Editor’s Note
This article was adapted by EUVC from an original insight published by Synaptic.
Physical AI is moving artificial intelligence from software into machines that can perceive, decide and act in the real world. For investors, the category is difficult to map because it spans foundation models, simulation, sensors, chips, robot hardware, deployment software and end applications.
Synaptic’s central argument is that two contests are unfolding at once: one for intelligence and one for embodiment. Progress in either is insufficient on its own. Commercial systems will depend on models, hardware and deployment economics converging in the same product.
Key takeaways
Physical AI systems follow a sense, think and act loop, connecting perception and reasoning to physical movement.
Capital and attention have concentrated in the cognition layer, particularly vision-language-action models, world models and robot foundation models.
The market includes model-led, hardware-led, specialist and integrated companies, each with different capital needs and sources of defensibility.
Data, simulation, edge compute and observability are enabling layers, not secondary features.
The sim-to-real gap, uncertain unit economics, scarce interaction data and on-device compute remain material constraints.
What is physical AI?
Physical AI refers to systems that combine artificial intelligence with a robot, machine or device so they can perform tasks autonomously in the physical world.
Unlike generative AI, which produces digital outputs such as text or images, physical AI must interpret a changing environment and translate a decision into safe, reliable movement.
The sense, think and act loop
Sense: Cameras, LiDAR, tactile sensors and other inputs collect information about the environment.
Think: Models and planning systems interpret those inputs, reason about the situation and choose an action.
Act: Control systems and hardware execute the decision through arms, hands, wheels, legs or other mechanisms.
The model architectures shaping the cognition layer
Synaptic identifies four approaches that help machines understand and operate in the physical world.
Vision-language models: Combine visual understanding with natural language so a machine can interpret a scene and follow instructions.
Vision-language-action models: Extend visual and language understanding into motor control, translating inputs into physical actions.
World models: Predict how environments and objects may change, supporting planning, physical reasoning and error recovery.
Robot foundation models: An emerging approach that combines several model types into a more general system for manipulation, coordination and autonomous decisions.
Four company archetypes in physical AI
The competitive landscape becomes clearer when companies are grouped by what they build and what they control.
AI-native and model-led companies focus on the intelligence layer, often owning little hardware. Their bet is that model quality and generalisation will provide the strongest moat.
Hardware-led companies prioritise robot bodies, components, dexterity, durability and manufacturing. They face higher capital intensity and longer development cycles but control the embodiment layer.
Specialists pursue a narrow technical breakthrough or use case instead of broad, general-purpose capability.
Full-stack and integrated companies combine models, hardware and deployment infrastructure in an effort to control the data and learning loop end to end.
Large technology companies and established robotics businesses can span several categories. Synaptic highlights NVIDIA as an infrastructure provider across the stack, while companies such as Tesla are pursuing more vertically integrated systems.
Where activity is concentrating across the stack
Foundation models and cognition
Foundation models are one of the most crowded and heavily funded parts of the market. Model-led companies are trying to build intelligence that generalises across tasks, environments and robot forms.
World-model developers are working on spatial reasoning so machines can anticipate how objects and environments behave.
Data and simulation
Real-world robot interaction data is scarce and costly. Synthetic data and simulated environments can expand training without requiring a physical fleet for every experiment, making this layer central to model development and testing.
“Models are only as good as the data they train on, and real-world robot data is scarce and expensive to collect.”
Perception and edge execution
Physical systems must process visual, spatial, audio and tactile signals quickly enough to operate safely. Edge execution reduces latency and dependence on cloud connectivity, but it also creates tight constraints around power, heat and available compute.
Hardware and deployment
Embodied companies are competing on dexterity, reliability, manufacturing scale and cost. Deployment is already taking shape across logistics, manufacturing, autonomous vehicles, healthcare and defence, although maturity varies significantly by system and use case.
For related reading on the wider move from AI demonstrations to real-world industrial deployment, see Industrial AI builders to watch.
Observability and feedback
As fleets move into production, operators need to monitor performance, diagnose failures and feed operational data back into training. Observability tools therefore connect deployment with the next cycle of model improvement.
Capital is concentrating around frontier intelligence
In its April 2026 overview, Synaptic reported that robotics and physical AI attracted approximately $40.7 billion of investment in 2025, 74% more than in 2024 and close to 9% of global venture investment.
Synaptic also cited an increase in funding for world models and vision-language-action models from $1.4 billion in 2024 to $6.9 billion in 2025.
The overview describes capital concentrating in a relatively small group of model and humanoid-robotics companies. It highlights large rounds involving Physical Intelligence, AMI Labs, World Labs, Figure AI, Skild AI, Mind Robotics and Rhoda AI. Some of the figures cited involved reported funding talks rather than completed financings.
Strategic investors were also seeking access to capabilities they may later need to manufacture, deploy or integrate. Synaptic points to venture firms, technology companies and industrial groups investing across the stack, while acquisitions involving robotics technology and business units indicate that incumbents were not waiting for a single winner to emerge.
Synaptic argues that infrastructure providers may capture durable value regardless of which end application wins, although the outcome remains uncertain.
Four constraints still stand between progress and scale
The sim-to-real gap
A system that performs well in simulation can deteriorate in uncontrolled environments with different lighting, surfaces, objects and human behaviour. Better simulation and synthetic data can narrow the gap, but Synaptic considers it unresolved.
Unproven economics
Actuators, sensors and compute remain expensive, while battery life, manufacturing capacity, certification and supply-chain concentration add cost and operational risk. Technical performance does not yet establish attractive unit economics.
A shortage of interaction data
Robots need extensive data from physical interaction. Collecting it is slow and costly, which is why companies are investing in teleoperation, synthetic data and simulation.
Edge-compute limits
Large multimodal models must process vision, force and body-position data with low latency under strict power and thermal limits. Reliance on cloud processing can restrict safety-critical or disconnected use cases.
Physical AI depends on intelligence and execution converging
Physical AI is advancing because models can interpret more complex environments and hardware is becoming more capable. Yet the decisive test is not whether either layer improves independently. It is whether intelligence, embodiment, data and deployment economics can work together reliably outside controlled settings.
“Breakthroughs in intelligence and physical execution must converge for the technology to actually deliver on its promise.”
For investors, Synaptic’s map therefore frames physical AI as a connected system rather than a single robotics category.
The models may attract the largest rounds, but progress also depends on the less visible layers that generate data, execute models at the edge, monitor performance and turn prototypes into repeatable deployments.
Frequently asked questions
What is the difference between generative AI and physical AI?
Generative AI produces digital outputs. Physical AI connects perception and reasoning to a machine that acts in the real world, where reliability, latency and safety become part of the problem.
What is a vision-language-action model?
A vision-language-action model combines visual inputs and language instructions with motor control, allowing a machine to translate what it sees and hears into a physical action.
Why is simulation important for robotics?
Simulation allows models to practise across many environments without collecting every example through a physical robot. The remaining challenge is ensuring that simulated performance transfers to uncontrolled real-world settings.
What are the main barriers to commercial physical AI?
Synaptic identifies four: the sim-to-real gap, uncertain unit economics, scarce physical-interaction data and the limits of low-latency compute on the device.
About the original author
Synaptic is the corporate author of the original article. Synaptic is a private-markets data and intelligence platform that turns company signals, including financial information, company materials, websites, news and alternative metrics, into connected data for investors.
Learn more about Synaptic’s research and tools for private-market investors.
Source
This article was adapted by EUVC from an original article published by Synaptic.
