How large is the physical AI market without double-counting robotics?
Whether you're building or investing in the physical AI market, our report covers everything you need to know: the players, the money, and what's coming next.
SUMMARY
The physical AI market is currently about $12 billion to $20 billion globally once we remove the clearest robotics overlap, while the cleaner intelligence layer is closer to $8 billion to $15 billion and probably around $10 billion to $12 billion at the center of that range.
The biggest published physical-AI figures are mostly ecosystem numbers, not clean software-market numbers. In several reports, robotics or hardware alone already represents roughly half or more of the total.
Fortune Business Insights makes the double-counting problem especially visible: robotics represented 46.9% of its $33.42 billion 2025 physical-AI market, so removing that category alone brings the total down to about $17.7 billion.
Autonomous driving and advanced ADAS are already the largest mature non-robot physical-AI category. A reasonable current range is roughly $4 billion to $7 billion once infotainment, cockpit systems and unrelated connectivity are excluded.
Computer vision, edge AI and digital twins are useful cross-checks, but they cannot be stacked as separate markets. The same perception model can appear in all three categories while representing only one customer purchase.
Simulation, synthetic data and physical-AI development infrastructure are commercially meaningful but still much smaller than the hardware ecosystems built around them. Roughly $1 billion to $2 billion today looks like a defensible standalone contribution.
General-purpose physical-AI models are attracting valuations far ahead of current revenue. Skild AI, Physical Intelligence and FieldAI show how aggressively investors are pricing the possibility that one intelligence layer could eventually control many different machines.
Company revenue supports a low-tens-of-billions market better than either extreme. Mobileye, NVIDIA Automotive and Robotics, Cognex and FieldAI already represent several billion dollars of real commercial activity tied to perception, autonomy and machine intelligence.
Hardware still dominates physical-AI deployment spending today, but the newer value pool sits in reusable intelligence: perception, planning, control, autonomous-driving software, simulation and models that can work across more than one machine type.
The market is probably growing around 20% to 25% annually rather than 30% to 40% across the board. That is still enough for a $12 billion market to become roughly $32 billion in five years at 22% annual growth.
The cleanest way to think about physical AI is therefore as a separate intelligence layer attached to existing machines, not as another label for the entire robotics, automotive, semiconductor or digital-twin industries.
Some countries are moving much faster than others. Our physical AI market report shows where the lead is building.
What should we actually count as physical AI?
The physical AI market should mainly cover the intelligence that lets machines understand and act in the real world, rather than every robot, vehicle, sensor and industrial machine that happens to use AI.
That sounds simple until we look at how the market is currently measured. NVIDIA uses physical AI as a broad umbrella covering robots, autonomous vehicles and smart spaces, with simulation, synthetic data, world models, edge computing and runtime software underneath. IBM also describes physical AI as systems that combine AI with sensors, actuators and control systems to interact with their surroundings.
For market sizing, we need a tighter boundary.
The robot arm, humanoid, warehouse robot, drone or autonomous vehicle is the physical machine. Above that sits the intelligence layer: perception, planning, control, autonomous-driving software, vision-language-action models and edge inference. There is also a development layer around it, including simulation, synthetic data, digital twins and tools used to train and validate these systems.
Why do physical AI market estimates disagree so much?
Physical AI market estimates range from less than $5 billion to more than $30 billion for the same year because researchers are often measuring completely different things.
Fortune Business Insights currently values the 2025 physical AI market at $33.42 billion and puts 2026 at $42.06 billion. Its definition includes hardware, software and services across robotics, autonomous vehicles and machines, vision AI and smart spaces.
Mordor Intelligence estimates a smaller $22.58 billion physical AI ecosystem in 2025, rising to $26.06 billion in 2026. Yet its separate physical AI platforms market is only $8.12 billion in 2025 and $9.71 billion in 2026 because that category focuses more closely on simulation, edge computing, robotics software and intelligent machines.
Grand View Research gets an even smaller number for embodied AI: $4.7 billion in 2025 and $6.5 billion in 2026.
A range from $4.7 billion to $33.4 billion is much too wide to blame on normal forecasting differences. The denominator has changed.
That is why we cannot seriously answer “How large is physical AI?” by picking whichever research report has the newest publication date. We first have to decide how much robotics, hardware and existing automation those numbers already contain.
