Learned Simulation as Infrastructure

This pushes General Intuition beyond selling agent behavior into selling the testbed where behavior gets trained and measured. A learned simulator lets a robotics or game team run thousands of cheap virtual rollouts, watch how an agent reacts when parts of the world are hidden or other agents behave unpredictably, and generate synthetic trajectories before risking hardware time. That is valuable because standard physics simulators are strong at rigid rules, but weaker at messy real scenes and social interaction.

• The closest analogue is NVIDIA’s Cosmos plus Omniverse stack. NVIDIA is positioning world foundation models as a layer for future state prediction, synthetic data generation, and physical AI simulation, which validates learned simulation as a real infrastructure category, not just a model feature.
  2 nvidia 3 nvidia 4 nvidia
• Applied Intuition shows how big the simulation layer can become once it sits inside customer workflows. It sells simulation, testing, and deployment software across autonomy programs, reached an estimated $830M of ARR in 2025, and partnered with NVIDIA to raise simulation fidelity with Cosmos based models.
  6 sacra 7 sacra 8 nvidia
• Siemens represents the incumbent path. Its digital twin and robotics simulation products let engineers test machines in virtual factories, but the workflow is still rooted in explicit models of equipment and control systems. Learned simulators fit beside that stack when behavior depends on partial observability, soft objects, or many interacting agents.
  5 runwayml 9 siemens 10 siemens

The next step is for learned simulation to become a standard pre deployment layer for physical AI teams. If General Intuition ships a broadly available API by late summer 2026, the company can expand from a niche model vendor into infrastructure that game studios, robotics labs, and autonomy teams use every day to train, score, and debug agents before they touch the real world.