Who actually makes money from humanoid robots: the robot makers or the companies that sell them parts and brains?

Humanoid robots are machines with two legs and two arms, built to work in places designed for people. Their story pulls together several separate threads: sixty years of industrial automation, half a century of university research into walking and grasping, cheaper sensors and much better AI, precision parts from Japan, China and Europe, and employers looking for new ways to get physical work done. Interest surged in 2024 and 2025 as chipmakers, carmakers and Chinese industrial policy all took up the idea, but robots shipped and tasks completed are not the same thing as profitable fleets. So far no public evidence shows a large profit pool that belongs to humanoids alone. Robot makers are starting to sell robots, mostly at a loss. The parts and AI companies earn large profits, but almost all of that money comes from their much bigger businesses elsewhere.

A factory robot learns to walk

In 1961 a machine went to work on a General Motors die-casting line. It was a heavy hydraulic arm called Unimate, and its job was dull, hot and dangerous: lifting glowing metal parts out of a casting machine and stacking them, the same way, all day.¹ It had no legs, no face and no ambitions beyond that one task. That is why it worked.

Unimate came from two men with different gifts. George Devol was an inventor who had patented a "programmed article transfer" device, a machine that could store a sequence of movements and repeat them. Joseph Engelberger was an engineer with a salesman's instinct, and he saw that Devol's device could take over factory jobs nobody wanted.¹ Their company, Unimation, sold the first industrial robots to carmakers. Those buyers already had something rare: production lines that repeated the same motions millions of times.

Nearly everything later in this story comes back to this insight. The first durable robot market grew by making the job narrower. A factory arm does not need to understand the world. It needs to be in the right place, grip the right part and move it the same way every time. The factory is built around the machine: guarded cells, fixed fixtures, parts that always arrive in the same orientation. Repeatability mattered far more than looking like a person.

The market that grew from that logic is now very large. The International Federation of Robotics counted about 542,000 industrial robots installed around the world in 2024, more than double the figure a decade earlier, and more than half of them went to China.² Almost all of them are arms bolted to a floor or a ceiling, doing one job in one place.

While Unimation was teaching arms to stack castings, a different question was taking shape in Tokyo. In 1970 Professor Ichiro Kato of Waseda University and his colleagues began the WABOT Project. It brought together engineers from several disciplines to study a machine shaped like a person.³ Kato was a mechanical engineer interested in artificial limbs and the mechanics of human movement, and Waseda's program reflected that. It treated the human body as an engineering problem to understand, not a costume to put on a machine.

By 1973 the team had finished WABOT-1, a full-scale humanoid that combined a limb-control system, a vision system and a conversation system in one research platform.³ It could walk a few steps, grip objects with its hands and respond to simple spoken Japanese. It was slow and fragile, and nobody bought one. That was never the point. WABOT-1 was a research question built from metal and wiring: what would it take for a machine to move through rooms, stairs and doorways made for people, and to use tools shaped for human hands?

So in the early 1970s two kinds of robot existed side by side, built on opposite philosophies. The industrial arm made money by changing the workplace to suit the machine. The humanoid was meant to suit the workplace people had already built. The first approach was cheap to justify: a carmaker could count the labour hours an arm replaced and the injuries it prevented. The second promised something larger, one machine that could do many jobs, but it had no customer and no price.

That gap is the subject of this story. Today's pitch for humanoid robots is the WABOT question backed by venture capital and AI: rather than rebuild a factory around a robot, build a robot that fits the factory. The pitch is persuasive because the industrial-robot business really is large and still growing. But sixty years of successful factory arms show that automation pays when the task is defined. They say nothing about whether a general-purpose biped can beat a task-specific machine on cost or reliability. If anything, the history points the other way: the winners in automation have been the companies that made jobs narrower, not broader.

Before anyone could test whether a humanoid could earn its keep, someone had to solve a more basic problem. A machine on two legs falls over. The next chapter follows the company that spent fifteen years and a great deal of money teaching one to stay up.

↑ Back to Top

The expensive promise of a human-shaped machine

In 1986 engineers at Honda began a project that sounded almost absurd for a carmaker. They built a pair of mechanical legs called E0 and tried to make it walk on its own. E0 moved forward one deliberate step at a time, with a long pause between steps while it reset its balance.⁴ A toddler would have overtaken it. It was the start of a fifteen-year effort, and the slowness was the point: Honda's engineers were learning how hard walking really is.

Picture carrying a tray of full glasses across a crowded room. Every step moves the support under the tray. Your body shifts the tray, your ankles, hips and arms in small corrections you never think about. A two-legged robot faces the same problem with no instinct at all. It has to sense its own tilt and the force under each foot, calculate a correction, and send commands to its motors fast enough that a stumble is caught before it becomes a fall. The analogy breaks down at the hardware. A person's correction is free and automatic. A robot's correction runs through sensors, motors, gears and control software, and each of them adds weight, cost, delay and a chance of failure.

Honda worked through that problem one generation at a time. The E-series moved from static walking, where the robot kept its weight over one foot before each step, to dynamic walking, where it let itself start to fall forward and caught itself, much as people do. Over the following years Honda added a torso and arms and produced the P-series prototypes.⁴

The public result arrived in 1996. Honda presented P2, a human-sized robot that walked by itself, climbed stairs and pushed a cart, with its computer and batteries carried on board instead of trailing cables.⁴ To engineers it was striking. Until then, independent two-legged walking at human scale had been mostly a laboratory dream. A year later P3 did the same with a smaller, lighter body.⁴ The sources credit these milestones to Honda's engineering teams, not to a single inventor, and that seems fair: the walking came from years of accumulated control work, not one breakthrough.

In November 2000 Honda announced ASIMO. It was about the height of a child, lighter than its ancestors, and presented as a robot that could one day act as a partner to people.⁵ Hiroyuki Yoshino was Honda's president at the time, and the company treated the robot as a statement of what its engineering culture could achieve.⁵ ASIMO became one of the most photographed machines in the world. It waved, climbed stairs, ran in later versions, and appeared at museums and trade shows.