How much of the reported physical AI market is just robotics?
Almost half of one of the broadest physical AI estimates is already robotics, which is enough to make double counting a major problem rather than a small technical adjustment.
Fortune Business Insights says robotics represented 46.9% of its $33.42 billion physical AI market in 2025. That works out to roughly $15.7 billion.
Remove that one category and the physical AI market drops immediately to about $17.7 billion.
Mordor Intelligence points in the same direction. Industrial robots alone represented 44.59% of its $22.58 billion physical AI ecosystem in 2025, equivalent to about $10.1 billion before we even consider service robots and other robotic systems.
Grand View Research's narrower embodied AI market also remains heavily tied to robotics: hardware accounted for 51.2% of revenue in 2025 and robots were the largest product category at 41.3%.
So when someone quotes a $20 billion, $30 billion or $40 billion physical AI market today, there is a good chance that a large chunk of that number already exists inside the robotics market.
Sometimes the money comes before the interest, sometimes after. We compare both in our physical AI market report.
What is left after we remove robotics from physical AI?
Once the obvious robotics overlap is removed, today's broader physical AI market looks closer to roughly $12 billion to $20 billion than $30 billion to $40 billion.
Fortune gives us the cleanest first adjustment. Its $33.42 billion 2025 market falls to approximately $17.7 billion once the reported robotics share is removed.
Mordor gives us a lower cross-check. Its physical AI ecosystem was worth $22.58 billion in 2025, but hardware represented 71.08% and industrial robots alone made up 44.59%. We cannot mechanically subtract every hardware dollar because autonomous vehicles and edge systems also need hardware, but the figures make clear that the genuinely separate software-and-intelligence market is much smaller than the headline ecosystem number.
Its $8.12 billion physical AI platforms market provides another useful anchor because it already focuses much more closely on simulation, edge computing, software and intelligent-machine infrastructure.
These estimates point us toward two different numbers worth keeping. A broad physical AI market after removing robotics is roughly $12 billion to $20 billion today. A cleaner intelligence layer is smaller, probably somewhere around $8 billion to $15 billion.
| Market definition | 2025 estimate | What is inside it | How useful is it? |
|---|---|---|---|
| Broad physical AI | $33.42B | Robots, AVs, vision, hardware, software, services | Too broad if robotics is counted separately |
| Broad physical AI less reported robotics | ~$17.7B | Robotics category removed | Useful upper anchor |
| Physical AI ecosystem | $22.58B | Very hardware- and robot-heavy | Useful cross-check |
| Physical AI platforms | $8.12B | Simulation, edge, software, intelligent machines | Better intelligence-layer proxy |
| Embodied AI | $4.7B | Robots, autonomous systems, hardware, software | Narrower but still robotics-heavy |
| Our broad non-robot estimate | ~$12B-$20B | Intelligence plus selected autonomous systems | Best answer for the broad market |
| Our clean intelligence-layer estimate | ~$8B-$15B | Models, autonomy, perception, control, simulation | Best answer for incremental physical AI |
If you want more recent data on this point, please see our latest physical AI market report.
Is autonomous driving already the biggest physical AI market outside robotics?
Autonomous-driving and ADAS technology is currently one of the clearest multibillion-dollar physical AI markets outside traditional robotics.
Published autonomous-driving software estimates still disagree considerably. Depending on what functions are included, 2025 estimates range from roughly $2.4 billion to nearly $7 billion.
Company revenue gives us a better reality check.
Mobileye generated $1.89 billion in 2025 revenue, up 15%, and says substantially all of its revenue still comes from commercially deployed ADAS systems. Through the first half of 2026, the business remained around a $2 billion annual revenue run rate, with management raising its full-year revenue outlook to $1.97 billion-$2.02 billion.
NVIDIA's Automotive and Robotics division generated a record $2.3 billion in fiscal 2026, up 39%. The company said growth was driven by adoption of its self-driving platforms, although robotics revenue prevents us from treating the entire $2.3 billion as automotive physical AI.
Qualcomm's automotive business is larger again, but its number includes digital cockpits and connectivity alongside ADAS and autonomous-driving technology.
Taken together, these companies make a sub-$3-billion global autonomous-driving software market look too narrow. A current range of roughly $4 billion to $7 billion for autonomous-driving and advanced ADAS intelligence is more credible once we remove infotainment and unrelated connectivity.
