There's A Spectrum, And We're Only Looking At One End

World models can be thought of as existing on a spectrum. Both ends share the goal of understanding the physical world, but they take very different approaches to doing so.​

At one end is simulation. World Labs generates purely virtual, fictional environments. NVIDIA's Cosmos and Omniverse simulate the real world with enough fidelity that robots and autonomous vehicles can be trained in synthetic environments before operating in the real one. Google DeepMind's Genie pushes further still, simulating not just visual scenes, but richer scenarios incorporating driving conditions, weather events, radar and LiDAR. That last step signals that the field understands vision alone isn't sufficient.​

At the other end of the spectrum is runtime world modeling. This approach doesn't simulate the world in order to train a model to eventually deploy into it. Instead, it runs AI inference directly on real-world, real-time sensor feeds as events actually happen, where they happen. While simulation-centric approaches aim to eventually reach the real world, they still learn indirectly, through synthetic proxies rather than from the machines themselves.​

Almost all of the investment and attention in world models today sits at the simulation end. The other end remains largely unexplored, though it may be where the biggest opportunity lies.​

The Data Has Always Been There

The irony is that the data needed to build a truly comprehensive world model already exists. It's just been overlooked.​

Every modern industrial machine already has a perception system. Wind turbines, drill rigs, HVAC systems—these machines continuously measure their own physical state through dozens, sometimes hundreds, of sensors. Every rotation, temperature fluctuation and voltage change is recorded in real time. The result is a continuous, high-resolution record that encodes the physical behavior of the machine, its environment and the relationships between them.​

Most of this data either goes unused entirely or feeds into narrow, hand-crafted models designed to detect only what engineers already know to look for. A truly useful world model cannot be built from vision alone, or even from a narrow set of modalities like radar and LiDAR. It must learn from the full richness of the physical world: hundreds of signals across real machines, in real time.​​

The Case For A Physics-First World Model

Real physical systems don't fail cleanly. They degrade, shift and drift across dozens of interdependent variables simultaneously. A turbine doesn't fail because a single sensor crosses a threshold. It fails because the relationship between wind speed, rotation rate, oil temperature and voltage output shifts across dozens of signals in ways no human analyst or classical model can track.​

What's needed is a fundamentally different kind of model: A model that understands the world not through pixels, but through physics.​

This means a model that runs on the actual machine, processing actual sensor data, in real time: a live, continuously updating understanding of the physical world as it unfolds, not a simulation of it. The runtime end of the world model spectrum is where the least investment has gone, but it has a key property that simulation-based approaches overlook. There's no sim-to-real gap, because there’s no simulation.​

But runtime physical sensor understanding isn't enough on its own. The same degradation and drift that defeats hand-engineered rules also defeats global generalization. Two machines of the same make and model diverge over time through wear, installation and environment. A world model trained globally learns the canonical machine, which in reality is no exact machine. ​

The answer is a universal architecture with local adaptation: one foundation model that understands physical dynamics broadly, paired with an ontology that each machine learns autonomously from its own localized sensor history. It's the same architecture everywhere, but with grounded understanding for every asset.​

The Untapped Edge

Every organization running modern physical infrastructure is sitting on decades of sensor data that encodes an extraordinarily precise picture of how its systems actually behave. The bigger opportunity lies in moving from AI that finds known problems to AI that understands physical systems well enough to discover unknown ones.​

This shift can only happen if the model can actually get to the machine. Industrial environments often have limited or no connectivity. Routing inference through a centralized GPU cluster isn't always secure, cost-effective or fast enough. The real test of a physical world model is whether it can run on the hardware that operations teams actually have: on-prem rack servers and edge devices with limited edge GPU or CPU-only configurations. When a model runs at the edge, the economics change: Egress costs drop, data stays protected and uptime is locally controlled.​​

What Comes Next

Simulation world models are making rapid progress, but physical runtime world models remain an open frontier. The organizations that move early will have a significant advantage.

​For teams evaluating solutions in this space, ask a few questions to cut through the noise: Can this model run on the hardware you already have? Does it adapt to your specific machines, or does it assume a canonical one that doesn't exist in practice? Does it learn from your own sensor history?

Modern machines have always had the data. The question is whether the AI we build is finally ready to unlock its potential.​​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?