World Models Market Outlook 2026–2036: Market Size, Growth Drivers, Key Players, Strategic Developments & Adoption Forecast for Physical AI, Robotics and Autonomous Systems — A Meticulous Next™ Foresight Brief

This Meticulous Next™ brief examines how world models i.e. AI systems that learn an internal representation of how the physical world behaves and use it to predict, plan and act, will move artificial intelligence from screens into machines over the next 5–15 years. Language models describe the world. World models simulate it. That difference is what allows a robot, a vehicle or a factory to act in situations it has never encountered. The brief maps the technology, its indicative market size and forecast, the factors behind its growth, the developments of the last 24 months, the key players operating in the space, and the adoption trajectory to 2036.

It is a focused 32-page decision brief for robotics and automation OEMs, automotive and mobility companies, industrial operators, semiconductor and platform vendors, and investors who need to separate the physical-AI applications that will scale from those that remain demonstrations. It presents an indicative trajectory rather than a segmented market model. Its purpose is to identify which applications world models unlock first, where the value chain forms, and who captures it.

A world model is an AI system trained on video, sensor, simulation and text data to learn how objects, forces and agents behave over time. Given a state of the world and a proposed action, it predicts what happens next. This lets a machine plan by imagining outcomes before acting, and lets developers train control policies inside a learned simulator instead of on physical hardware. The approach differs from conventional simulation, which is hand-built from physics equations, and from vision-language models, which recognize and describe scenes but do not model dynamics.

Three uses are emerging. First, world models generate synthetic physical data at scale, closing the data gap that has limited robot learning. Second, they act as training environments in which policies for grasping, navigation and driving are learned and stress-tested. Third, they run inside the machine as a planning component, allowing it to reason about unfamiliar situations. The World Economic Forum's Top 10 Emerging Technologies 2026 identifies this shift from pattern recognition to physical reasoning as the reason world models matter now, with NVIDIA's Cosmos platform cited as the example: it trains robots on physical-world data so that they adapt to new environments using an internal model of how things work.

The category sits at the intersection of three industries. Semiconductor and platform vendors supply the training compute and world foundation models. Robot, vehicle and equipment OEMs supply physical data and embed the resulting policies. Simulation and synthetic-data companies supply the tooling that connects the two. Deloitte's Tech Trends 2026 describes the result as intelligence that is embodied and autonomous rather than confined to screens, citing Amazon's fleet of a million robots coordinated by an AI layer and BMW factories in which cars drive themselves through production routes.

The world-models market, including world foundation model platforms, simulation and synthetic-data tooling, and physical-AI software licensed into robots, vehicles and industrial systems, is estimated at USD 1.5–2.5 billion in 2026. Meticulous Next™ expects it to reach USD 40–60 billion by 2036, an indicative CAGR of 35–40%. Growth is led by two applications with existing budgets: autonomous-driving development and warehouse and factory robotics. General-purpose and humanoid robots add a second wave from the early 2030s once world-model-trained control proves reliable outside structured environments. The market is software and platform revenue; the much larger hardware markets it enables — robots, vehicles, automation equipment — are addressed in adjacent Meticulous reports. North America leads on platform supply; East Asia leads on deployment volume; Europe leads on industrial and automotive integration.

The base case assumes steady gains in long-horizon prediction accuracy and continued decline in training compute cost. An accelerated case adds a breakthrough in sample-efficient physical learning and rapid humanoid deployment, pulling the inflection to ~2029 and the 2036 value to the top of the range. A delayed case assumes simulation-to-reality transfer remains unreliable outside structured settings, or safety incidents slow autonomous deployment, pushing the inflection to ~2034.

Growth drivers

• Data scarcity in robotics: physical demonstration data is expensive and slow to collect; world models generate it at scale.
• Labour shortages in logistics, manufacturing and care create demand for robots that handle variation rather than fixed tasks.
• Autonomous-driving programmes need to test rare scenarios that cannot be driven safely or often enough in the real world.
• Compute and platform vendors are investing to make physical AI the next growth market after language models.

Enablers

• World foundation model platforms available as developer tooling rather than research code.
• Falling cost of GPU compute and growth of video and sensor datasets from fleets and facilities.
• Digital-twin infrastructure in factories and cities that supplies structured physical data.
• Robotics foundation models that turn world-model outputs into executable control policies.

Restraints and barriers

• Simulation-to-reality gap: policies trained in learned simulators can fail on contact, deformable objects and edge cases.
• Long-horizon prediction remains error-prone; models compound mistakes over extended sequences.
• Safety validation and certification frameworks for learned physical reasoning do not yet exist.
• Concentration of training compute and models in a few vendors raises dependency and cost concerns for OEMs.

Five converging forces will determine how fast, and how far, world models reshape physical AI:

• Accuracy of physical prediction over long horizons
• The supply of physical data from fleets, facilities and simulation
• The economics of training compute
• Safety validation and certification of learned control
• The balance of power between platform vendors and the OEMs that own physical data.

The brief assesses each force for direction, speed and confidence.

The key players operating in world models and physical AI include NVIDIA Corporation, Alphabet Inc. (Google DeepMind), Meta Platforms Inc., Microsoft Corporation, Amazon.com Inc., Tesla Inc., World Labs, Wayve Technologies Ltd., Physical Intelligence, Figure AI Inc., Skild AI, 1X Technologies AS, Agility Robotics Inc., Boston Dynamics Inc., Covariant (Amazon), Toyota Research Institute, Waymo LLC, Applied Intuition Inc., Foretellix Ltd., Niantic Spatial Inc., Runway AI Inc., Decart, Odyssey and Unitree Robotics. The brief profiles representative players in each archetype and assesses which are positioned to own the world-model layer.

The competitive landscape is forming around five archetypes. Compute and platform vendors supply world foundation models bundled with training infrastructure. Frontier AI labs develop general world models as a path toward physical reasoning. Robotics foundation-model and humanoid developers build the policies and machines that consume world models. Automotive and mobility developers apply world models to driving and scenario generation. Simulation and synthetic-data specialists supply the tooling and validation layer. Competitive intensity is moderate in 2026 and is expected to rise as OEMs decide whether to license platform models or build proprietary ones on their own data.

Where value migrates:

In 2026 value sits in training compute and world foundation model platforms. By 2031 it moves to the control policies and machines that world models make possible, and to the validation tooling that certifies them. By 2036 it settles in proprietary physical data — the fleets, facilities and deployments that keep a world model current — and the platform vendors that capture that data as a service. OEMs that license models without retaining their data risk becoming hardware suppliers to someone else's intelligence.

Capital is concentrating in robotics foundation-model and humanoid companies, which raised large growth rounds through 2025–2026, and in the platform layer, where compute vendors are investing to make physical AI the next growth market [add named rounds]. Patent activity is concentrated in video prediction, simulation-to-reality transfer, sensor fusion and policy learning. The brief tracks three indicators: number of commercial robot and vehicle programmes trained primarily in learned simulators, published benchmarks for long-horizon physical prediction, and the share of OEMs building proprietary world models versus licensing.

North America leads on platform supply because compute vendors, frontier labs and autonomous-driving developers are concentrated there. East Asia leads on deployment volume, with China's humanoid and industrial-robot programmes and Japan and South Korea's factory automation providing the physical data flywheel. Europe leads on industrial and automotive integration and on certification frameworks under the AI Act and Machinery Regulation, which makes it the proving ground for validated physical AI.