Curious whether the hype matches reality? Our physical AI market report puts the trends next to the deals.
Should normal ADAS count as physical AI?
Modern ADAS should count as physical AI when the system perceives the road and directly changes what the vehicle does.
Automatic emergency braking is a straightforward example. The system detects an obstacle, decides that intervention is needed and applies the brakes. Automated lane keeping sees lane boundaries and changes steering. More advanced hands-free systems combine cameras, radar, maps and neural networks to make continuous driving decisions.
Those functions fit the physical AI definition even when a human still supervises the vehicle.
A digital dashboard or 5G modem does not automatically qualify merely because it sits inside a car.
This distinction becomes important when we use company automotive revenue. Qualcomm's Snapdragon Digital Chassis, for example, covers connectivity, cockpit systems and ADAS. Counting all of that revenue as physical AI would overstate the market.
We use a simple test throughout the calculation: does the AI understand something about the physical environment and then influence an action in that environment? If yes, it belongs much closer to physical AI.
How much computer vision really belongs to physical AI?
Only part of computer vision belongs in physical AI, and counting the whole category would inflate the market by tens of billions of dollars.
Grand View Research estimates the global computer-vision market at $23.6 billion in 2025 and $28.2 billion in 2026. Hardware represented 71.5% of the 2025 total, leaving roughly $6.7 billion of software.
Even that $6.7 billion cannot simply be added to physical AI.
Document OCR, social-media image analysis and many medical-imaging applications use computer vision without controlling a machine in the physical world.
Robot grasping clearly qualifies. Autonomous-driving perception qualifies. A factory camera that detects a bad component and automatically rejects it also fits well because perception changes what the production system does.
A surveillance camera that merely sends someone an alert sits much closer to the boundary.
Computer vision therefore gives us a useful upper bound for physical perception software, but the relevant physical AI portion is only the subset connected to autonomous decisions or physical control.
| Computer-vision application | Count as physical AI? | Reason |
|---|---|---|
| Robot grasping | Yes | Vision directly controls movement |
| Autonomous driving | Yes | Perception feeds vehicle decisions |
| Automated factory rejection | Yes | Detection triggers a physical action |
| Warehouse navigation | Yes | Vision changes machine movement |
| Autonomous agricultural equipment | Yes | Perception controls equipment |
| Safety camera sending alerts | Partly | Understands the environment but may not act |
| Medical-image diagnosis | Usually no | Analysis generally stops before machine action |
| Document OCR | No | Pure information processing |
| Social-media image recognition | No | No physical interaction |
If you want more recent data on this point, please see our latest physical AI market report.
Investors don't value every company the same way. Our physical AI market report explains who gets the premium and why.
How much edge AI should we include in physical AI?
A meaningful part of edge AI belongs to physical AI, but adding the entire edge AI market would count the same software several times.
The global edge AI software market was around $3.7 billion in 2025 and is estimated at roughly $4.5 billion in 2026.
Computer vision accounted for about $1.37 billion of edge AI software in 2025. Manufacturing and industrial customers represented roughly $890 million, automotive and transportation around $567 million, and healthcare and life sciences about $405 million.
Those numbers show that physical environments already account for a large portion of edge AI spending.
The overlap is the problem. An autonomous-driving perception model may simultaneously appear in autonomous-driving software, computer vision and edge AI. Adding all three categories would turn one customer purchase into three separate markets.
We therefore treat edge AI as supporting evidence rather than another full block of revenue. Roughly $2 billion to $3 billion of edge AI software appears closely related to physical-world applications, but much of that is already captured in autonomous driving, industrial perception and machine control.
How much of the digital-twin and simulation market belongs to physical AI?
Physical-AI simulation is probably a low-single-digit-billion-dollar market today, while the much larger digital-twin market contains too many unrelated applications to include wholesale.
The global digital-twin market already exceeds $20 billion by most current estimates. Yet a digital model of a power plant used for maintenance forecasting is quite different from a simulated environment used to train an autonomous machine.
The strongest physical AI applications use simulation to create experience that would be difficult, expensive or dangerous to gather in reality.
NVIDIA's Omniverse and Isaac stack is a good example. Developers can recreate factories or warehouses, simulate machines and sensors, generate synthetic data, train policies and then transfer the resulting behavior to physical systems. Autonomous-vehicle developers do something similar when they generate unusual weather, traffic or crash scenarios in simulation.
