Physical AI · Supply-chain maps · AS OF 2026-09-02
Physical AI
The supply chains behind robotics and autonomy
AI is making more physical tasks practical to automate. The opportunity reaches far beyond humanoids—and depends on much more than the intelligence inside a machine.
Physical AI is intelligence acting through machines: robots handling goods, vehicles navigating roads, equipment responding to crops and terrain, and aircraft inspecting or delivering. Our view is that these systems will become more prevalent as perception and control become capable, reliable and economical enough for more tasks.
The opportunity is not confined to machines that replace an entire human job. A robot can remove walking while a person still picks the goods. A sprayer can recognize weeds while a person still drives. These narrower improvements can create useful markets before general-purpose autonomy is ready.
This report follows what enables that change: components, software, data, production, integration and operating services. A customer using a robot is evidence of demand; it is not automatically a supplier node. Conversely, a distinct enabling activity can belong in the chain while it is still performed inside the machine producer.
What is changing
Less of the world has to be specified in advance
Automation already works well in many stable, repeatable settings. The opening for AI is variability: recognizing a differently placed pallet, finding a part in a changing scene, or adapting movement to an unfamiliar situation. Better perception, learned policies and development tools can reduce the engineering needed for each new task. They do not eliminate it.
The family evidence spans different stages. Warehouse robots already support repeated operating workflows. Selective spraying changes a working tool without making the whole vehicle driverless. Humanoid deployments test whether more adaptable manipulation can become dependable production work. These should not be counted as one adoption curve.
The machine base matters too. AI can draw on existing motors, sensors, controls, manufacturing and service networks. As a capability becomes useful, the constraint may move toward repeatable production, installation, reliability or support. That is why the maps continue past the model and the computer.
What is easy to misread
A capable demonstration is not a working business
A successful task shows what a machine can do under the demonstrated conditions. Adoption requires repeated useful work; the economic case also includes integration, supervision, charging, maintenance and failures. The possibility of working around the clock is valuable, but useful operating time depends on all of those inputs—and on enough work to do.
Nor does shared technology make the machines interchangeable. A hand actuator, a truck’s steering interface and an aircraft’s flight controller enable physical action in different ways. The same supplier can participate in several chains without turning its entire company into a shared node.
Our view
The likely expansion is broad, but uneven
We expect AI to make a widening set of physical tasks worth automating. Some will be performed by familiar machines with better capabilities; others may require new bodies and supply chains. General-purpose robots are one possible route, not the condition on which the whole theme depends.
The important uncertainty is how much learning and integration can be reused. If every deployment still needs extensive bespoke engineering and human intervention, the economics may remain narrow. If useful capabilities become easier to transfer and maintain, demand can broaden for the inputs and services that support them. Neither outcome tells us in advance which suppliers will earn durable margins.
Across the family chains
Where their growth meets
Different machines can create demand for the same enabling businesses. A supplier of computing, sensing or precision motion may serve several robot markets even when the machines themselves have little in common. This is the main investment question behind the family maps: which inputs become more valuable as more kinds of machines become useful?
Demand can grow through more machines, more content in each machine, more development work, or a larger installed base to support. The strongest opportunities need more than recurrence: the supplier must provide something difficult to reproduce or replace, and retain value as production scales. The sections below distinguish shared products and capabilities from needs that still require different solutions.
The companies are selected examples of the relevant work, not a market-share ranking. A product offering, design win and delivered volume are different evidence.
More adaptable machines need experience they can learn from.
A grasping robot needs examples of objects, actions and outcomes; a driving system needs road scenes and difficult situations. Collecting, aligning and evaluating those records creates work before a model becomes useful. If more tasks can be learned, demand can spread across data collection, training tools and policy development rather than remaining inside each machine’s programming team.
Some of that work can be reused. Google DeepMind’s October 2023 Open X-Embodiment research pooled experience from different robot bodies and demonstrated transfer in the tested systems. It does not make a manipulation dataset interchangeable with driving data. The overlap is in collection infrastructure, representations and learning methods; valuable task-specific experience can remain with the operator or robot builder.
The economic question is who can supply useful new experience or a reliably adaptable skill. Open datasets and model weights lower entry costs, while collection quality, difficult cases and access to operating data can remain differentiated. A larger dataset is valuable only if it improves performance on the intended task.
Relevant businesses
Scale AI (private) — robot demonstrations, collection operations and data preparation for developers.
Google DeepMind / Alphabet — cross-embodiment learning research and models; the cited open corpus is not a paid data product.
Siemens — task-specific Robot Pick AI software, illustrating a commercial capability narrower than a general robot model.
Learning and testing more behaviours creates work before another machine is built.
Simulation lets developers rehearse situations that are slow, costly or dangerous to reproduce physically. A digital twin models the relevant machine, cell or environment so engineering can be tested against its intended setting. These tools serve policy training, sensor testing, robot programming and commissioning—not just attractive virtual demonstrations.
Rendering, physics engines and development tools can support several families, while the models of contact, roads, terrain and flight remain different. Applied Intuition’s partnerships with TRATON in March 2025 and Komatsu in September 2025 illustrate a software supplier crossing road and mining applications. They establish commercial relationships, not that the same scenarios or approval evidence work in both.