The muscles and joints

Underneath ASIMO's white shell was a supply chain that still matters for this story. Each joint needs a motor to produce motion, a gear reducer to turn the motor's fast, weak spin into slow, strong and precise movement, bearings so the joint turns smoothly under load, and an encoder that reports exactly where the joint is. Think of these as the robot's muscles and joints. The comparison leaves out the electronics and software that coordinate them, which act more like a nervous system, and it leaves out the fact that a robot joint wears out and has to be serviced.

Japan was well placed to supply those parts. Companies such as Harmonic Drive Systems had spent decades building strain-wave gearing: a compact reducer in which a flexible metal ring meshes with a rigid one, giving large gear reductions with almost no slack. That precision matters when a robot has to put its foot exactly where its controller expects. Harmonic Drive and other precision-motion firms grew up serving factory automation, machine tools and aerospace. Humanoids were a showcase for their parts, not a market.

Applause is not a customer

What happened to ASIMO commercially is the most important lesson of this chapter. Honda offered it for rental at events and showrooms. It drew crowds, but there was never a buyer who needed an ASIMO to do work. Human-like ability and huge public attention did not turn into a mass-market worker.

The reasons are clear in hindsight. ASIMO could walk and climb but had little ability to handle unfamiliar objects. Its batteries ran for a limited time. It was expensive to build and maintain, and it worked best in prepared settings. A factory buyer had cheaper and more reliable options for nearly any job ASIMO could do. Honda later directed the research behind ASIMO toward more practical uses, a decision covered in the next chapter.

The precise lesson matters, because the industry keeps relearning it. Honda proved that a human-sized biped could balance, walk, climb and carry. It did not prove that anyone would pay for it. Mobility was necessary but not enough. A robot that can reach a workstation still has to do something useful once it gets there, every shift, cheaply enough to beat the alternative.

By the late 2000s, then, humanoids could walk on flat floors in prepared rooms. The next test was much harsher: could a robot do useful work when the world around it was broken?

↑ Back to Top

A disaster test, a famous failure, and a new kind of brain

In June 2015, in a fairground arena in Pomona, California, some of the world's most advanced robots tried to open a door. They were in the finals of the DARPA Robotics Challenge, a competition built around tasks a rescue robot might face after a disaster: drive a vehicle, get out of it, open a door, turn a valve, cross rubble, climb stairs. Several robots stalled for long minutes before simple tasks. Some fell over, sometimes in slow motion, and crashed onto the floor.⁶ The falls became an internet joke. They also showed, more clearly than any lab paper had, the distance between a robot that can move and a robot that can work.

The challenge grew out of a real disaster. In March 2011 the Fukushima Daiichi nuclear plant failed after an earthquake and tsunami, and people had to enter places too dangerous for them. The episode exposed how few machines could go into a damaged human building and do what a person would do: open doors, climb ladders, close valves. DARPA, the US Defense Department's research agency, launched its Robotics Challenge in 2012 to fund hardware and software that could.⁷

At the centre of the contest was Atlas, a hydraulic humanoid built by Boston Dynamics. It made its public debut in July 2013 as a DARPA-funded platform. At first it was tethered to an external power supply and handed to competing teams.⁷ That arrangement is worth noticing. Boston Dynamics supplied a powerful body, and university and laboratory teams wrote much of the software that decided what the body did. Hardware and "brains" were separate pieces of the product, made by different organisations, more than a decade before investors began arguing about which of the two would capture the profits.

The finals answered one question and raised another. Robots could finish disaster-style tasks, but slowly, with heavy human supervision over communication links and with frequent failures in the physical details: a handle that was not quite where the model expected it, or a step that shifted underfoot. The hard part was no longer walking on a flat floor. It was coping with a world that did not match the plan.

Two famous robots leave the stage

The commercial record during the same years was blunt. In June 2018 Honda said it would stop developing ASIMO and apply what it had learned to more practical technologies, such as mobility assistance and robot components.⁸ The most famous humanoid of its era ended as a research archive, not a product line.

SoftBank Group $9984.T, the Japanese telecoms and technology investor, provided a second warning. Its Pepper robot was a waist-high machine with a tablet on its chest, sold to shops, banks and homes as a friendly assistant. In 2021 Reuters reported, citing sources, that SoftBank had stopped producing Pepper after it struggled to build a lasting global customer base.⁹ Pepper rolled on wheels, so it is not a direct humanoid comparison, and its failure belongs more to the social-robot market than to factory work. But the pattern was the same: huge visibility, a human-like presence, and too few customers who kept paying once the novelty wore off.

A new kind of brain

While the bodies struggled, a separate thread began moving fast. For decades robots had used hand-written rules and simple cameras. An engineer would program the robot to find an edge, measure a distance and grip at a set point. That worked in tidy factories and failed in messy ones.

In 2012 three researchers at the University of Toronto, Alex Krizhevsky, Ilya Sutskever and their supervisor Geoffrey Hinton, trained a deep neural network that beat rival methods by a wide margin in a major image-recognition competition, ImageNet. Hinton had spent decades backing neural networks while most of the field had given up on them. The result revived interest in deep learning across computing, including machine vision, though critics soon warned about the limits of systems that learn patterns from data without understanding them.¹⁰ Graphics processors, the chips that made the training practical, became central to AI. Over the following decade, simulation software and much larger datasets gave robot developers new ways to train behaviour without writing every rule by hand.

The limit was just as important as the gain. Recognising a mug in a photo is a different problem from picking up that mug when it is half full, slippery and in an awkward spot on a cluttered shelf. Perception tells a robot what is there. It does not tell the robot how hard to squeeze, how to recover when the mug slips, or how to avoid hitting the worker standing beside it. Better seeing did not automatically produce safe, dexterous action, and that remains the central technical gap in 2026.