Published robot-simulation estimates still range enormously, from hundreds of millions of dollars to several billion, largely because some reports include integration services and large parts of the robotics stack.
For our calculation, roughly $1 billion to $2 billion of current simulation, synthetic-data and development infrastructure looks genuinely attributable to reusable physical AI outside the robot hardware itself.
Deadlines slip more often than headlines suggest. We track them in our physical AI market report.
Are world models and robot foundation models a real market yet?
Physical AI foundation models are becoming commercially real, but their current revenue remains far smaller than their valuations suggest.
Skild AI raised $1.4 billion earlier this year at a valuation above $14 billion while pursuing what it calls an omni-bodied intelligence system: one general-purpose brain intended to control many different robots.
Physical Intelligence has also been discussing funding around an $11 billion valuation while building general-purpose models for robotic manipulation.
A fresher and more useful commercial example comes from FieldAI. The company builds a general-purpose autonomy system that can work across humanoids, quadrupeds, wheeled robots and other machines. Recent reporting says FieldAI has now passed $135 million in revenue and customer contracts across more than 30 customers in construction, energy, data centers, defense and other industrial settings.
That number is still tiny beside Mobileye's roughly $2 billion annual revenue, but it tells us something important: reusable physical-AI intelligence has started generating meaningful contracts outside the major automotive and industrial incumbents.
We should still avoid converting startup valuations into market size. A $10 billion or $14 billion valuation reflects what investors think these companies could become, rather than the revenue they generate today.
Which physical AI products should we exclude to avoid counting robotics twice?
Robot bodies, actuators, conventional robot hardware and robot revenue already captured by the robotics industry should stay outside a separate physical AI estimate.
The same logic applies to much of the supporting hardware. Fortune says hardware represents 55.2% of its physical AI market. Grand View puts hardware at 51.2% of embodied AI. Mordor's physical AI ecosystem is 71.08% hardware.
Those figures explain why the headline markets become so large.
Sensors and processors are obviously necessary for physical AI, but calling every camera, actuator or edge processor a new physical-AI market makes comparisons almost useless. The same hardware dollar may already exist in robotics, automotive electronics, industrial automation or semiconductors.
For the cleanest estimate, we count reusable intelligence, autonomous control, physical-world perception, simulation and supporting software. Specialized hardware can be shown separately when needed.
| Product or revenue stream | Count separately? | Why |
|---|---|---|
| Industrial robot body | No | Already belongs to robotics |
| Humanoid robot | No | Already belongs to robotics |
| Warehouse AMR | No | Already belongs to robotics |
| Robot actuator | No | Robotics component |
| Conventional robot controller | Usually no | Existing robotics stack |
| General-purpose robot foundation model | Yes | Distinct intelligence layer |
| Autonomous-driving software | Yes | Separate physical-AI intelligence |
| Physical-AI simulation tools | Yes | Reusable development infrastructure |
| Synthetic physical-world training data | Yes | AI-specific data layer |
| Machine vision linked to autonomous action | Yes | Core perception function |
| Generic computer vision | No | Includes many non-physical use cases |
| Generic edge AI | No | Large overlap with other categories |
| Entire digital-twin market | No | Much broader than physical AI |
| Robot integration services | Usually no | Already part of robotics deployment |
If you want more recent data on this point, please see our latest physical AI market report.
Want to know which startups could overtake the giants next? It's all in our physical AI market report.
Do actual company revenues support a $10-billion-plus physical AI software market?
Yes, real company revenues support a physical AI intelligence market measured in the low tens of billions, although they do not support the idea that $30 billion or $40 billion of completely new software revenue has suddenly appeared.
Mobileye remains one of the clearest anchors. Its 2025 revenue reached $1.89 billion, and more than 230 million vehicles had been built with Mobileye EyeQ technology by the end of that year.
NVIDIA's Automotive and Robotics business generated $2.3 billion in fiscal 2026.
Cognex produces around $1 billion of annual revenue from machine vision, although some of that revenue belongs to conventional inspection rather than autonomous physical AI.
FieldAI has now reportedly passed $135 million in revenue and contracted business despite being founded only a few years ago.
