World models will force a rethink of AI infrastructure – Ready or not

AI is crossing an important threshold. For years, progress was driven primarily by systems that reasoned digitally, predicting the next word, image, or recommendation. Today, AI is increasingly expected to act in the physical world: navigating streets, manipulating objects, operating factories, and interacting with dynamic, uncertain, and physics-constrained environments.

This shift was on full display at CES. Across industries, from autonomous vehicles to robotics to industrial automation to embodied AI, a consistent message emerged: real-world AI systems must be trained, tested, and validated long before they ever touch reality.

During my time at Rivian, working closely with teams focused on autonomous driving and advanced driver-assistance systems, I saw firsthand how important simulation and world modeling had become to real-world autonomy. Long before a self-driving car ever touches public roads, it must learn how to interpret complex, unpredictable environments, not just once, but continuously, across millions of scenarios, especially unusual and edge cases.

Jensen Huang demonstrated this at CES with his video of a car autonomously navigating the streets of San Francisco. What made the moment compelling wasn’t just the demo, but what it represented: years of simulation, validation, and learning happening long before a car drives itself in the real world. As these capabilities move from prototypes to production, world models will increasingly shape experiences we encounter every day, even when we don’t see them.

World models are AI systems that simulate environments, behaviors, and physical laws. They allow machines to explore virtually limitless numbers of scenarios safely, learn cause-and-effect, and make decisions under uncertainty. They are quickly becoming the backbone of physical AI, the domain of AI systems that operate in the real world.

It’s worth comparing world models with the AI systems that most people are already familiar with. Large language models learn from linear sequences of tokens, words, and symbols. World models learn from spatial, temporal, and physical state: frames, sensor readings, object interactions, and cause-and-effect relationships unfolding over time. The result is data that is fundamentally richer and heavier, not just larger in volume but more complex in structure and meaning, and therefore harder to manage, reuse, and operationalize effectively without the right data infrastructure.

Physical AI systems act directly in the world. When they fail, those failures take shape immediately through motion, force, and interactions with people and environments, leaving far less room for trial-and-error learning after deployment. As a result, these systems must be trained, tested, and stress-tested extensively before they ever operate in real conditions.

That requirement fundamentally changes the data management equation. Learning must happen in simulation, across vast numbers of scenarios, and the resulting data, state, outcomes, edge cases, and failures must be retained, compared, and reused as models evolve. The complexity of this data, not just its volume, raises the bar for how it must be managed over time.

World models enable this shift by making it possible to:

• Simulate rare, dangerous, or expensive scenarios
• Stress-test systems before deployment
• Accelerate learning without real-world risk

Real-world physical systems already generate enormous volumes of data. Fleets of vehicles, robots, or industrial machines produce continuous streams of video, sensor telemetry, logs, and state data that quickly scale into petabytes per year.

Simulation multiplies that footprint. Once a system can generate synthetic scenarios on demand, teams routinely create orders of magnitude more simulated data than real-world data to capture edge cases, rare failures, and “what-if” conditions that may never occur naturally. Because simulation is unconstrained by safety or cost, data generation scales faster than reality ever could.

Just as important, this data is not ephemeral. World models don’t train once and move on. They accumulate data over time. Without infrastructure designed to manage data accumulation, teams quickly find that learning slows as data becomes harder to manage, reuse, and trust.

World models succeed not because they exist, but because they improve continuously. That improvement depends on feedback loops:

• Compare simulated outcomes to real-world telemetry
• Identify divergence and failure modes
• Refine the model
• Re-run simulations
• Retrain downstream systems

Each loop produces new data simulation outputs, annotations, validation results, and metadata that must be retained to ensure reproducibility, safety, and trust. Rare events become more valuable, not less. Old data is reused as models evolve. Versions of environments, assumptions, and labels must be preserved.

Over time, this creates long-lived, multi-petabyte data estates that grow continuously throughout the system's lifetime.

At this point, the bottleneck shifts. The limiting factor is no longer just model architecture or compute availability. It also includes iteration velocity, which measures how quickly teams can refine world models using both synthetic and real-world data.

World models place unique demands on data systems. They require infrastructure that can:

• Handle massive volumes of unstructured, multi-modal data
• Preserve lineage, versions, and reproducibility
• Enable reuse across simulation, training, validation, and audit
• Operate across on-premises, cloud, and edge environments
• Enforce governance, security, and sovereignty over time

This is not simply a storage problem. It is a data lifecycle problem, where data must remain accessible, trustworthy, and meaningful for years, often decades, as systems evolve.

In physical AI, data is no longer just an input to training. It becomes a persistent asset that compounds in value as models improve.

As AI systems move from tools to long-lived, learning systems, infrastructure decisions made today will constrain what is possible tomorrow. Leaders must shift their thinking from “Which model should we use?” to “How do we design systems that can learn safely and continuously over time?”

Early advantages quickly fade. The models that looked impressive at launch may not be the ones that matter years later. What matters is how quickly those systems can learn and how reliably they can improve as new data accumulates.

In a world shaped by world models, organizations will generate vast amounts of rich stateful data across simulation and real-world operations. Some will struggle under the weight, slowed by fragmented datasets, limited reuse, and governance challenges that make iteration costly and slow. Others will treat that data as a strategic asset, managed, versioned, governed, and reused deliberately, enabling faster learning and safe evolution.

Over time, that is where advantage compounds, not in a single model or moment but in the infrastructure and practices that allow systems to keep improving. As world models become foundational to physical, AI success will ultimately be decided by how well organizations manage, govern, and reuse complex data at scale.

Learn how NetApp helps you manage, govern, and reuse your data at scale.