Why World Model Startups Are Sitting on Billions and Telling No One Anything

TL;DR

• World model startups like World Labs, Odyssey and Decart have raised hundreds of millions on the promise of AI that understands 3D space and physics, but have revealed almost no technical details about how their systems actually work.
• Founders and even their data partners are refusing interviews and staying vague on roadmaps to protect proprietary data deals, avoid tipping off Google DeepMind and Meta, and buy time while the tech remains unreliable.
• That secrecy is slowing independent evaluation and raising fears of a hype bubble, legal fights over training data, and a future where spatial intelligence is locked behind a few closed APIs.

The Billions Are Real

Silicon Valley has not been this excited — or this confused — about a category since the first ChatGPT boom. World Labs, the startup founded by Stanford AI pioneer Fei-Fei Li, has raised more than $230 million at a valuation north of $1 billion from Andreessen Horowitz, NEA and Radical Ventures. Israeli startup Decart has pulled in over $150 million for its real-time interactive models. Odyssey, founded by former Cruise and Voyage leaders Oliver Cameron and Jeff Hawke, has raised more than $45 million from EQT Ventures, GV and Air Street Capital. Luma, Runway and a wave of younger labs are all rebranding around the same phrase: world models.

The pitch is enormous. Not chatbots that predict the next word, but AI systems that generate persistent, navigable 3D worlds, understand physics, object permanence and spatial reasoning, and eventually power everything from video games and films to robots and self-driving cars. Investors call it the next step to artificial general intelligence.

But ask what is actually under the hood, and the conversation stops.

A Demo, Not A Paper

Unlike the large language model era, where startups and labs competed to publish papers, release weights and top leaderboards, world model companies are shipping tightly controlled demos with almost no technical disclosure.

World Labs released Marble, its browser-based text-to-3D environment tool, in June 2025, followed by a first large world model preview that turns a single image into an explorable 3D scene. Odyssey launched its Explorer tool for interactive streaming video worlds. Decart went viral with Oasis, a playable AI-generated Minecraft clone, and later a real-time conversational video model. DeepMind upped the pressure in August 2025 with Genie 3, a real-time interactive world generator.

All are impressive in a browser tab. None came with architecture details, training compute numbers, evaluation benchmarks, or clear roadmaps for pricing, API access, or when these toys become reliable products. Requests for model cards, data sources and failure rates are met with waitlists and marketing language about spatial intelligence.

The Founders Who Won't Talk

That silence is deliberate. In recent months, founders of three of the most-funded world model startups have declined in-depth technical interviews or pulled out of conference panels, even as they continue to raise and hire aggressively. Off the record, employees and investors give the same reasons: no one wants to show their hand.

First, the race is unusually paranoid. With Google DeepMind, Meta, Nvidia and OpenAI's Sora team all building competing video-to-world systems, a single detail about data pipelines, rendering engines or hybrid diffusion-transformer designs could be copied in weeks.

Second, the tech is still brittle. Anyone who has spent time in an AI-generated world knows the problem: walls melt, objects disappear when you turn around, avatars fall through floors. Founders prefer to talk about long-term vision — persistent worlds for creators and robots — rather than admit their current models can only hold coherence for a few minutes.

Third, hype is the strategy. In a market where talent and GPU capacity are scarce, mystery helps. Being the stealthy Fei-Fei Li company or the secretive ex-Cruise team is better for recruiting and fundraising than publishing a benchmark where you lose to Genie 3.

Even The Data Suppliers Have Gone Quiet

More telling is who else won't talk: the data suppliers. World models need something LLMs never did — massive libraries of 3D geometry, video with camera movements, drone footage, game engine renders, and licensed scans of real-world places.

Companies that provide that fuel, from stock video platforms and photogrammetry studios to game asset marketplaces and autonomous vehicle fleets, have signed exclusive deals with world labs but now refuse to name clients or discuss terms. Two data executives told reporters this summer they were barred by NDAs from confirming partnerships, citing fears of copyright lawsuits and backlash from creators who never consented to having their 3D work used for training.

It echoes the early LLM data wars, but with higher stakes. A text scrape is one thing. A licensed 3D replica of a city, a national park or someone's Unreal Engine environment raises new questions about ownership, consent and price. No startup wants to be the test case, so no one discloses where the worlds come from.

What Secrecy Means For The Future

The short-term winner from secrecy is the incumbents. DeepMind, Meta's WorldGen efforts and Nvidia's Cosmos platform can afford to experiment in public because they already own data, distribution and compute. Startups, by contrast, are trying to build moats out of mystery.

But researchers warn of three long-term costs.

One is evaluation chaos. Without open benchmarks for physics consistency, memory and controllability, there is no way to compare World Labs vs. Odyssey vs. Decart vs. Genie 3 except vibes from cherry-picked demos. That makes enterprise adoption almost impossible for gaming studios and robotics firms that need reliability guarantees.

Two is a legal time bomb. If training data deals stay hidden, Hollywood studios, game publishers and 3D artists say they cannot audit infringement or negotiate fair licensing. Lawsuits that hit image and video generators in 2023-2024 could return even larger for 3D worlds.

Three is centralization. The dream of world models was open, creator-driven universes anyone could build. The current path points to the opposite: a handful of closed APIs where you rent explorable worlds by the minute, with pricing, limits and safety filters controlled entirely by the lab.

Until one company breaks ranks and actually explains how its worlds are made — what data, what model, what roadmap — customers, researchers and even investors are left betting billions on faith.

AndroGuider Team

Articles written by the AndroGuider team. We try to make them thorough and informational while being easy to read.