Where is the money in the physical AI market?

Why is it so hard to tell where the money is in physical AI?

Physical AI is already making serious money today, but mostly in established automation markets rather than the humanoid robots attracting the most attention.

The difficulty starts with the label itself. “Physical AI” now covers everything from industrial robots and warehouse systems to autonomous vehicles, surgical robots, drones and general-purpose robot brains. These businesses share AI that perceives the real world and acts on it, but commercially they are at completely different stages.

The latest World Robotics data from the International Federation of Robotics makes that gap unusually clear. Factories installed 603,000 industrial robots in one year, taking the global installed base above five million. Professional service robot shipments reached almost 250,000 units, including 117,500 used in transportation and logistics.

Full-size humanoid robots were at roughly 7,000 units.

So when we ask where the money is in physical AI, we need to separate a large automation industry that already exists from the newer bet that smarter AI will let machines handle far more types of physical work.

The first market is measured in hundreds of thousands of machines and billions of dollars of revenue. The second could eventually become much bigger, but a lot of its value still comes from what investors think these robots will be able to do later.

What actually counts as physical AI?

Physical AI is AI that can perceive a physical environment, decide what to do and then make a machine act on that decision.

A warehouse robot choosing its own path around an obstacle qualifies. So does an autonomous truck deciding how to react to traffic, a surgical robot assisting with a procedure or a robot arm using vision to manipulate unfamiliar objects.

A conventional machine following the same fixed sequence every time sits closer to traditional automation.

The line is getting blurrier because old robotics companies are adding better vision, machine learning and AI control to products that have existed for years. ABB, FANUC, KUKA, Universal Robots and warehouse automation suppliers increasingly overlap with what newer companies call physical AI.

That commercial overlap is important. If we only count newly founded AI robotics startups, we miss a large part of where customers are spending money today.

Where is physical AI already generating billions of dollars?

The biggest physical AI businesses currently sit in factories, warehouses, cars and operating rooms.

Intuitive Surgical generated more than $10 billion of annual revenue from robotic surgery. KION's automation business generated more than €3 billion last year, and its renamed Intelligent Automation Solutions division reached €1.63 billion in the first half of its latest financial year, up 18%.

Symbotic has also moved well beyond startup-sized revenue. Its latest reported quarter produced $721 million, 22% more than a year earlier, while net income reached $55 million. AutoStore generated roughly $539 million last year from automated storage systems while keeping a gross margin above 72%.

Automotive physical AI is already sizeable too. Mobileye expects close to $2 billion of annual revenue from driver-assistance and autonomy technology. Hesai shipped almost 1.1 million lidar units in the first half of its current financial year alone.

None of these businesses requires a human-shaped robot.

Customers are already spending billions on physical AI when it solves a narrow, expensive problem reliably.

Company or business	Recent revenue scale	What customers pay for	Commercial position
Intuitive Surgical	More than $10B annually	Surgical robots, instruments and service	Large-scale
KION Intelligent Automation Solutions	€1.63B in H1	Warehouse automation and software	Large-scale
Symbotic	$721M latest quarter	AI warehouse systems	Scaling quickly
Mobileye	About $2B expected annually	Automotive AI processors and software	Large-scale
AutoStore	$539M annually	Automated storage systems	Large-scale
Hesai	$127M latest quarter	Lidar for vehicles and robots	Scaling quickly
Aurora	$3M in first commercial year	Autonomous trucking	Very early

Why is warehouse robotics making so much money right now?

Warehouse robotics currently has one of the clearest business cases in physical AI because warehouses combine expensive labor, repetitive movement and environments that are much easier to automate than streets or homes.

The latest IFR figures show 117,500 transportation and logistics robots sold in one year, up 21%. They represented 47% of all professional service robot shipments.

Customers do not need a robot with human-level intelligence here. A machine that moves pallets, totes or products for thousands of hours can already save labor and increase throughput.

Symbotic shows how large that demand can become. Quarterly revenue has now reached $721 million, up 22% year over year, and adjusted EBITDA more than doubled to $95 million. KION's Intelligent Automation Solutions revenue rose almost 18% in the first half of its latest year, while its order book approached €3 billion.