Those four examples already represent several billion dollars of revenue tied to perception, autonomy or machine intelligence, and they leave out major spending at Qualcomm, Tesla, Waymo, industrial automation suppliers, simulation vendors and private robotics-AI companies.
The revenue evidence fits much better with an $8 billion-$15 billion intelligence layer than with either extreme. A $2 billion market is clearly too small. A separate $40 billion software market is equally hard to reconcile with today's disclosed company revenue.
Is physical AI mostly software today?
Hardware still takes the largest share of physical AI deployment spending today, but software is growing faster and is where much of the genuinely new market is forming.
Fortune puts hardware at 55.2% of physical AI spending. Grand View estimates 51.2% for embodied AI. Mordor's broader ecosystem reaches 71.08%.
Even Mordor's narrower physical AI platforms category remains 45.12% hardware.
So anyone describing today's physical AI economy as mostly AI models and software is getting ahead of the market.
What has changed is the possibility of reusing intelligence across more machines.
A conventional industrial robot usually ships with software designed around that specific system. A general-purpose model could eventually control multiple robot designs. Simulation platforms can serve thousands of machines. Fleet software becomes more valuable as deployments grow. Autonomous-driving software can spread across millions of vehicles.
The International Federation of Robotics now counts 5.079 million industrial robots operating worldwide, up 9% in one year and more than twice the number seven years earlier. Factories installed a record 603,000 new industrial robots in 2025.
Professional service robots are expanding too. IFR says shipments rose 24% to almost 250,000 units in 2025, led by transport and logistics.
That installed base gives physical AI software a very large pool of machines to address. Many older robots lack the sensors, compute or mechanical flexibility needed for sophisticated AI, so the opportunity will develop gradually rather than through a simple software upgrade of five million existing machines.
In many markets, a handful of startups take most of the money. Our physical AI market report shows who's getting the biggest slice.
Where is physical AI already making real money?
Physical AI is already commercially strong in autonomous driving, logistics, industrial vision and other controlled environments where customers can measure whether the system actually saves money.
Cars are already shipping in enormous volumes with AI perception and driver-assistance systems.
Warehouses use autonomous mobile systems because routes, shelves and operating areas can be standardized.
Factories use machine vision and autonomous inspection because defect detection and repetitive processes are relatively easy to measure.
Construction, mining and energy are becoming another interesting cluster. FieldAI's latest commercial progress is useful here because its customers are using autonomous systems for tasks such as site inspection, mapping and operating in environments where conventional pre-programmed robots struggle.
The pattern is fairly clear today. Physical AI works best when the job has economic value, enough structure to limit the number of surprises and a customer that can measure the improvement.
General-purpose household robots face a much harder problem because homes are unpredictable and the value of each automated task is lower.
Is physical AI funding running ahead of the actual market?
Physical AI valuations are far ahead of current physical AI revenue, especially in general-purpose robotics models and humanoids.
Skild AI's $1.4 billion financing valued the company above $14 billion.
Physical Intelligence has discussed a roughly $1 billion round at a valuation above $11 billion.
FieldAI is reportedly raising around $700 million at roughly a $10 billion valuation, up sharply from its previous $2 billion valuation.
Waymo has raised capital on an entirely different scale around autonomous driving, while humanoid companies such as Figure and Apptronik have also attracted very large funding rounds.
Compare those valuations with the current commercial base. FieldAI's more than $135 million in revenue and contracts is impressive for such a young company, yet the reported $10 billion valuation is still dozens of times that figure. Some foundation-model startups disclose even less revenue.
Investors are effectively pricing the possibility that one reusable intelligence platform could eventually run across millions of machines.
That could happen. Today's market revenue does not justify treating that future value as if it already exists.
If you want more recent data on this point, please see our latest physical AI market report.
Which physical AI markets can we safely add together?
We can build a defensible physical AI estimate only by adding purchases that represent different revenue pools and using overlapping markets such as edge AI and computer vision as cross-checks rather than extra blocks.
Autonomous-driving software is one relatively distinct pool.
Non-robot industrial perception and control is another.
Simulation, synthetic data and physical-AI development infrastructure add a smaller but separate category.
World models and general-purpose machine intelligence are now starting to generate commercial contracts, although they remain small.
Smart-space systems and autonomous physical infrastructure add more revenue at the edges.
