World models have moved from research curiosity to a funded commercial category in under a year. World Labs and Advanced Machine Intelligence (AMI) each raised around $1 billion in February and March 2026 respectively, while Runway pulled in $315 million in February. The pitch across all three: AI systems that simulate physical environments, not just generate text.
The category now spans Google DeepMind's Genie 3, unveiled in August 2026; World Labs' Marble, introduced in November; and Runway's GWM-1 family, announced in December. Each takes a different architectural path — Genie 3 layers interactivity onto video generation, Marble outputs 3D assets from text or images, and GWM-1 extends Runway's video-model work into action-conditioned simulation.
The funding surge tracks a growing dissent inside the AI field about whether large language models can carry the industry to general intelligence. Yann LeCun, who ran AI research at Meta for years, has been the loudest name-brand skeptic and has now started AMI to build an alternative stack.
“The idea that you're going to extend the capabilities of LLMs to the point that they're going to have human-level intelligence is complete nonsense.”— Yann LeCun, Former Meta chief AI scientist, founder of AMI
Key facts
- 01World Labs and AMI each raised roughly $1 billion in February and March 2026; Runway added $315 million in February.
- 02Google DeepMind unveiled Genie 3 in August 2026, followed by World Labs' Marble in November and Runway's GWM-1 in December.
- 03Yann LeCun left Meta to found Advanced Machine Intelligence (AMI), betting the company on world models over LLMs.
- 04Fei-Fei Li defines a world model by three criteria: perceptual/geometric/physical consistency, multimodality, and next-state generation from input actions.
- 05MIT's Vincent Sitzmann calls the term 'overloaded' — practitioners at Runway, World Labs, and MIT each give different definitions.
Fei-Fei Li, the computer vision pioneer who co-founded World Labs, framed the same critique differently in a Substack post late last year. LLMs, she wrote, "remain wordsmiths in the dark; eloquent but inexperienced, knowledgeable but ungrounded." She calls spatial intelligence "AI's next frontier" and pitches world models as the technology that will reshape robotics, storytelling, and scientific simulation.
The skepticism is not confined to companies building alternatives. Hugging Face CEO Clem Delangue, whose platform hosts thousands of open LLMs, has publicly argued the language-model boom is nearing its financial ceiling. He sees the wider AI field — biology, chemistry, image, audio, video — as the multi-year growth story.
“I think we're in an LLM bubble, and I think the LLM bubble might be bursting next year.”— Clem Delangue, CEO of Hugging Face
What a world model actually is remains contested. Vincent Sitzmann, who leads MIT CSAIL's Scene Representation Group, defines it broadly as any model that takes an interaction as input and simulates what happens next in an environment. Runway's own definition, published with GWM-1, describes an AI system that builds an internal representation of an environment and uses it to simulate future events within it.
Ben Mildenhall, World Labs co-founder and co-creator of neural radiance fields (NeRF), emphasizes the interaction model as the key differentiator from LLMs. "A very distinguishing aspect of interacting with an LLM is they are turn-based," he said. A world model, by contrast, is meant to be synchronous — the user or agent takes continuous actions in a spatial environment with parallel events happening at once.
“Today, what most people mean when they say 'world model' is generating pixels, so, like, generating a realistic video conditional of the actions.”— Vincent Sitzmann, MIT CSAIL assistant professor, Scene Representation Group lead
Today most systems marketed as world models are, functionally, action-conditioned video generators — they produce pixels that respond to user input. That is a narrower thing than the term suggests, and Sitzmann acknowledges the label is "definitely an overloaded term." Mildenhall goes further: "Ask Runway, ask us, ask whoever — we're all gonna give you a little bit of a different term."
The practical use cases the companies point to are more concrete than the definitions. World models are being pitched for training and testing robots in simulation, generating 3D assets for games and film, and running scientific and physical simulations. Runway's filmmaker customer base is a natural first market; World Labs' Marble tool exports 3D assets that game studios and 3D artists can drop into existing pipelines.
The open question is whether one foundation model can plausibly cover all of that, or whether "world model" will fragment into several distinct product categories held together mostly by branding. The term already carries marketing weight disproportionate to any settled technical meaning, and the companies raising billions are each solving somewhat different problems with different architectures.
There is also the sober version of the bull case. World models are, in research terms, still early — the systems shipping today are impressive video generators with interaction layers, not general-purpose physics simulators. The gap between a playable Genie 3 demo and a robot policy that survives contact with the real world is large, and closing it will require far more than scaling current video architectures.
The category matters less as a bet against LLMs than as a bet on where the next round of AI infrastructure spend goes. If World Labs, AMI, Runway, and Google DeepMind are correct that spatial and physical simulation is the next foundation layer, then the compute, data, and talent flows that built the LLM stack over the past four years get replicated for a distinct set of models — and the companies already positioned in that layer capture the equivalent of the OpenAI and Anthropic seats in the next cycle. That is what the $2 billion-plus raised across three companies in eight weeks is really pricing in.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