AutoStore shows a different side of the same market. Its revenue is smaller, but its standardized cube-storage system produced a 72.4% full-year gross margin. In its latest reported fourth quarter, gross margin reached 73.7% and adjusted EBITDA margin was 43.3%.

Those are very different economics from custom engineering projects.

Warehousing gives us an early view of how physical AI value gets divided. Big bespoke systems can produce enormous revenue, while standardized proprietary platforms can capture much more profit from each dollar of sales.

Are humanoid robots making real money yet?

Humanoid robots are finally becoming a commercial market, but current sales are still tiny beside conventional industrial and logistics robotics.

According to the latest IFR market data, around 7,000 full-size humanoids were sold globally in one year for professional applications. That compares with 117,500 transportation and logistics service robots and more than 600,000 new industrial robots installed in factories.

The interesting part is that commercial evidence is starting to improve.

Unitree says it delivered more than 5,500 humanoid robots to end customers in its latest reported year, while total production exceeded 6,500 units. Agility Robotics says Digit has accumulated more than 65,000 hours of real-world operation and has secured more than $300 million of multi-year orders for its next generation, subject to contractual milestones.

At GXO, Digit passed 100,000 tote movements while maintaining roughly 98% on-task accuracy. Agility now reports commercial deployments with companies including GXO, Schaeffler and Toyota Motor Manufacturing Canada.

That gives the humanoid market much more substance than it had a couple of years ago.

Still, the gap with established automation remains enormous. Humanoids are currently an early industrial product category with some serious customers, rather than a major source of physical AI revenue.

Why are factories becoming the first real market for humanoids?

Factories give humanoid robots a much better chance of earning money today than homes because the tasks, environment and economics are easier to control.

Figure's work with BMW is a good example. Figure 02 robots performed sheet-metal loading at the Spartanburg factory and accumulated more than 1,250 operating hours while handling over 90,000 parts. The deployment contributed to production of more than 30,000 vehicles.

Agility is following the same route through manufacturing and logistics. Unitree's growing shipment volume also shows that buyers are willing to purchase humanoid hardware before anything close to household-level general intelligence exists.

Industrial customers can tolerate a much narrower robot.

A factory does not need the machine to cook dinner, fold laundry and walk the dog. It may only need the same robot body to move containers in one part of the plant and load components somewhere else.

That smaller jump from specialized automation to flexible automation is much easier to monetize.

For now, humanoid robotics looks far more like the next chapter of industrial automation than the beginning of a consumer robot market.

Is the real money in robot hardware or in the AI brain?

The AI brain could eventually become one of the most profitable layers of physical AI, and we are now seeing the first meaningful revenue evidence behind that idea.

Skild AI is the strongest example. The company says it crossed $100 million in annual recurring revenue only ten months after its first commercial deployment, with more than 60 paying customers.

Around 90% of that revenue comes mainly from manipulation rather than mobile robots. Skild has also moved beyond lab partnerships: it acquired Zebra Technologies' robotics business, formerly Fetch Robotics, and is working with ABB, Universal Robots, NVIDIA and Foxconn.

The model is attractive because Skild wants the same intelligence system to control many different robot bodies.

If that works broadly, customers would no longer need to program every machine and every task almost from scratch. A common robotic intelligence layer could be deployed across arms, mobile robots, humanoids and other machines.

Investors are already pricing in a huge amount of that future. Skild raised $1.4 billion at a valuation above $14 billion, which is more than 140 times the $100 million ARR figure disclosed later in the year.

The revenue is now real. The hard part is proving that one robot brain can generalize widely enough to deserve software-like economics across the physical world.

If you want more recent data on this point, please see our latest physical AI market report.

Are chips, sensors and robot parts better businesses than complete robots?

Physical AI components can be excellent businesses when they remain technically difficult to replace, but rapidly falling hardware prices can destroy that advantage surprisingly quickly.

Mobileye shows the attractive version. The company expects roughly $2 billion of revenue this year, mostly from technology embedded inside vehicles rather than from building complete cars. Its business combines EyeQ processors, perception software and deep integration with automakers.

Hesai shows what happens when volume explodes while prices fall.

