The World Cup Is Now, Are World Models Next?

In March 2026, AMI Labs CEO Alexandre LeBrun predicted that within six months every AI firm would relabel itself a world model start-up to attract funding. Although that claim may be hyperbolic, world models have quickly become one of AI's most capitalized frontiers – vendors such as Dyna Robotics, FieldAI, Odyssey and World Labs have raised around $2 billion collectively to build systems that understand and predict the physical world. For enterprise leaders, the challenge is separating commercial reality from market hype and understanding where this capability is actually useful.

That begins with defining what world models are. Unlike large language models (LLMs), which operate on text and code, world models learn how environments change over time – making them relevant to robotics, autonomous vehicles, industrial operations and immersive video. Achieving this requires architectures purpose-built for physical prediction, with two leading approaches emerging. Video-prediction networks (Google DeepMind's Genie 3, NVIDIA Cosmos, Odyssey) adapt diffusion and transformer architectures to generate future visual frames from past frames. Meanwhile, Joint Embedding Predictive Architecture (JEPA, pursued by AMI Labs) is a genuinely new architecture, predicting outcomes in an abstract latent space rather than at the pixel level, on the basis that this mirrors how humans anticipate the world.

Critically, physics is not programmed into these systems as equations. Instead, behaviours such as gravity or fluid dynamics emerge implicitly from training data, giving world models greater flexibility in messy real-world environments, but also creating a significant dependency on high-quality physical-world data.

Importantly, not every use case is fit for world models. In bounded, well-mapped environments such as factories, refineries and warehouses, digital twins paired with physics simulators and LLM-driven planning are often sufficient, and easier to certify under functional safety standards. World models are most valuable where those systems break down – in environments such as construction sites, mines, outdoor inspection and autonomous driving – where conditions can change rapidly and dynamically, beyond the scope of the models physics engines are trained on.

The technology is very much in its infancy, but momentum is building across three layers. The first is synthetic environment generation, which is already in production. NVIDIA's Cosmos platform is used by firms such as LG Electronics and Samsung to train robotics and industrial vision systems without expensive real-world trials. Similarly, Waymo used Google DeepMind's Genie 3 to build its world model for autonomous driving simulation. The second layer is autonomous operation in unstructured industrial environments. For example, FieldAI has deployed world models into live construction and industrial workflows for inspection, material handling and progress tracking. The third layer is general-purpose embodied intelligence, where AMI Labs sits. These systems reason and act across any physical environment with minimal retraining – although this remains commercially distant.

Enterprise technology vendors must stay abreast of these tech developments and continuously assess the partner network needed to take advantage. Dassault Systèmes moved early, announcing an expanded NVIDIA partnership to build "industry world models" that combine its virtual twin technologies with NVIDIA's foundation models. Other industrial software incumbents have signalled interest, but most have yet to make equivalent public product commitments.

For more research on emerging AI innovations and their real-world impact on enterprise and industrial firms, explore Verdantix AI Applied research or book an inquiry with our analyst team.