By the early 2020s the threads were coming together. AI models had become far more capable, the chips to run them were more compact, sensors were cheaper, and a few factories and warehouses were willing to test robots on their own floors. The next chapter shows what happened when a chip company, Chinese planners and a handful of customers all decided at once that the humanoid's time had come.

↑ Back to Top

From lab demonstrations to production floors

On 18 March 2024, at NVIDIA's $NVDA annual developer conference in San Jose, chief executive Jensen Huang shared the stage with a row of humanoid robots and announced Project GR00T, a general-purpose foundation model for humanoids. Alongside it came updates to Isaac, NVIDIA's robot simulation platform, and a new onboard computer for robots called Jetson Thor.¹¹ The scene marked a change in how the industry pitched itself. The question was no longer only "can we build an impressive body?" It became "can we make robot behaviour cheaper to develop and reuse, the way software is?"

Huang was a natural messenger for that pitch. He co-founded NVIDIA in 1993 to make graphics chips for video games. Over three decades he turned it into the main supplier of the processors used to train AI, partly by building CUDA, a software toolkit that made developers comfortable programming NVIDIA hardware. Robotics fitted the same pattern: give developers tools, simulation and models that run best on NVIDIA chips, and let the ecosystem build on top.

A foundation model is a large AI model trained on a broad range of data so that it can be adapted to many tasks, rather than built for one. Think of it as a broad starting library of skills that a robot developer can tune for a particular job. The analogy stops at the physical world. A language model that gives a slightly wrong answer can be corrected in the next sentence. A robot that makes a slightly wrong motion can drop a part, damage a machine or hurt someone, and its skills have to hold up when lighting, objects and floors change.

NVIDIA's announcement also showed the business model. NVIDIA sells a platform: chips, software and models that robot makers use. It does not build humanoids. Its robotics revenue is not reported separately; it sits inside much larger segments. Later, Agility Robotics said it planned to use Jetson Thor in a future generation of its Digit robot,¹² a named design-in. Platform availability is confirmed. How much money NVIDIA makes from humanoids is not disclosed.

Policy as an accelerator

China was moving at the same time, and on the supply side. In 2023 the Ministry of Industry and Information Technology (工业和信息化部, MIIT) issued national guidance on humanoid-robot innovation. It set goals for an initial innovation system and progress on core technologies by 2025, and a stronger industrial ecosystem by 2027.¹³ Cities turned that guidance into money and sites. Beijing published a 2025–2027 embodied-intelligence plan in March 2025 covering technology, local supply chains and application scenarios.¹⁴ Shanghai followed in August 2025 with a plan aiming for a 500-billion-yuan core industry by 2027 and support for qualifying projects.¹⁵

Policy of this kind is a real force. It brings capital, test sites, training data and supplier attention to an early industry. It also carries a familiar risk from China's earlier pushes into solar panels and electric vehicles: capacity built for targets rather than for customers, and orders placed for demonstration instead of profit. A target is not an order book.

From claims to workplaces

The more useful evidence of the period came from named customers. GXO Logistics $GXO, a contract-logistics company that runs warehouses for other businesses, signed what it called an industry-first multi-year agreement with Agility Robotics to deploy Digit, a two-legged robot built to move totes. GXO described it as a robots-as-a-service arrangement, meaning GXO pays for the robots' use instead of buying them.¹⁶ Amazon $AMZN announced it was testing Digit to move empty totes, and it had invested in Agility.¹⁷ Amazon's statement describes a test, not a procurement programme, and Amazon has many alternatives, including its own fleets of wheeled robots.

In the car industry, BMW Group $BMW.DE confirmed that Figure AI's humanoid had been used in production at its Spartanburg plant in South Carolina, and later announced next-stage humanoid work in Germany.¹⁸ Figure, a private Californian start-up, reported that its robot had loaded more than 90,000 parts and run for more than 1,250 hours over an eleven-month deployment.¹⁹ The distinction matters: BMW confirms the deployment, while the operating figures come from Figure. Neither company has published the contract value, robot count or site economics.

Apptronik, a Texas start-up spun out of university robotics research, took a manufacturing route. Its Apollo robot gained a production partner when Jabil $JBL, one of the world's largest contract electronics manufacturers, announced a collaboration to build Apollo robots and use them in its own operations.²⁰ Apptronik also ran pilots with Mercedes-Benz Group $MBG.DE and GXO, and announced work with Google DeepMind on AI models. Hyundai Motor Group $005380.KS, which bought Boston Dynamics in 2021, showed the production version of a new all-electric Atlas in January 2026 and described a staged plan to put it into Hyundai plants, starting with parts sequencing.²¹ Boston Dynamics named Hyundai Mobis $012330.KS, the group's parts maker, as the Atlas actuator supplier, so the robot's joints are a family affair.

Magnets become geopolitics

One input then became a political question. Robot joints rely on electric motors, and the best compact motors use permanent magnets made from rare earths such as neodymium, with heavier rare earths added to keep them stable at high temperatures. China dominates the processing of these materials and the making of magnets. In April 2025 China's Ministry of Commerce and customs authority imposed export licensing on selected medium and heavy rare-earth items, with State Council approval.²² Overnight, magnet sourcing became a supply-security issue for carmakers, defence firms and robot makers alike.

That shift gave strategic weight to the few non-Chinese suppliers. MP Materials $MP runs the main US rare-earth mine and is building toward magnet production. Lynas Rare Earths $LYC.AX of Australia is the largest separated rare-earth producer outside China. Japan's Shin-Etsu Chemical $4063.T is a major magnet maker. None of these companies discloses meaningful humanoid revenue. Their relevance to robots is an option on the future, and in the near term it is outweighed by cars, wind turbines and defence.