The difficult part is making simulated results predict the real task. More testing is not useful if sensor errors, friction or unusual events are represented poorly. Revenue can accrue through engineering software, support and computing: Isaac Sim is free and open source, while cloud resources and some associated distribution rights are paid separately.
Relevant businesses
NVIDIA — robot simulation and synthetic-data tools, alongside the computing used to run them.
Siemens — Process Simulate for robot programming, manufacturing simulation and commissioning work.
Applied Intuition (private) — autonomy development, simulation and validation tools.
Development workloads and deployed machines create two sources of computing demand.
Training, simulation and evaluation use computing before a robot reaches a customer. On the machine, processors and memory run perception and decisions within its latency, power and cooling limits. Development demand depends on the work needed to build and improve capabilities; onboard demand depends on machine volume and computing content. Neither grows mechanically with the other.
This is one of the clearest places for different robot chains to share an upstream supplier. NVIDIA’s August 2025 Jetson Thor launch named early adopters spanning logistics, humanoids, heavy equipment and medical technology. Qualcomm offers another route. Shared processors and development ecosystems can coexist with different boards, interfaces and qualification requirements.
Replacing a compute module can mean adapting software and retesting the machine. NVIDIA’s published module lifecycles illustrate why continued availability matters alongside processing speed. That can make an established platform valuable without establishing a shortage. Memory also differs: low-power onboard memory is not the HBM beside a data-centre accelerator. The training-infrastructure chain is traced in AI — The supply chain behind AI compute; this research does not establish how much of its demand or scarcity is attributable to robots.
Relevant businesses
NVIDIA — accelerated development computing and Jetson onboard modules.
Qualcomm — Dragonwing industrial and robotics computing platforms.
Micron — memory and storage for edge computing, distinct from its data-centre memory products.
Machines working in less predictable settings need better information about the world and themselves.
Cameras, lidar and radar observe surroundings. Encoders, inertial sensors and force or tactile sensors report position, movement and contact. AI can make these signals useful for more tasks, expanding demand beyond simply counting the sensors on today’s robots. The opportunity can be in better sensing, calibration or processing rather than always adding more devices.
Hesai provides a concrete cross-family example. Its March 2026 results reported 239,273 robotics lidar shipments in 2025, up from 45,503 in 2024, and named orders spanning robotaxis, delivery vehicles, humanoids/quadrupeds and lawn mowers. Those are sensor shipments and order examples—not a count of robots or evidence that every architecture uses lidar.
The common market breaks into meaningful specialties. Road sensing needs range and environmental robustness; manipulation needs information about close contact; a mobile robot’s navigation sensor may not provide its protective safety function. Calibration, reliability and qualification can be harder to reproduce than a specification sheet. Competition, lower component costs and software that works with fewer sensors can offset rising volumes.
Relevant businesses
Hesai — lidar products supplied into several autonomous-machine markets.
SICK (private) — industrial sensing, localization and protective sensing.
Novanta — force/torque sensing for contact-sensitive robotic work.
A learned decision still has to become a timely, controlled physical action.
Runtime software moves information between sensors, models and hardware. Controllers execute motion and manage faults. These functions remain necessary whether a machine uses a large learned policy, a small perception model or mostly conventional automation. Greater capability can increase the work of coordinating them rather than replacing them with one AI processor.
There is actual product overlap beneath the different machines. Infineon’s August 2025 robotics announcement pairs motor-control microcontrollers with Jetson Thor. QNX offers pre-certified embedded software under several sector-specific standards. Common components and support can be reused, while system hazards, stopping behaviour and the finished machine’s validation remain local.
The potential advantage is dependable execution, supported interfaces and qualification work customers would otherwise repeat. Switching can require redesign and retesting. That is different from a shortage of processors, and it does not guarantee high margins: open middleware such as ROS 2 and in-house implementations remain alternatives for parts of the work.
Relevant businesses
Infineon — motor-control processors, power electronics and feedback sensing.
Texas Instruments — processing, control and supporting electronics for robotic systems.
QNX / BlackBerry — embedded operating systems and safety-qualified software components.
Motors, transmissions, bearings, drives and joint modules turn electrical commands into movement. A humanoid may combine many tightly packed joints; an industrial arm needs repeated precision; a mobile base needs efficient wheel drives. Supplier know-how can cross these applications even when torque, speed, shock tolerance and cost produce different components.
The important split is between making individual motion parts and delivering a tested module whose parts work together. Miniaturization, manufacturing consistency and long service life can make that work valuable. But aircraft propulsion, heavy-machine hydraulics and small finger mechanisms do not become one interchangeable actuator market merely because all produce motion.
Supply can improve as demand grows. In August 2026, Schaeffler described a forming process for strain-wave gearboxes, with series production planned for 2027. Its aim is to reduce machining time and cost. This is a specific potential constraint-clearing mechanism; current qualified capacity, yields and customer orders still matter more than assuming that precision parts must remain scarce.
Useful operating time depends on supplying energy without excessive weight, heat or downtime.
