In January 2026, Yann LeCun walked away from Meta after 12 years to start something new. Two months later, his startup AMI Labs closed a $1.03 billion funding round at a $3.5 billion valuation. The mission was not to build a better chatbot. It was to build something that understands the physical world the way humans do. LeCun calls it a world model, and he believes it is the next frontier of artificial intelligence.
He is not alone. Fei-Fei Li’s World Labs launched its first commercial product in November 2025. Google DeepMind released Genie 3, its real-time interactive world model, in August. Collectively, these three projects have attracted billions in funding from investors who believe that large language models have hit a ceiling. For founders watching from the sidelines, this shift represents both a warning and an opportunity.
What World Models Actually Are and Why They Matter
Large language models like GPT and Claude predict the next word in a sentence. They are powerful, but they do not understand physical reality. Ask a language model to predict what happens when you push a glass off a table and it will give you a plausible answer. But it does not actually know what gravity is. It has never seen a glass fall. It is pattern matching, not reasoning.
World models take a fundamentally different approach. Instead of learning from text, they learn from video and sensory data. They build internal representations of how objects move, how physics works, and how actions lead to consequences. LeCun’s architecture, called JEPA (Joint Embedding Predictive Architecture), trains AI systems to predict what will happen next in a visual scene without needing explicit labels or instructions.
The practical difference is enormous. A world model can simulate a warehouse floor before a robot ever touches a box. It can generate rare driving scenarios, like a pedestrian stepping into traffic during a snowstorm, to train autonomous vehicles without waiting for those events to happen in real life. It can model chemical reactions, climate patterns, or surgical procedures in ways that text-based AI simply cannot.
The Three Companies Leading the Race
AMI Labs is the highest-profile entrant. LeCun serves as executive chairman, with Alex LeBrun (formerly of medical AI company Nabla) as CEO. The company is headquartered in Paris and plans to target healthcare, robotics, wearables, and industrial automation first. Its first partner is Nabla, which will use world models to advance AI-driven medical applications beyond what current language models can offer.
World Labs, founded by Stanford professor Fei-Fei Li, took a different route by shipping a product early. Its platform called Marble generates 3D environments from text prompts, photos, or videos. Pricing starts free with a $20 per month standard tier and goes up to $95 per month for commercial rights. The output is compatible with Unreal Engine and Unity, making it immediately useful for game developers, VFX studios, and VR creators building for Vision Pro and Quest 3.
Google DeepMind’s Genie 3 is the most technically ambitious. Released as a research preview, it runs at 24 frames per second at 720p resolution with up to one minute of lookback memory. It is the first general-purpose world model that operates in real time, learning physics through training data rather than hard-coded rules. DeepMind is gradually expanding access to academic and creative communities.
Why Founders Should Pay Attention Now
PitchBook projects the market for world models in gaming alone could grow from $1.2 billion (2022 to 2025) to $276 billion by 2030. That is not a typo. The venture capital market is already responding, with world model startups attracting some of the largest seed rounds in AI history. SpAItial, a European startup building spatial AI, raised a $13 million seed round, unusually large for the region.
But the opportunity is not limited to building world models. It extends to building on top of them. Just as thousands of startups were built on top of GPT and Claude, the next wave will be built on world model APIs. The applications span robotics training (generating synthetic environments instead of expensive real-world data), autonomous vehicle simulation, architectural visualization, medical procedure modeling, and industrial automation.
Gartner predicts that 40% of enterprise applications will embed AI agents by the end of 2026, up from less than 5% in 2025. World models provide the foundation for agents that can actually understand their environment rather than just process text about it. For founders building in AI-powered businesses, this is the infrastructure layer that makes the next generation possible.
What This Means for the LLM Boom
LeCun has been blunt about his view. He told MIT Technology Review that scaling up language models will not lead to artificial general intelligence. The limitations are structural, not just computational. Language models hallucinate because they have no grounding in physical reality. They cannot plan, they cannot reason about cause and effect, and they cannot learn from experience the way humans do.
That does not mean LLMs are going away. The more likely outcome is a hybrid approach where language models handle communication and reasoning about text while world models handle understanding of physical environments. The companies that figure out how to combine both will have a significant advantage.
For startup founders, the strategic takeaway is clear. The AI stack is getting deeper. The winners of the next five years will not just be the ones who build clever wrappers around language APIs. They will be the ones who understand where world models fit and build products that leverage spatial understanding, physics simulation, and environmental reasoning alongside traditional language capabilities.
The Window Is Open
Every major platform shift creates a window where startups can compete with incumbents. The shift from web to mobile created Uber, Instagram, and Snapchat. The shift from on-premise to cloud created Slack, Zoom, and Stripe. The world model shift is earlier but the pattern is the same. The infrastructure is being built right now by LeCun, Li, and DeepMind. The application layer is wide open.
The founders who move first will not need to build world models from scratch. They will need to understand the capabilities, identify the highest-value applications, and ship products while the market is still forming. That has always been the startup advantage: speed, focus, and the willingness to build before the playbook exists.