The first scaling signals

By 2025 the counting had begun. The research firm Omdia estimated that more than 13,000 humanoids shipped worldwide that year. It ranked three Chinese makers first, second and third: AgiBot (智元机器人) with 5,168 units, Unitree (宇树科技) with 4,200 and UBTech (优必选) with about 1,000, as reported by the South China Morning Post.²³ Counterpoint Research separately estimated 16,000 humanoid installations.²⁴ On 22 April 2026 Tesla $TSLA said its first-generation Optimus production lines were being installed in preparation for volume production.²⁵ On 19 August 2026 Unitree began trading on Shanghai's STAR Market under the code 688836, becoming a listed, pure humanoid maker.²⁶

Each of these raised the stakes. None of them answers whether a deployed humanoid earns back its cost. To see who might eventually be paid, follow the money from the magnet to the paid task.

↑ Back to Top

The money trail, from magnet to paid task

Start inside a single robot elbow. The robot's computer decides the forearm should rise ten degrees. It sends a signal to a motor driver chip, which feeds carefully timed current into a motor. Inside the motor, copper coils push against rare-earth magnets and the shaft spins quickly. A reducer trades that speed for force, bearings carry the load, an encoder reports the exact angle back to the computer, and cameras and force sensors check that the hand is where it should be. A robot has dozens of joints like this. Every one of them is a small purchase order.

Now follow the robot outward and the chain takes shape in six steps.

At the start are materials and magnets: rare-earth miners and separators, and the magnet makers that turn their output into finished magnets. They sell to motor and actuator makers, not to robot companies. Next come precision motion parts: motors, reducers, bearings, roller screws and complete actuator modules, sold by Japanese specialists, Chinese car-parts groups and European bearing companies. Alongside them sit sensing suppliers: image sensors, depth cameras, lidar, force and position sensors, and the chips that convert their signals. Then comes compute and models: onboard processors, power and motor-control chips, simulation software and AI models, sold or licensed by chip companies and cloud providers.

Those inputs meet at the robot maker, which designs the machine, integrates the parts, trains its behaviour, proves it is safe and builds it, sometimes in its own factory and sometimes through a contract manufacturer such as Jabil. Finally there is deployment and service: mapping a site, changing processes, writing a safety case, training workers, keeping spare parts and fixing broken robots. A maker, an integrator or a rental provider does that work and invoices the customer: a carmaker, a logistics company, one day perhaps a household.

What Empor's numbers do and do not say

Empor tracks more than sixty listed companies across these six layers. Their combined results look spectacular: revenue up 29% in the quarter to June 2026 and an operating margin of 24.6%. Read naively, that sounds like a booming robot industry. It is not, and the reason is the most important point in this chapter.

Those companies are counted whole. The chips-and-intelligence layer reported combined revenue of about $744 billion and a 46% operating margin in Empor's data, which is overwhelmingly NVIDIA's data-centre business and Microsoft's $MSFT cloud and software. The "humanoid makers" layer grew 61.9% in the June quarter, but that layer includes Samsung Electronics, whose quarterly revenue more than doubled on a memory-chip boom. Samsung's robot work does not drive Samsung's results. Strip out the giants and the humanoid-specific money is small, scattered and mostly undisclosed. The available data does not show a large, recurring profit pool that belongs to humanoids.

So the question is not who looks profitable. It is who is profitable because of humanoids. Layer by layer, the answer looks different.

Where today's money actually sits

The AI and chip companies have the highest margins in the chain. NVIDIA earned a 60.4% operating margin on $216 billion of revenue in its fiscal year to January 2026. But robotics is a small slice that NVIDIA does not report separately, so its humanoid earnings cannot be measured. For now, being the most visible platform is a position, not a profit stream.

The precision-component suppliers sell into large, established markets: cars, machine tools, factory robots, appliances. Their parts may be essential to a humanoid, yet as a group they earn thin profits. Empor's precision layer generated about $138 billion of revenue at a net margin below 1% in the latest fiscal years, weighed down by restructuring at large groups such as Nidec $6594.T and Schaeffler. No public record shows a component supplier with a sole-source humanoid contract or a take-or-pay deal, where the buyer must pay for volumes whether it uses them or not.

The robot makers invoice the customer, so they carry the product risk: warranties, inventory, field service and the cost of a robot that falls over. UBTech $9880.HK, a Shenzhen robot company listed in Hong Kong, is one of the clearest tests. In 2025 it reported about CN¥821 million ($114 million) of humanoid and related revenue, roughly 41% of its sales, and its fiscal-year operating margin was minus 38.9%. Unitree's prospectus tells a different story. Group revenue was CN¥1.70 billion in 2025 with net profit of CN¥278 million; adjusted profit excluding non-recurring items was CN¥591 million, a gap explained mainly by share-based pay.²⁷ Humanoids made up about half of Unitree's revenue, alongside its quadruped "robot dogs" and components. The company-wide profit cannot be treated as a humanoid margin.

The buyers may end up with the largest share of the gain if robots work, through lower task costs, more output or safer jobs. Public payback data is close to non-existent.

Renting, not buying

GXO's arrangement with Agility illustrates a design choice in the value chain. Under robots-as-a-service, the customer pays for robot time or output instead of buying the machine. That lowers the customer's risk and makes the first purchase easier. It also moves the risks to the provider: if the robot sits idle, breaks down or becomes obsolete, the provider absorbs the loss and has to finance the fleet. Rental helps adoption but makes the maker's economics harder.

How the profits could move

The chain's profits are not fixed. Early on, scarce parts can command high prices: a reducer good enough for a humanoid's knee, or a magnet that does not need Chinese heavy rare earths. If volumes grow and many suppliers qualify, those parts become standard and procurement teams push prices down, as has happened with car components. Profit could then shift toward whoever makes robots reliable across many sites: fleet software, integration and service, plus the task data gathered along the way. If makers compete hard on price, as Chinese makers already do, much of the gain could flow to buyers instead.

Each handover in the chain has its own trap. A component order is not a lasting margin. A robot delivery is not productive use. An AI platform design-in is not a licence fee. Rising customer capital spending is not humanoid demand: carmakers and logistics firms spend on many things.