We should be much more careful with computer vision, edge AI and digital twins because the same software purchase can appear in several of those categories simultaneously.
| Physical AI value pool | Approximate current market | Treatment |
|---|---|---|
| Autonomous-driving and advanced ADAS intelligence | ~$4B-$7B | Major additive component |
| Non-robot industrial perception and control | ~$2B-$4B | Add after removing robotics |
| Simulation and synthetic-data infrastructure | ~$1B-$2B | Add cautiously |
| World models and general-purpose physical-AI models | <$1B-$2B | Emerging commercial market |
| Smart spaces and autonomous infrastructure | ~$1B-$2B | Include selected use cases |
| Incremental physical edge software | ~$1B-$2B | Include only after overlap removal |
| Robotics hardware and robot bodies | Excluded | Already robotics |
| Generic edge AI | Not additive | Cross-check |
| Generic computer vision | Not additive | Cross-check |
| Entire digital-twin market | Not additive | Cross-check |
Those pieces produce roughly $8 billion to $15 billion for the cleaner intelligence layer.
A broader definition that also includes more autonomous-machine hardware, services and borderline physical-AI infrastructure brings us to roughly $12 billion-$20 billion after obvious robotics overlap is removed.
Don't trust a single market forecast. We compare them and build our own view in our physical AI market report.
How fast is physical AI actually growing?
The non-robot physical AI market is probably growing around 20%-25% annually today, fast enough to become several times larger within a few years without relying on the most aggressive forecasts.
The underlying categories grow at different speeds.
Mordor expects physical AI platforms to grow 15.81% annually through 2031.
Its broader physical AI ecosystem forecast is 14.58%.
Fortune's broader market forecast reaches 27.5%.
Grand View expects embodied AI to grow 39.7%, although that market starts from a much smaller base.
Edge AI software forecasts sit closer to the high twenties, while autonomous-driving software generally lands somewhere in the mid-to-high teens.
A single 30% or 40% growth rate for physical AI therefore looks too aggressive for today's whole market. The mature automotive and industrial pieces are already too large for that.
Something around 20%-25% gives us a more realistic blend.
At 22% annual growth, a $12 billion market becomes roughly $32 billion after five years. A $15 billion market reaches roughly $41 billion.
Physical AI can become a very large market surprisingly quickly without pretending that every robot, semiconductor and digital twin represents new physical-AI revenue.
Could the physical AI software layer eventually become bigger than robotics?
Physical AI software could eventually capture as much economic value as large parts of the robotics industry, but that depends on whether the same intelligence can really work across many different machines.
Traditional robotics companies such as FANUC, ABB, Yaskawa and KUKA mainly sell integrated hardware and software.
The newer physical AI companies are aiming for something broader.
Skild AI wants one intelligence system that can control different embodiments. Physical Intelligence is building general-purpose robot models. FieldAI already deploys its autonomy stack across several kinds of machines. NVIDIA is building common simulation, model and compute infrastructure around the whole ecosystem.
If this approach works, the economics change substantially. A model developed once could generate revenue across many robot manufacturers and millions of machines instead of being engineered separately for every platform.
We still do not know how far that portability can go. A humanoid, warehouse vehicle, industrial arm and construction robot have completely different dynamics, sensors and safety requirements.
The software-layer opportunity is therefore large enough to take seriously today, while the idea of a universal physical-AI platform remains something companies still have to prove commercially.
Some businesses here depend on just a few buyers. Find out who in our physical AI market report.
So how large is the physical AI market without double-counting robotics?
The physical AI market is currently about $12 billion to $20 billion globally once we remove the clearest robotics overlap, while the cleaner intelligence layer is closer to $8 billion-$15 billion and probably around $10 billion-$12 billion at the center of that range.
That is our best answer today.
The broad $30 billion-$40 billion figures published for physical AI are useful ecosystem estimates, but they usually include large amounts of robot hardware, autonomous machines, sensors, processors and other spending already captured elsewhere.
Fortune's $33.42 billion 2025 estimate shows the problem especially clearly because 46.9% of the market is explicitly robotics. Removing that portion alone cuts the number to around $17.7 billion.
Our bottom-up calculation lands in the same general area. Autonomous-driving intelligence contributes roughly $4 billion-$7 billion. Non-robot industrial perception and control adds another few billion. Simulation, synthetic data, smart infrastructure, edge software and the emerging general-purpose model layer fill out the rest, with overlaps removed rather than stacked.