The company shipped 628,275 lidar units in its latest quarter, up 78% from a year earlier. Robotics lidar shipments alone rose 193%. Hesai was profitable and generated roughly $127 million of quarterly revenue.

At the same time, the broader lidar market has experienced a steep decline in average selling prices as production scales.

Hesai is already reacting by moving further into robotics hardware. Its new robotic actuation business had delivered more than 10,000 modules by the end of the second quarter, with the company targeting six-digit annual shipment volumes later.

The component opportunity is attractive when a supplier owns difficult technology and improves faster than prices fall. Once the part becomes interchangeable, enormous physical AI growth can still produce a mediocre business.

Is autonomous driving finally a real physical AI business?

Autonomous driving is now unquestionably a real commercial service, although we still do not know whether robotaxis will become a great business financially.

Waymo provides more than 400,000 autonomous rides per week across six major US metropolitan areas. The company completed about 15 million rides last year, more than triple the previous year's volume, and its fleet has accumulated more than 127 million fully autonomous miles.

The operating scale is also becoming easier to understand. Waymo disclosed that around 70 remote-assistance agents are on duty globally at any given time for a fleet of roughly 3,000 vehicles. Those vehicles cover more than four million miles per week.

That is a functioning transport network, not an experimental pilot.

Profitability remains the missing piece. Alphabet does not publish a separate Waymo income statement, so we cannot see the full cost of vehicles, sensors, depots, cleaning, charging, insurance, operations and continuing R&D.

Autonomous trucking is earlier. Aurora generated only $3 million in revenue during its first commercial year while spending $745 million on R&D.

Autonomous driving has passed the “does anyone actually use this?” test. The next one is whether autonomous miles can produce attractive returns after all the physical infrastructure around them is paid for.

What does surgical robotics teach us about physical AI profits?

Surgical robotics shows how powerful physical AI becomes when one robot sale leads to years of recurring revenue.

Intuitive Surgical generated roughly 84% of revenue from recurring sources in its previous full-year disclosure. That includes instruments and accessories, service contracts and operating leases.

The business therefore earns far more than the price of the da Vinci system itself.

As hospitals perform more procedures, they keep purchasing proprietary instruments and paying for service. A growing installed base produces more recurring revenue without Intuitive having to sell a completely new robot every time.

This is one of the strongest physical AI business models we can already observe at scale.

A robot manufacturer that only gets paid when another machine leaves the factory will always carry heavy manufacturing and sales costs. A company that keeps earning when an installed machine performs another surgery, mile, pick or productive hour has much better economics.

That lesson is now spreading well beyond healthcare.

Will Robot-as-a-Service become the main way companies pay for physical AI?

Robot-as-a-Service is becoming a major physical AI business model because customers increasingly want to pay for useful work rather than own an expensive robot upfront.

The latest IFR service robotics report says traditional equipment sales are still the largest monetization model globally, but RaaS continues to gain ground and has already become the dominant model in the United States.

The appeal for customers is straightforward. A warehouse can compare a monthly robot fee with the cost of labor. A trucking company can compare an autonomous-driving fee per mile with the cost of a human-driven truck.

Suppliers like the recurring revenue too, although owning thousands of expensive robots can quickly absorb cash.

That is why some of the more interesting models separate ownership of the machine from ownership of the intelligence. Aurora eventually wants truck manufacturers, carriers and other partners to own much of the physical fleet while Aurora charges for its Driver-as-a-Service technology.

We expect that structure to spread.

Physical AI becomes financially much more attractive when outside capital finances the machine and the technology provider keeps collecting a software, service or usage fee.

If you want more recent data on this point, please see our latest physical AI market report.

Where does the money go when robots need lots of integration and service?

Integration and service create a large physical AI business, but the margins are usually less exciting than the revenue.

KION's warehouse automation business makes the trade-off easy to see. Its Intelligent Automation Solutions division reached €1.63 billion of revenue in the first half of its latest year and grew 18%, yet adjusted EBIT margin was 6.5%.

Building automated warehouses involves much more than delivering robots. Engineers need to connect conveyors, storage systems, software, sensors and existing customer infrastructure. Installations can take months or years, and every site is slightly different.