That raises a harder question than "who is connected to humanoids?" It is "who has evidence of real execution?" Layer by layer, the contests look quite different.

↑ Back to Top

The contests: scarce joints, cheap bodies, and brains for hire

Which constraint decides who gets paid: the joint that moves the robot, or the software that tells it what to do? Investors have answered both ways, and share prices have swung accordingly. In Empor's 9 October snapshot, shares in the humanoid-maker layer had roughly doubled over a year while magnet and materials shares had fallen by about a third. Prices tell you what the market believes. To judge whether that belief is earned, go layer by layer.

Magnets: strategic, but not about robots yet

The magnet layer looks strong on paper: combined revenue up about 50% in the June quarter and operating margins back to 12.6% from 2.1% in 2024. China Northern Rare Earth $600111.SS, the state-linked giant of light rare earths, and JL Mag Rare-Earth, a large magnet maker, both grew. Lynas grew fastest among the profitable names, with revenue up 80% in its year to June 2026 and a net margin above 22%, as higher prices and non-Chinese demand fed through.

The cause is mostly not humanoids. Rare-earth prices, export controls, electric cars and industrial motors drive this layer, and a few thousand robots a year barely register against tens of millions of cars. MP Materials shows the gap between strategic value and earnings: its operating margin in 2025 was deeply negative while it builds US magnet capacity. Its value rests on supply security, a different idea from robot demand. In this layer the humanoid story is a real long-term option attached to businesses whose results are driven elsewhere.

Joints and actuators: the precision contest

Harmonic Drive Systems is the established name in strain-wave reducers outside China. Its position was built over decades on precision, low wear and long qualification records with robot and machine-tool makers. Its results show a recovery without much profit yet: sales rose 7% in the year to March 2026 and the operating margin was only 4.3%, though the June quarter improved, with revenue up about 24% and an 11% operating margin. Its shares traded at about 200 times trailing earnings in Empor's 9 October data. That price reflects hopes for humanoids far more than current earnings.

Its closest challenger is Suzhou Leaderdrive (绿的谐波), a Chinese reducer maker that has won domestic customers with lower prices and local supply. On Empor's annual numbers Leaderdrive is growing faster (revenue up 47% in 2025) and reported a higher operating margin, about 26%, but from a base around one-fifth of Harmonic's. Its margin fell to about 13% in the latest quarters and its free cash flow turned negative. Leaderdrive is closing the gap on growth; Harmonic still leads on scale and record. The contest is open, and the market prices both richly.

Two large Chinese car-parts groups, Zhejiang Sanhua $002050.SZ and Tuopu $601689.SS, have been reported as Optimus actuator suppliers. Neither discloses how much humanoids contribute. Their results are driven by heat pumps, valves and car chassis parts; Tuopu's quarterly growth slowed to about 6% and it missed revenue forecasts in each of its last four quarters. Shenzhen Inovance $300124.SZ, China's leading factory-automation drives company, is developing joint modules alongside a profitable core business. Hyundai Mobis has the one confirmed actuator contract in this layer, for Atlas, but at $43 billion of revenue the robot business is too small to show up in its results.

Other suppliers add to the picture without settling it. Japanese bearing and motion companies such as Minebea Mitsumi $6479.T and NSK, Taiwan's Hiwin, Timken in the US, Regal Rexnord $RRX with its frameless motors, and South Korea's tiny Robotis each sell parts that fit humanoids. Robotis shows how far expectations can run ahead: a company with about $28 million of revenue was valued at nearly 90 times its sales. In this layer, the deciding evidence would be repeat orders at prices that survive cost-cutting and second suppliers.

Sensors: many eyes, few disclosed wins

In sensing, Analog Devices $ADI, which makes precision signal-processing chips, is the most profitable large company, with a company-wide operating margin of about 27%. Specialists grow faster from small bases. Ouster $OUST, a US lidar maker, grew revenue by 52% in 2025 while losing heavily; Hesai and RoboSense in China sell lidar mainly to carmakers. TDK $6762.T, STMicroelectronics $STMPA.PA, Goertek $002241.SZ and Sunny Optical supply sensors, components and optics with large exposure to phones and cars. No company in the layer reports a material humanoid revenue share. The architecture is also uncertain: if makers settle on cheap cameras plus AI, expensive sensors may lose their place. For now this layer has technical relevance without a measurable stake.

Brains for hire

NVIDIA has the most visible developer platform for humanoids. Its position comes from its developer tools and a complete package of simulation, models and onboard compute, and Agility's planned use of Jetson Thor is one named example. Rivals compete at every level. Qualcomm $QCOM is pushing robotics processors. Arm Holdings $ARM collects royalties on many chip designs. AMD $AMD offers embedded processors. Alphabet's $GOOGL DeepMind has announced model collaborations with Apptronik and Boston Dynamics. Meta Platforms $META funds robotics research. In China, Horizon Robotics (地平线) has launched robot chips but still earns nearly all its revenue from cars, with an operating margin of about minus 109% in 2025.

The chip firms that supply motor control and power, including Texas Instruments $TXN, Infineon $IFX.DE and NXP $NXPI, sell into every robot and every electric car, so humanoids add little to their results. A general-purpose robot model only becomes a business once physical tasks run reliably and licence fees appear in someone's accounts. No company's accounts show that yet.

The makers

Among listed makers, Unitree and UBTech are the most direct tests because their disclosures separate humanoid sales. Unitree's shipment lead is real but needs careful reading. Omdia estimated 4,200 shipments in 2025. Unitree reported more than 5,500 shipments that year under its own definition, which excludes wheeled dual-arm robots, and 5,632 humanoid sales over 2023 to 2025.²⁷ These are different measures from different sources, and neither counts robots doing paid work. Many go to universities, labs and entertainment. AgiBot's top place in Omdia's ranking carries the same caveat, and AgiBot is private.