Mobile robots and electric aircraft need batteries, protection, charging and power conversion. Humanoid designs must share a limited energy and cooling budget between computing and motion. Fixed robots can use mains power, while many field machines retain other energy sources. Shared demand is therefore broader than batteries, and battery demand is not universal.
Cells can come from much larger electronics or vehicle markets. The differentiated work may lie in a qualified pack, battery management, compact power electronics or reliable automatic charging. Texas Instruments describes monitoring, protection and balancing for mobile-robot packs; EnerSys supplies wireless AGV charging. Repeated use also creates replacement and service demand.
Skydio’s October 2024 battery-supply disruption shows how a specific sourcing dependency can constrain a machine supplier. It does not establish today’s industry-wide battery shortage. Better efficiency, automated charging, alternative suppliers and suitable pack designs can all expand useful operating time without a battery-chemistry breakthrough.
Relevant businesses
Texas Instruments — battery-management and power-control electronics.
Infineon — power conversion, motor-drive and battery-management components.
EnerSys — industrial batteries and charging systems, including automated charging.
A growing installed base needs repeatable manufacturing, integration and continuing support.
Manufacture turns a design into consistent machines. Deployment joins them to the customer’s task, facilities and software. Service keeps them useful. These activities create demand even when the core AI is developed internally—and can determine whether cheaper, more capable hardware leads to wider adoption.
There is cross-family supplier participation here too. Magna appears in both Serve’s 2024 delivery-robot manufacturing agreement and Waymo’s May 2025 vehicle-integration account. This establishes participation in two chains; how much manufacturing capability transfers between them is a further question. It does not establish a common production line or identical customer economics. Factory controls and logistics software offer other routes to repeatable deployment.
Overlap becomes weaker in local work: commissioning a welding cell, maintaining a mine fleet and operating a drone service require different people and obligations. The attractive question is which tools, interfaces and production methods reduce repeated engineering across customers. Revenue that still requires proportional additions of field labour has a different expansion path from reusable software or standardized components.
Relevant businesses
Magna — contract manufacturing and integration; the cited customers remain separate programs.
Siemens — industrial engineering, controls and commissioning tools.
Rockwell Automation — factory automation and OTTO autonomous material-handling systems.
The family pages are detailed examples, not the limits of the theme. Novanta’s surgical-robot force sensing illustrates how an enabling business can also serve machines not separately mapped here. That extends our view of a supplier’s demand; it does not substitute for research into medical robots’ own production, adoption and operating requirements.
Across these markets, the promising pattern is broader demand meeting a capability that takes time to reproduce. The evidence above establishes several overlaps, not one universal Physical AI bottleneck. The family chains below let us trace where each input is used, what can replace it, and which downstream conditions must hold for that demand to arrive.
Six provisional families · seven route maps
Follow the machine’s supply chain
These complete chains are the foundation for the cross-family reading above. They group machines by their enabling chains, not every application or body size. Mobile robots have two maps—managed-site transport and public-facing service—because their deployment and operating work differ. Shared inputs repeat so each map is understandable on its own.
Family 1
Industrial manipulators & robot cells
Industrial arms and cobots become robot cells only after task tooling, guarding, controls and workflow engineering adapt them to a production application. The arm can cross the market as equipment, or an integrator can carry application risk through acceptance of a commissioned cell.
AMRs, guided vehicles and autonomous lift trucks move goods through warehouses and factories. The useful unit is a working material flow: robots, loads, charging, fleet software and the connections to people and other machines.
Robots move goods or perform services around pedestrians, customers and visitors. Street delivery adds route operations and remote assistance; indoor services can add doors, lifts and building-system integration. The machine and its operating service both matter.
Robots that move through a workplace and handle objects with arms or hands. Their promise is to use spaces and tools already built for people; their distinctive supply chain combines many joints, whole-body control, dexterity and task learning.
Systems that drive passenger vehicles or freight trucks on public roads. A driving system must be integrated with a suitable vehicle, validated for its intended operation and supported by transport operations. Robotaxis and highway freight share inputs but have different service economics.
Agricultural, mining and construction machines use AI and automation to drive, handle material or control a working tool. Rugged hardware and machine interfaces recur, but autonomous haulage, tractor fieldwork and precision implements have different supply and deployment routes.
Autonomous aircraft carry sensors, parcels or other task payloads. Flight control, energy and airspace awareness meet a mission-specific body; launch infrastructure, supervision and data processing determine whether the customer receives useful work rather than just a flight.
These references explain the difference between selling equipment and selling work, and how the current scope was chosen.
Economics
Commercial routes
The same technical system creates different economics depending on who owns the asset, integrates it, performs the work and carries obligations after delivery.
The theme and family reports were reviewed on 2 September 2026. Their source links distinguish operating evidence, product descriptions and future plans. The supporting references retain their own dates of record. The maps are provisional and non-exhaustive: an absent activity or family is not thereby excluded from Physical AI.
A node is a product, service or value-adding activity, not a company proxy. The reports remain a guide to understanding, not a market-size forecast or a place to accumulate news. A timeline is useful only when it explains an evidenced change in demand, adoption or a bottleneck; background history can remain in the prose.