The newest commercial evidence also fits that range. Mobileye operates around a $2 billion annual revenue base, NVIDIA's combined Automotive and Robotics division reached $2.3 billion, Cognex remains roughly a billion-dollar machine-vision company, and FieldAI has already reported more than $135 million in revenue and contracts. None of those figures alone defines the market, but together they show that physical AI intelligence is already a serious commercial category.
| Definition | Current market we would use | What it represents |
|---|---|---|
| Broad published physical AI | ~$25B-$40B+ | Usually includes large amounts of robotics and hardware |
| Broad physical AI excluding robotics overlap | ~$12B-$20B | Best broad answer |
| Clean physical-AI intelligence layer | ~$8B-$15B | Autonomy, perception, models, control, simulation |
| Central intelligence-layer estimate | ~$10B-$12B | Most useful current working number |
| Autonomous-driving intelligence | ~$4B-$7B | Largest mature non-robot category |
| General-purpose physical-AI models | Still relatively small | Rapid commercial growth from a small base |
| Robotics market | Separate | Robot bodies, systems and robotics deployment revenue |
So the clean physical AI market is already in the low tens of billions of dollars, yet it remains much smaller than the industries it is being attached to.
The part worth watching now is the reusable intelligence layer. Cars, factory equipment and robots already exist at massive scale. The newer market comes from software that can give those machines better perception, reasoning and autonomy without rebuilding the entire machine around every new task.
That is where physical AI starts becoming a genuinely separate market rather than another label for robotics.
If you want more recent data on this point, please see our latest physical AI market report.
OUR METHODOLOGY
The central difficulty in sizing physical AI is not finding a market forecast. It is deciding what the market actually contains. “Physical AI” is still used to describe several overlapping things: robotics, autonomous vehicles, machine vision, edge AI, simulation, smart spaces, embodied AI and the emerging intelligence layer that sits across these systems.
We broke the market into the dimensions that actually change the answer: what qualifies as physical AI, how much reported spending is already robotics, which non-robot markets represent separate revenue, where categories overlap, what companies are already generating in real revenue, and how quickly the underlying commercial markets are expanding.
The key rule was to avoid stacking the same economic activity more than once. A perception system used inside an autonomous vehicle can appear in autonomous-driving, computer-vision and edge-AI estimates at the same time, so broad adjacent markets were used mainly as cross-checks unless we could isolate a distinct revenue pool.
Robot bodies, conventional robotics hardware, actuators and deployment revenue were separated from the incremental physical-AI layer. We focused instead on reusable intelligence, autonomous control, physical-world perception, simulation, synthetic data and supporting software, while showing specialized hardware separately where it helped explain the ecosystem.
We also used disclosed company revenue and deployment data as a reality check. These figures were not simply added together as though they represented the whole market; they were used to test whether the scale implied by the top-down estimates was commercially plausible.
Published market studies were most useful when they disclosed what was inside the number. Fortune Business Insights, Mordor Intelligence and Grand View Research provided the main composition checks because they split physical AI or embodied AI by robotics, hardware, software or product category rather than giving only a headline total.
Key sources used for this analysis include: NVIDIA on the definition and scope of physical AI, IBM on physical AI systems, NVIDIA on physical-AI tools across robotics, autonomous vehicles, vision AI and digital twins, NVIDIA on physical-AI data and synthetic-data infrastructure, NVIDIA Isaac Sim, and NVIDIA's autonomous-driving platform.
For market composition and company-level checks, we used Fortune Business Insights on physical AI market size and robotics share, Mordor Intelligence on the physical AI ecosystem, Grand View Research on embodied AI, Mobileye's 2025 results, Mobileye's 2026 guidance update, NVIDIA's fiscal 2026 results, Qualcomm on automotive scale, Qualcomm on the boundaries of Snapdragon Digital Chassis, and Qualcomm's fiscal 2026 automotive momentum.
We used the International Federation of Robotics as the installed-base check: IFR on the global industrial robot installed base and 2025 installations and IFR on professional service-robot shipments. The final ranges come from triangulating adjusted top-down estimates, bottom-up value pools, disclosed company revenue and the scale of machines already deployed, rather than treating any single forecast as definitive.
Some startups are growing their valuation much faster than others. Our physical AI market report shows who's pulling ahead.