UBTech's figures point in a better direction while it keeps losing money: revenue in the June 2026 quarter roughly doubled and the quarterly gross margin rose to 39%. Tesla's Optimus remains the flagship that has not yet sold anything. Tesla has disclosed line installation, not units delivered or revenue. Boston Dynamics has a guaranteed first customer in Hyundai, but internal sales between group companies can hide whether Atlas would earn a return on its own. Figure's last funding round valued it at $39 billion after the money came in, in September 2025.²⁸ That is a private figure and says little about what ordinary shareholders would get. Mobileye bought Mentee Robotics in February 2026,²⁹ adding another corporate humanoid project.

The diversified groups mostly hold options rather than businesses: Xiaomi $1810.HK, XPeng, LG Electronics $066570.KS, Midea $000333.SZ, Doosan Robotics, Kawasaki Heavy Industries $7012.T and Rainbow Robotics $277810.KQ, which Samsung controls. Rainbow's market value of about 184 times sales shows how much hope these options can carry.

Builders and buyers

Jabil's work with Apptronik is a real production relationship, but its scale and revenue are not disclosed, and it makes no visible difference to a company with $36 billion of annual sales. Hon Hai $2317.TW, the company best known as Foxconn, could be both a user and a builder of humanoids. Hyundai is both owner and buyer. These overlapping roles may speed up learning, but they make it harder to see the maker's standalone economics.

If the evidence is this patchy company by company, does it look any clearer when the layers are watched together as volumes rise?

↑ Back to Top

When more robots move through the chain

When a robot maker decides to build a batch, its first orders go upstream: reducers, motors, magnets, sensors. Those orders arrive months before any customer pilot becomes a line in the maker's accounts. In principle that makes suppliers an early signal of robot volumes. In practice a supplier's sales are dominated by cars, chips or consumer devices, and a few hundred extra robot joints disappear in the noise.

Empor tests these links by comparing how each layer's results move against the humanoid makers' combined revenue and against customers' capital spending, quarter by quarter, at different time lags. The results follow a story, but not the simple one.

Magnets show the clearest pattern. Magnet makers' revenue growth has tracked the makers' revenue a quarter later quite closely, and the pattern holds for every company with enough history. The cause is almost certainly not humanoid demand. Magnet volume for robots is tiny next to cars, and the "maker" revenue series is itself dominated by Samsung, Xiaomi and Midea, whose businesses respond to the same electronics and industrial cycles that drive magnet demand. Currency adds to the confusion: as the yuan strengthened about 6% against the dollar over the past year, Chinese magnet makers' margins rose, the opposite of what a simple export story would predict. This pattern is worth investigating, not proof that robots caused the growth.

Sensors move only loosely with makers. STMicroelectronics, TDK and Analog Devices tracked maker revenue in the same quarter, a sign of a shared electronics cycle. The lidar specialists went the other way: Ouster and RoboSense slowed while maker revenue rose. If robots were driving lidar demand, those two would be the first to show it.

Actuator margins show no consistent response. The intuitive story says each new robot needs dozens of reducers and screws, so supplier margins should widen as maker volumes grow. In the data, fewer than half of the twenty component companies moved that way. Sanhua, Inovance and Wuzhou Xinchun moved the opposite way. Volume gains in China are colliding with price competition, and in any case the humanoid volumes are too small to move these companies' results.

Chip margins show no humanoid effect. Chip-company revenue moves with customers' capital spending, as you would expect for firms that sell into factories and data centres. But operating margins do not rise when humanoid makers grow. Horizon Robotics, the chip firm most focused on robots and cars, moved opposite to the expected pattern. NVIDIA's results cannot be read as humanoid results.

Customer spending is not a robot order book. The seven buyers Empor tracks spent about $69 billion on capital projects in the June 2026 quarter, up 68% on a year earlier. Most of that was Amazon's data centres and SoftBank's AI investments. Maker revenue does tend to rise a couple of quarters after customer spending, but loosely, and LG showed no link at all. A carmaker's new paint shop is not a humanoid purchase.

All of this rests on a short history. Humanoid revenue was negligible before about 2023, so most of these tests measure supplier and customer cycles that happen to include a few robot companies. Acquisitions, changes in product mix and currency moves distort them further. A few years of numbers moving together is evidence, not proof of cause. What the data does show is narrower: there is no sign yet of humanoid volumes passing through to supplier margins.

So the money trail gives no answer on its own. The answer depends on whether the robots themselves start to earn their place, and there the optimistic and pessimistic stories diverge sharply.

↑ Back to Top

The robot must earn its place

Stand in a warehouse aisle and watch a wheeled robot carry a tote from one station to another. It has no legs to balance, no hands to calibrate, and a battery that lasts a full shift. For a fixed transport job on flat floors it is cheaper, safer and easier to maintain than any humanoid. The IFR's service-robot survey shows where buyers have actually spent: about 199,000 professional service robots sold in 2024, led by transport and logistics machines, most of them wheeled, with the robots-as-a-service fleet up 31%.³⁰ A humanoid has to earn its extra complexity. It earns it only where legs, hands and human-like reach make a real difference: stairs, tools, existing workstations, processes no one wants to rebuild.

The optimistic chain

The hopeful version runs like this. Better models and simulation make robots more reliable at defined tasks. Reliable robots need fewer human interventions. Fewer interventions mean robots work more of each shift and customers earn back their cost sooner. Customers then order more robots for more sites. Every deployment produces task data that improves the models, and every production run lowers manufacturing costs. Cost per productive hour falls, and the robot takes on neighbouring tasks. In that world, a general platform beats a collection of single-purpose machines.

Each link in that chain needs proof from customers, not from launches. Videos of robots folding shirts show capability. They say nothing about how often the robot fails, how many people supervise it, or what it costs when it does.

The pessimistic chain

The cautious version is just as coherent. Robots perform well in structured tasks and fail outside them. Buyers add supervision, safety barriers and integration work. Downtime and service visits eat up the labour savings. Customers choose wheeled robots, fixed arms or a redesigned process instead. Makers cut hardware prices to win orders, and suppliers compete their margins away. In that world the humanoid becomes an expensive research platform, and the profit stays with established automation companies.