The service work that follows is valuable because customers cannot simply stop a warehouse when something breaks.

GXO shows another version of the integration opportunity. The logistics company uses autonomous mobile robots, robotic arms, sortation systems and other technologies inside customer facilities without needing to invent every machine itself.

There will be plenty of money in making different physical AI systems work together.

We would simply expect more of the long-term profit to sit with companies that own differentiated software, components or platforms than with businesses doing mostly custom engineering.

Which physical AI jobs make the most economic sense today?

Physical AI makes the most money today in jobs where machines can work for long hours, labor is expensive and the task is repetitive enough to automate without solving general intelligence.

Warehouses fit almost perfectly. Factories do too. Mining adds another useful example because autonomous machines can operate in closed environments while moving huge quantities of material.

Caterpillar already has hundreds of autonomous haul trucks operating around the world. Mining customers care about tonnes moved, equipment utilization and safety rather than whether the machine can imitate a person.

Surgical robotics works for another reason: each procedure has high economic value, and the machine supports highly skilled labor rather than trying to replace every part of a hospital worker's job.

Autonomous highway trucking also looks attractive on paper because driver costs are significant and highway driving is more structured than unrestricted urban driving.

Across these markets, the common feature is easy to see. The customer can put the robot into a financial model and estimate what it saves or produces.

That remains a much better predictor of physical AI revenue than how impressive a robot looks in a demo.

Why will recurring software and service become so important in physical AI?

Recurring software and service are likely to capture a growing share of physical AI profits because they let companies earn more from machines that are already installed.

The gap between business models is already large.

Symbotic still earns most of its revenue from deploying large automation systems. AutoStore sells hardware too, but its standardized proprietary architecture produces unusually high margins. Intuitive has pushed the model much further by generating most of its revenue repeatedly from its installed surgical systems.

Ocado offers another useful variation. Its technology agreements can tie ongoing fees to the retail sales flowing through an automated customer fulfillment center rather than simply charging for the equipment once.

The attractive endpoint is clear: hardware gets the technology into the customer's operation, then software, maintenance, consumables or usage fees keep producing revenue.

As seen above with surgical robotics, that structure can completely change the economics of a robot company.

Physical AI companies will probably experiment heavily with fees per mile, procedure, pick, productive hour or robot under management as deployments become larger.

Which part of physical AI is most at risk of becoming a commodity?

Generic robot hardware faces the biggest long-term risk because falling component prices can make impressive shipment growth much less profitable.

Lidar already gives us a useful preview. Hesai has been shipping far more sensors while industry pricing has fallen sharply. Robot actuators, cameras, mobile bases and arms could eventually go through the same process.

Chinese manufacturers are particularly important here because they are pushing production volumes up and robot prices down.

Cheaper hardware is excellent for physical AI adoption. It also makes life harder for manufacturers whose main advantage is simply producing a physical machine.

The companies in a stronger position will own something customers struggle to replace: proprietary control software, unusually good actuators, specialized processors, valuable training data, a large installed base or a tightly integrated ecosystem.

Apple showed that hardware can still capture exceptional economics when the product, software and distribution reinforce one another. Physical AI could produce similar companies.

A generic robot body with no lasting advantage will have a much tougher time.

If you want more recent data on this point, please see our latest physical AI market report.

Are investors putting money into the same physical AI markets as customers?

Investors are currently putting much more weight on general-purpose robot intelligence and humanoids than today's customer spending would justify on its own.

Skild's valuation above $14 billion is one obvious example. Figure has raised billions for general-purpose humanoids. Apptronik has attracted close to $1 billion of financing. Agility is preparing to enter public markets at a $2.5 billion pre-money equity value while it moves from early deployments toward larger orders.

Customer spending is still far more concentrated in established industrial automation, warehouse systems, automotive technology and medical robotics.

SoftBank's agreed $5.375 billion acquisition of ABB Robotics captures both sides of the bet unusually well. ABB brings an installed industrial robotics business with decades of customer relationships, while SoftBank explicitly wants to combine that footprint with AI and next-generation computing.

Instead of choosing between old robotics and new physical AI, SoftBank is buying an existing route into factories and betting that better intelligence makes that installed platform more valuable.

