make-money scale.</cite> When pressed for specifics, even co-founders clam up. <cite index="3-1,3-2,3-3">Michael Rabbat, a co-founder of AMI Labs, was cagey when asked what the company was working on, saying "We'll talk about it when we're ready to talk about it."</cite>

It's not just the founders being tight-lipped. <cite index="6-6,6-7">Alex de Vigan, CEO of Physicl—a data supplier for the world model business—knows his data has been useful, but he's in the dark about what exactly they're building.</cite>

What Are World Models, Actually?

Before we can understand why everyone's being secretive, we need to know what we're talking about. <cite index="10-1">World models are AI models that can understand basic principles of how the physical world works.</cite> Unlike large language models that predict the next word, <cite index="10-9">world models are trained on multimodal data and can predict the consequences of actions and changes in the environment.</cite>

<cite index="2-6">At their core, world models are about automating spatial intelligence, and could head in lucrative directions from robotics to interactive video to more complex self-driving systems.</cite> <cite index="26-5">World Labs' Marble product already generates persistent 3D environments from simple inputs.</cite> Google DeepMind's Genie, NVIDIA's Cosmos, and Wayve's GAIA-2 are all examples already in the wild.

The Funding Frenzy

<cite index="23-2,23-3">In March 2026, AMI Labs closed a $1.03B seed, while a few weeks earlier, World Labs raised another billion.</cite> These aren't unusual outliers. <cite index="27-1">During 2026, General Intuition announced a $320 million Series A, Odyssey raised $310 million, and Decart raised $300 million.</cite> <cite index="27-8">World Labs and AMI Labs account for about two-thirds of the $3.05 billion raised by six pure-play companies.</cite>

Globally, the momentum is accelerating. <cite index="20-3,20-4">In the first seven months of 2026, China saw 23 new world model startups launch, with 18 of the 23 securing early-stage capital rapidly.</cite>

The Dark Forest Strategy

So why the silence? The answer is strategic and ruthless. <cite index="1-5">It's best to delay competition for as long as possible, which means keeping quiet about exactly what you're building.</cite> <cite index="22-4,22-5,22-6,22-7">No one doubts there are viable businesses to be built on world model tech. If AMI announced they'd built a humanoid system or Hollywood rendering engine, other world model companies, neolabs, and even OpenAI and Anthropic would suddenly be very interested in the space.</cite>

<cite index="1-6">This is what's known as a dark forest scenario: if you don't know who else is in the woods, it's best not to attract attention.</cite> <cite index="1-4">The same money that lets you build under the radar is also funding lots of potential rivals once the path to market becomes clear.</cite>

The Real Bottleneck: Training Data

The secrecy isn't just about competitive advantage—it's about data too. <cite index="27-2,27-3,27-4,27-5">Physical AI has a training problem that language models never had: robots cannot cheaply make billions of mistakes in the real world. An LLM can process enormous amounts of text without breaking anything, but a warehouse robot or autonomous car learning through trial and error can damage equipment or hurt someone. Simulation offers a way to compress experience into a safer training environment.</cite>

<cite index="25-1,25-3">World Model Data, a Cambridge startup, raised £7 million and aims to assemble one million hours of training data by the end of 2026, contrasting that with the current largest comparable database of about 40,000 hours.</cite> Whoever controls the best datasets controls a massive advantage—which companies want to hide from each other.

What This Means for Builders

For developers trying to build on top of world model tech, this silence is frustrating but temporary. <cite index="26-5,26-6">When Marble's capabilities arrive as APIs, no-code builders will be able to add physics-aware rendering to their apps without custom engineering.</cite> The infrastructure is coming, but it's still early.

Right now, you can experiment with some public implementations. Google DeepMind's Genie and NVIDIA's open-source Cosmos are available. If you're building AI agents or robotics applications, try integrating world model reasoning into your planning layer—don't rely solely on language models for spatial decisions. You can also look into platforms like Runway which offer video generation with world model properties.

For enterprises, this is also a signal: the robots and systems you'll deploy in 2027-2028 will be trained differently than today's models. Start thinking about simulation infrastructure now.

The Bet VCs Are Making

<cite index="28-2">Several world model startups have closed Series A rounds north of $50 million without revealing much beyond high-level vision statements.</cite> That kind of bet suggests investors believe the technology will prove so valuable that whoever reaches the finish line first wins everything. <cite index="27-9">Major world-model startups are being financed more like frontier AI labs than normal software startups.</cite>

The silence will eventually break. Product launches will come. APIs will emerge. But for now, we're watching a race where the runners won't tell us where they're going—they just keep running. If you want to keep pace, start building with the world models that have already shipped (like Marble and Genie), and position your stack to integrate spatial reasoning from day one. You can also try platforms like NeonCodex AI, which is building tools to help developers work with emerging AI architectures.

Source: [TechCrunch](https://techcrunch.com/2026/09/20/world-model-companies-are-keeping-a-lot-of-secrets/)