The most consequential question in AI right now is not which frontier lab produces the smartest chatbot. It is who controls the data that teaches machines how the physical world behaves. Text-trained models gave us language intelligence. World models will give us physical intelligence, and the input that matters changes completely: from sentences to sensor streams, from web crawls to factory floors, from scraped documents to flight telemetry and satellite imagery. In our view, this transition constitutes the next platform shift in AI, one where the moat moves away from internet-scale text corpora and toward something far harder to replicate.
From Language to Physics
A language model learns by predicting the next token in a sequence of text. The output is fluency. A world model learns by predicting what happens next in video and simulation: how objects move, how forces propagate, how cause precedes effect in space and time. The output is physical intuition.
That distinction matters because physical intuition is the missing ingredient for the applications that actually transform industrial economies. Robotics, autonomous vehicles, and automated manufacturing are all bottlenecked not by compute or algorithmic cleverness, but by the absence of models that understand what happens when a robot arm exerts torque on an unfamiliar object, or when an aircraft enters an edge-case aerodynamic regime. Text cannot teach that. Physics must.
The field has converged rapidly. A leading chip and AI infrastructure provider released a world foundation model in June 2026 specifically designed for physical AI, unifying synthetic world generation, vision reasoning, and action simulation in a single architecture.1 Separately, a major UK AI research lab has produced interactive environment generators capable of producing photorealistic physics-grounded worlds from text prompts.2 A European AI startup raised over $1 billion to pursue a fundamentally different model architecture built around physical world prediction rather than language modeling.3 This is not incremental progress on a single research front. It is convergence from multiple directions.
The Moat Moves
In the language AI era, the moat was access to text at scale: internet crawls, digitized books, licensed content libraries. That data, while unequally distributed, was at least theoretically accessible. Physical-world data is structurally different. The aerospace sector alone is generating an unstructured flood of flight logs, satellite imagery, maintenance records, and supply chain data that no public index captures.
Artificial intelligence is compounding the value of aerospace and defense telemetry by converting high-volume sensor streams into faster diagnostics, anomaly detection, and predictive insights. But the deeper point is not what AI can do with that data today. It is that the data itself becomes a training asset. An operator sitting on years of flight telemetry across varied aircraft types, weather envelopes, and failure modes holds a dataset that no amount of funding can quickly replicate. The same logic applies to factory floor sensor data from a precision manufacturer, or to satellite imagery collected by a constellation operating over years at global scale.
The space-based synthetic data market for AI training expanded from $2.08 billion in 2025 to an anticipated $2.71 billion in 2026, a 30.7% jump attributed to increased satellite deployments, rising demand for AI training data, and adoption of earth observation technologies. That market exists because satellite imagery is now recognized as a training input, not just an analytical output. The asset class of proprietary physical-world data is forming in real time.
The global synthetic data for physical AI market was valued at $2.03 billion in 2025 and is projected to reach $63.95 billion by 2035, growing at a 41.25% CAGR. The direction and magnitude of that trajectory reflect a structural recognition: real-world data collection, even at scale, cannot keep pace with the training requirements of world models. Synthetic data generated by high-fidelity simulation fills the gap, which is precisely why simulation itself is becoming a product category, not just a research tool.
Simulation as Product
The economic logic here is underappreciated. In previous technology cycles, simulation was a cost center: you ran simulations to avoid physical testing, and the value was measured in avoided expense. World model simulation is something categorically different. World model simulators are becoming infrastructure for physical AI, and the market will grow where teams can move from static scenarios to interactive worlds that support training, testing, and policy improvement.
The global world model simulators market is valued at approximately $5.25 billion in 2026, with robotics and physical AI capturing 38.5% of total revenue, and is forecast to reach $48.9 billion by 2036 at a 25.1% CAGR. That growth is not a rounding effect of general AI enthusiasm. It reflects a specific structural reality: an agent that can simulate its world can plan, learn from imagined experience, and handle situations missing from its training data, which means that simulation quality directly determines model capability in deployment. Better simulation environments do not just reduce testing cost. They produce better-performing autonomous systems, full stop.