History gives the pessimistic chain real weight. ASIMO and Pepper drew huge attention without lasting demand, and the industrial-robot business grew by narrowing tasks. Safety rules add friction too: the international industrial-robot standard ISO 10218-1 was revised in 2025, and US safety guidance from OSHA still asks for a hazard analysis for each application. A machine that can fall, pinch or move unexpectedly near workers needs a safety case for every task and every site.³¹³²

Does the labour shortage settle it?

The usual case for humanoids starts with demographics, and the starting point is solid. The OECD reports that the working-age population across its member countries has begun to decline and is projected to keep shrinking through 2060. In April 2025 about one in six euro-area industrial firms and one in four service firms said labour shortages were limiting production.³³ That is a lasting reason to automate.

But labour scarcity varies by place and time. US data from the Bureau of Labor Statistics showed manufacturing job openings at about 324,000 in July 2025, lower than a year earlier.³⁴ A labour problem also does not imply a humanoid solution. Employers short of workers can buy wheeled robots, conveyors or fixed cells, or redesign the work. The demographic case supports automation in general. It does not decide which form wins.

Does the shipment count settle it?

The shipment figures look like momentum, and in one sense they are. Omdia's estimate of more than 13,000 shipments in 2025 sits well above its own 2024 forecast, which expected shipments to pass 10,000 only by 2027 and reach 38,000 by 2030.³⁵ Counterpoint's figure of 16,000 installations measures something else: robots placed at a site, not robots delivered. Unitree's own count uses a third definition. None of them says how many machines work paid shifts. Long-range forecasts are even shakier. Goldman Sachs raised its estimate of the 2035 market from at least $6 billion to $38 billion between research updates, a sign of how much those totals depend on assumptions.³⁶ Faster shipments show that makers can build robots. They do not show that buyers keep using them.

China's double edge

China has real advantages: dense supplier networks, manufacturing skill, cheaper components and active state support. These lower costs and speed up trials, and they explain why Chinese makers lead on shipments. The same forces can produce too many factories and orders that are placed to meet targets rather than to make money. The question is not whether China can make humanoids cheaply. It clearly can. The question is whether Chinese buyers keep using what they buy.

What could move the profits

Several changes could shift profit between layers. Open-source robot models, or AI that runs on the robot without a cloud connection, would limit how much model owners can charge. Integrated actuator modules could change which parts suppliers win. Standardised hardware would push profit toward fleet service and integration. Tighter export controls would raise the value of magnets made outside China. Each of these favours a different layer, and none has happened decisively.

With strong arguments on both sides, the useful question is what evidence would tip the balance.

↑ Back to Top

The signals that would settle it

The simplest test is a customer placing another order after the first robot has worked long enough to be judged. Until that becomes visible across many customers, a small number of published figures give earlier hints. The financial readings below come from Empor's tracking of June 2026 company results, and each one can be checked as new filings appear.

UBTech's revenue growth is the clearest listed test of whether announced humanoid orders become delivered, invoiced robots. It moves before profits because deliveries are booked as revenue first. UBTech reports twice a year. In the quarter to June 2026 revenue grew 99.4% on a year earlier. Sustained growth above 50% with a rising humanoid share would support the demand story. A slide back toward low growth would suggest the orders were mostly pilots. Revenue alone does not prove profitability, and UBTech's results still include education robots and other products.

UBTech's gross margin settles a separate argument: whether makers can price robots for profit or must buy orders with discounts. It reached 39.2% in the quarter to June 2026, and the company publishes it in its interim and annual reports.³⁷ A margin that holds or rises as volumes grow would suggest makers have some pricing power. A fall below the low 30s would point to price competition from other Chinese makers passing the gains to buyers.

Harmonic Drive's revenue growth sits upstream. Reducer orders come before robots ship, so this figure can move a few quarters ahead of maker deliveries. Harmonic reports quarterly. Revenue grew 23.6% in the quarter to June 2026. Double-digit growth tied by management to robotics orders would suggest build volumes are rising. Flat sales would say humanoid volumes are still too small to matter. Harmonic does not report humanoid sales separately, so growth from factory robots or chip equipment can look the same.

Tesla's capital spending as a share of revenue shows Optimus factory building before any Optimus revenue exists. It was 12.5% in the quarter to June 2026, published quarterly. This is a whole-company figure that also includes car plants and AI computing. A step-up clearly tied to Optimus lines in Tesla's own disclosures would support a volume push. Delays or cuts would weaken the claim that Tesla leads the US effort. Spending is not robots delivered.

NVIDIA's robotics revenue would show how many makers are moving from prototypes to products. NVIDIA reports quarterly but does not separate robotics, so the latest reading is not disclosed. A reported robotics, or automotive-and-robotics, line growing above 50% would support broad adoption. A flat reading would mean few robots are reaching production. Total NVIDIA growth, driven by data centres, is not a substitute.

Behind all five sit the figures that would settle the debate directly and are almost never published: paid repeat deployments by site, productive hours per robot, how often humans have to step in, and customer payback against the best alternative. When a maker or customer starts publishing those, it will tell investors more than any quarterly revenue figure.

↑ Back to Top

Who gets paid if the machines work?

So who makes money from humanoid robots today? The parts and AI suppliers make plenty of money. NVIDIA, Analog Devices, Texas Instruments, the magnet makers and the precision-component groups all run profitable businesses, some extremely so. But that money comes almost entirely from other markets: data centres, cars, phones, factory automation, electric motors. For them humanoids are a small, mostly undisclosed sideline, and Empor's data shows no sign yet of robot volumes lifting their margins.

The robot makers have the clearest claim on future robot revenue, because they send the invoice. They also carry the warranty, the inventory, the service vans and the risk that a robot fails in front of a customer. Most of them still lose money. Few disclose standalone economics, and several of the most important are private or part of larger groups.

