"Before machines can act in the world, they must learn to simulate it."
— The Cubed Thesis

Why Physical AI Hits a Compute Wall

World Models Need More Than Text and Static Images

Robots and autonomous systems must understand geometry, motion, object permanence, cause and effect, and the consequences of action. That requires multimodal training and persistent simulation far beyond ordinary language-model inference.

Illustrative Physical-AI Workload Mix

Real-World Data Is Expensive—and Failure Is Physical

Collecting robot trajectories, autonomous-driving scenes and industrial edge cases takes hardware, operators and time. Simulation can generate rare scenarios safely, but only when teams can access large, burstable GPU clusters without paying for idle capacity.

One workload, many compute stages

Data + Simulate + Train + Evaluate + Deploy

Opportunities: World Models & Physical AI

A New Demand Curve for GPU Compute

World models combine video understanding, generative simulation, reinforcement learning and real-time inference. Each stage needs a different mix of accelerators, memory, bandwidth and latency—creating demand for intelligent compute routing.

World simulation
Multimodal training
Synthetic data
Edge inference

Simulation Scales Intelligence Before Deployment

World models let physical AI systems learn from vast simulated environments before they face the cost and risk of real-world deployment.

Category Simulation-First Physical AI Real-World-Only Training
Scenario Coverage Generate rare, dangerous and long-tail conditions on demand. Wait for scenarios to occur and be captured.
Safety Test failures in virtual environments before deployment. Failures can damage hardware or endanger people.
Iteration Speed Run thousands of environments in parallel. Bound by fleet size, operators and elapsed time.
Compute Profile Bursty, accelerator-heavy training and simulation jobs. Continuous hardware and data-collection overhead.
Reproducibility Replay identical worlds, seeds and interventions. Physical conditions are difficult to reproduce exactly.
Data Scale Programmatically create labeled trajectories at scale. Manual collection and labeling remain bottlenecks.
Edge Cases Actively search for model weaknesses. Biased toward common, already-observed events.
Feedback Loop Real-world outcomes continuously refine simulated worlds. Learning remains tied to new physical collection cycles.