The investment community has caught up to this logic at the company level. In June 2026, a leading physics AI company announced a $300 million Series C at a $2.4 billion valuation, led by a major sovereign wealth fund with participation from a chip-and-AI infrastructure giant and two industrial conglomerates, cementing its position as the highest-valued dedicated physics AI company globally. A separate physics AI simulation company secured a $135 million Series B, with strategic investors drawn from industrial manufacturing and semiconductor fabrication. Both rounds signal that institutional capital is now treating simulation software as infrastructure, not tooling.
Aerospace at the Intersection
No sector sits more squarely at the center of this thesis than aerospace. It is simultaneously among the most data-rich industries on earth and among the most demanding consumers of physical AI. The aerospace and defense telemetry market, valued at $2.75 billion in 2026, is projected to reach $4.94 billion by 2032 at a 9.9% CAGR, and that figure captures only the telemetry infrastructure layer, not the compounding value of the data it generates as a training asset for world models.
The applications run in both directions. Flight telemetry from decades of commercial and defense operations is exactly the kind of time-series, physics-grounded data that world models require to learn aerodynamic behavior, system degradation patterns, and edge-case failure modes. At the same time, world model simulation is precisely what aerospace engineering teams need to validate new vehicle designs, train autonomous systems for GPS-denied environments, and compress certification timelines that currently require years of physical flight testing.
By simulating rare and dangerous edge cases, such as unexpected aerodynamic conditions, world models allow autonomous systems to be trained on scenarios that would be nearly impossible or too risky to replicate in the physical world. For aerospace, that is not a convenience. It is a qualification pathway.
What Determines Who Wins
The platform shift from text to physical-world data creates a new set of structural advantages, and they are not the ones that defined the previous cycle.
First, proprietary sensor infrastructure. Organizations with instrumented fleets, factory floors, or satellite constellations accumulate training data as a byproduct of operations. That flywheel does not exist for new entrants without existing physical deployments. Software is projected to grow at a 40.43% CAGR through 2031 in the physical AI market, with the shift tied to world foundation models, simulation frameworks, and fleet management tools that improve performance across deployed machines without replacing physical assets. The incumbents with large installed bases accumulate the data to fine-tune those models. The advantage compounds.
Second, simulation fidelity. Commercial platforms have demonstrated production-grade accuracy matching within 1 to 2% of high-fidelity reference simulations for computational fluid dynamics and structural analysis problems, but reaching that fidelity threshold requires sustained investment in physics-based modeling that pure software companies cannot shortcut. The simulation layer is not commoditizing quickly.
Third, the integration between real and synthetic data. Frameworks are emerging that use world models explicitly as low-cost, controllable virtual environments for training, substantially reducing dependence on real interaction while improving data efficiency and robustness. The teams that can close the real-to-simulation gap fastest will train models that generalize across physical environments, not just simulated ones.
Investment in Vision-Language-Action models reached $3.8 billion in 2025, nearly three times the 2023 level, while leading models now support zero-shot learning, indicating that the capability frontier is advancing faster than the data infrastructure supporting it. That gap between model ambition and data reality is where the structural competition is being decided.
Where This Lands
In our view, the transition to world models is not a continuation of the language AI buildout. It is a distinct platform shift with a different set of winners, different capital requirements, and a different answer to the question of where durable competitive advantage forms.
The companies best positioned are those that simultaneously generate proprietary physical-world data and build or access the simulation infrastructure to turn it into training signal. Aerospace, advanced manufacturing, and autonomous systems are the primary domains where those conditions overlap, and that overlap is not accidental. These are the sectors where the gap between what current AI can do and what physics demands is largest, and therefore where the prize for solving it is greatest.
What would change this view: a rapid commoditization of world model simulation platforms that eliminates the fidelity advantage of established physics AI software. Early evidence, specifically the concentration of funding into high-fidelity simulation companies rather than general-purpose world model infrastructure, argues against that outcome in the near term. The moat is real. The race to occupy it is underway.