There are early exceptions. Unitree's prospectus shows a maker can sell thousands of humanoids and report a profit for the whole company. Its robots, though, go largely to research and entertainment buyers, and its profit includes other products. UBTech has more visible humanoid revenue and an improving gross margin, but it still loses heavily. Neither shows that fleets of robots pay off for the customers who run them.

The question would be settled by evidence that is currently missing: repeat paid deployments across many sites, productivity confirmed by customers rather than vendors, makers whose cash flow improves as they grow, and segment reporting that separates robot sales from the parent company's results. Until then, the honest answer is that company-wide profit today and humanoid profit tomorrow are separate questions, and an investor can be right about the robots and still pick the wrong layer, the wrong company or the wrong price.

The first Unimate made its case on a single hot, dangerous job on a GM casting line. Factory automation then grew one well-defined task at a time. Humanoids promise to do many tasks, but they will have to win customers the same way: one task, one site and one repeat order at a time.

↑ Back to Top

Glossary

Actuator: A motor-and-gear assembly that moves a robot joint.

Bipedal walking: Walking on two legs while constantly correcting balance.

Foundation model: A broad AI model that can be adapted to perform different tasks.

Humanoid: In this story, a general-purpose robot with two legs and two arms.

Installation: A robot placed into use at a site. Installation alone does not prove productive work.

Manipulator: A robot arm or hand used to handle objects.

Motion policy: A set of learned or programmed rules that turns what a robot senses into movement.

RaaS: Robot-as-a-service, where customers pay for robot access or output instead of buying the machine outright.

Reducer: Gearing that trades a motor's speed for controlled force at a joint.

Task-specific automation: Machinery designed for one defined job, often easier to optimise than a general-purpose robot.

Teleoperation: A human remotely controlling or assisting a robot.

Utilization: The share of available time a robot spends doing useful work.

Working capital: Cash tied up in inventory, money owed by customers, and service parts.

References

1. AI & Robotics Timeline — Computer History Museum ↩↩

2. Global Robot Demand in Factories Doubles over 10 Years — International Federation of Robotics, 2025 ↩

3. History of Humanoid Robot — Waseda University Humanoid Robotics Institute ↩↩

4. History of Robotics Development — Honda ↩↩↩↩

5. Honda Debuts New Humanoid Robot "ASIMO" — Honda, 20 November 2000 ↩↩

6. The Iron Man Challenge — Time, 2015 ↩

7. Atlas Robot Debuts — DARPA ↩↩

8. Robotics Research Initiatives — Honda, 28 June 2018 ↩

9. Exclusive: SoftBank Shrinks Robotics Business, Stops Pepper Production — Reuters via Investing.com, 2021 ↩

10. Deep Learning: A Critical Appraisal — arXiv, 2018 ↩

11. NVIDIA Announces Project GR00T Foundation Model for Humanoid Robots and Major Isaac Robotics Platform Update — NVIDIA, 18 March 2024 ↩

12. NVIDIA Jetson Thor: The Ultimate Platform for Physical AI — NVIDIA Blog ↩

13. Interpretation of the Guidance on Humanoid Robot Innovation and Development — Ministry of Industry and Information Technology, 2023 ↩

14. Beijing Embodied Intelligence Technology Innovation and Industry Action Plan (2025–2027) — Beijing Municipal Government, 4 March 2025 ↩

15. Shanghai Embodied Intelligence Industry Implementation Plan — Shanghai Municipal Government, 6 August 2025 ↩

16. GXO Signs Industry-First Multi-Year Agreement with Agility Robotics — GXO Logistics ↩

17. Amazon Introduces New Robotics Solutions — Amazon ↩

18. BMW Group to Deploy Humanoid Robots in Production in Germany for the First Time — BMW Group ↩

19. Production at BMW — Figure AI ↩

20. Apptronik and Jabil Collaborate — Jabil ↩

21. Hyundai and Boston Dynamics Unveil Atlas — Associated Press, January 2026 ↩

22. Announcement No. 18 of 2025 on Export Control of Certain Medium and Heavy Rare-Earth Items — Ministry of Commerce and General Administration of Customs, April 2025 ↩

23. Chinese Firms Outpace US Rivals in 2025 Humanoid Robot Shipments as AgiBot Takes Lead — South China Morning Post, January 2026 ↩

24. Global Humanoid Robot Installations Reach 16,000 Units in 2025 as Mass Production Picks Pace — Counterpoint Research, 14 January 2026 ↩

25. Tesla Shareholder Update, Q1 2026 — Tesla Investor Relations, 22 April 2026 ↩

26. Unitree Robotics STAR Market Listing Notice — Shanghai Stock Exchange, 18 August 2026 ↩

27. Unitree Prospectus Summary — Shanghai Stock Exchange, 11 August 2026 ↩↩

28. Figure Exceeds $1B in Series C Funding at $39B Post-Money Valuation — Figure AI, September 2025 ↩

29. Mobileye Global 2026 Annual Report Filing — US Securities and Exchange Commission ↩

30. Service Robots See Global Growth Boom — International Federation of Robotics, 2025 ↩

31. ISO 10218-1:2025 Robotics — Safety Requirements — International Organization for Standardization ↩

32. OSHA Technical Manual, Section IV Chapter 4: Industrial Robot Systems — Occupational Safety and Health Administration ↩

33. OECD Employment Outlook 2025 — OECD, 2025 ↩

34. Job Openings and Labor Turnover Survey, July 2025 — US Bureau of Labor Statistics, 3 September 2025 ↩

35. Global Humanoid Robot Shipments to Exceed 10,000 Units by 2027 and Reach 38,000 Units in 2030 — Omdia, July 2024 ↩

36. The Global Market for Humanoid Robots Could Reach $38 Billion by 2035 — Goldman Sachs ↩

37. UBTech Robotics Filing — HKEXnews, 24 April 2026 ↩