That approach may prove easier than building both the robot intelligence and the commercial distribution network from zero.

Where are the best margins in physical AI right now?

The strongest physical AI margins currently appear where proprietary technology can be sold repeatedly without rebuilding a custom system for every customer.

AutoStore is a strong example. Its full-year gross margin remained above 72%, and its latest fourth-quarter adjusted EBITDA margin reached 43%.

Mobileye has historically produced semiconductor-like gross margins because automakers buy its processors and software at large scale.

Intuitive Surgical combines proprietary hardware with recurring instruments and service, allowing it to turn an installed robot into years of additional revenue.

Project-heavy warehouse integration sits much lower. KION's Intelligent Automation Solutions business produced a 6.5% adjusted EBIT margin in the first half of its latest year despite generating more than €1.6 billion of revenue.

Autonomous driving companies are currently at the other extreme because huge R&D and operating costs arrive long before meaningful revenue.

The most interesting future economics probably sit somewhere between hardware and pure software: proprietary physical systems deployed widely, then monetized again through intelligence, usage or service.

Physical AI model	Current revenue maturity	Margin potential	Main economic problem
Surgical robotics ecosystem	Very high	Very high	Regulation and clinical adoption
Standardized warehouse platform	High	Very high	Customer investment cycles
Automotive AI chips and software	High	High	Automaker pricing power
Large warehouse integration	Very high	Moderate	Custom engineering
Robot foundation models	Emerging	Potentially very high	Generalization still unproven
Robot-as-a-Service	Growing	Potentially high	Who finances the robots
Autonomous trucking	Very early	Potentially high	Huge upfront R&D
Humanoid robot sales	Early	Unknown	Reliability and hardware cost

If you want more recent data on this point, please see our latest physical AI market report.

Where will the biggest physical AI market eventually be?

The biggest physical AI market could eventually be general-purpose labor automation, but specialized automation will probably keep most of the money for several more years.

The theoretical humanoid opportunity is enormous because a sufficiently capable robot could work in buildings, factories and warehouses designed around people. Companies would not need to rebuild every environment around a specialized machine.

That possibility explains the billions flowing into humanoid startups.

The commercial evidence is still much stronger elsewhere. Factories now operate more than five million industrial robots. Logistics robot sales exceed 100,000 units annually. Robotaxis already complete hundreds of thousands of paid trips every week. Surgical robotics has become a multibillion-dollar recurring business.

Meanwhile, humanoids are only beginning to move from thousands of units toward larger commercial fleets.

Robot foundation models could change the speed of that transition. Skild reaching $100 million ARR so quickly suggests that customers will pay for more flexible robot intelligence before general-purpose humanoids reach mass deployment.

So the next physical AI boom may come partly from making the millions of machines that already exist much smarter, rather than waiting for an entirely new robot population.

So where is the money in the physical AI market?

The money in physical AI today is concentrated in warehouses, factories, medical robotics and automotive systems, while the biggest future upside sits in recurring robot intelligence and more flexible automation.

The current market rewards machines that solve narrow problems extremely well.

Warehouse automation already generates billions of dollars because customers can measure labor savings and throughput. Industrial robotics operates at massive scale. Surgical robotics shows that a physical AI platform can build a highly profitable recurring ecosystem. Automotive AI chips and sensors show how much value can sit inside one component rather than the complete machine.

Humanoid robots are starting to join that commercial market, and the latest deployments are much harder to dismiss than the demonstrations of a few years ago. Agility has more than $300 million of multi-year orders, Unitree says it has delivered thousands of humanoids, and industrial customers are accumulating real operating hours.

Their economics still have much more to prove.

The more interesting shift now is happening one layer higher. Skild's rapid move past $100 million in recurring revenue suggests that companies may pay heavily for robot intelligence that works across several types of hardware. If that model scales, physical AI software could eventually capture a large part of the value created by cheaper robot bodies.

For now, we would put the money in three places.

Specialized automation has the most revenue. Proprietary platforms with recurring service have the best proven economics. General-purpose robot intelligence has the largest credible upside.

Humanoids could eventually connect all three, but they have not done it yet.

If you want more recent data on this point, please see our latest physical AI market report.