World Models: The Next AI Platform
Large language models taught machines to speak. World models will teach them to understand -- and act.
For the past decade, AI progress has been dominated by one paradigm: predict the next token. It worked. LLMs unlocked language, reasoning, and abstraction at a level few expected this fast.
But language is not the world.
World models are systems that learn how reality behaves -- physics, spatial relationships, causality, and time. Instead of guessing what comes next in a sentence, they simulate what happens next in an environment.
That shift is subtle, but foundational.
From Chatbots to Simulation Engines
If you want a robot to fold laundry, drive a vehicle, or manipulate objects in a warehouse, language alone isn't enough. The system needs to understand:
- What objects are
- How they move
- What happens when actions are taken
- How sequences of actions compound over time
This is how humans operate. We imagine outcomes before acting. World models give machines that same capability.
Yann LeCun, Meta's Chief AI Scientist, has been the most vocal proponent of this shift. His argument: LLMs are impressive pattern matchers, but they lack a grounded understanding of the physical world. His proposed architecture -- JEPA (Joint Embedding Predictive Architecture) -- is designed to learn abstract representations of reality rather than generating pixels or tokens.
The Market Is Moving Fast
The capital is already flowing. Serious teams are racing to build the foundational infrastructure for world models:
Meta's V-JEPA 2 -- In June 2025, Meta released V-JEPA 2, a 1.2-billion-parameter world model trained primarily on video. It achieves state-of-the-art visual understanding and prediction, enabling zero-shot robot control in new environments. This is the clearest signal that one of the world's largest AI labs is betting heavily on world models as the next paradigm.
Physical Intelligence -- Backed by OpenAI, Khosla Ventures, and Thrive Capital, Physical Intelligence is training generalist models for robotics. Their goal: a foundation model that can control many different robots across many tasks by learning a shared representation of the physical world. They raised $400M at a $2B valuation in late 2024.
Figure AI -- Figure just closed a $1B+ Series C at a $39B post-money valuation in September 2025. They're building general-purpose humanoid robots that require sophisticated world understanding to navigate real environments. The funding signals massive conviction in embodied AI.
1X Technologies -- The Norwegian humanoid robotics company is reportedly seeking up to $1B at a $10B+ valuation. Their NEO humanoid is designed for home environments -- a use case that demands robust world modeling to handle the chaos of real domestic spaces.
Wayve -- Wayve's GAIA-1 is a 9-billion-parameter generative world model for autonomous driving. Instead of hand-coding rules for every scenario, GAIA learns to simulate driving environments and predict outcomes. This approach could finally crack the long tail of edge cases that has plagued self-driving for years.
Skild AI -- Building a "general purpose brain" for robots, trained in simulation-first environments. Their approach treats robotics as a world-modeling problem, not a narrow control problem.
World Labs -- Founded by Fei-Fei Li, the architect behind ImageNet, World Labs is explicitly focused on building large world models. When the person who created the dataset that launched the deep learning revolution starts a new company, pay attention.
Why This Becomes a Platform Shift
World models don't live in one vertical. They sit under many:
- Robotics -- Humanoids, warehouse automation, surgical robots
- Autonomous vehicles -- Cars, trucks, drones, delivery robots
- Defense -- Autonomous systems, simulation, mission planning
- Manufacturing -- Flexible automation, quality control
- Scientific discovery -- Simulated experimentation, drug discovery
- Games and virtual environments -- NPCs, procedural generation, training data
The pattern is familiar: a horizontal capability emerges, gets commoditized into infrastructure, and enables a Cambrian explosion of applications built on top. World models are following the same trajectory that cloud computing, mobile, and foundation models followed before them.
The Investment Thesis
The bet: In 5-10 years, "world model" will be as common a term as "foundation model" is today -- and the companies building them now will sit at the center of the next AI cycle.
The winners will likely be:
- Foundation model companies that successfully extend from language to world understanding
- Robotics companies that vertically integrate world models with hardware
- Simulation companies that provide the training environments
- Picks-and-shovels plays -- compute, data pipelines, evaluation infrastructure
We're still early. The equivalent of GPT-2 for world models probably hasn't been released yet. But the trajectory is clear, the capital is flowing, and the best teams in AI are pivoting toward this problem.
World models are the missing layer between raw intelligence and real-world action. The companies that build it will define the next decade of AI.
