How do humanoid robotics business models actually work?

How much revenue do humanoid robot companies really make?

The leading Chinese humanoid companies now make hundreds of millions of renminbi a year, while some well-known U.S. players still record only a few million dollars. Most of that money comes from selling robots and projects, not from proven autonomous labor at scale.

The first surprise is that even the size of the market depends on what we count. Omdia estimated 13,318 humanoid robots shipped globally in 2025, up almost 480% in one year. IDC's 2026 commercialization review puts 2025 shipments above 18,000. IDC also estimates that more than 85% of deployments went into performances, education, data collection and guided services, with industrial manufacturing and logistics still at the pilot stage.

Company filings make the gap concrete. Unitree's Shanghai prospectus says it shipped more than 5,500 humanoids in 2025, with humanoid revenue of RMB 868 million. In 2026, UBTECH reported RMB 590.3 million of full-size humanoid revenue from 921 units in the first half alone. Agility's SEC filing, by contrast, shows net sales of $1.78 million in 2025 and $174,000 in the first half of 2026. Revenue is not profit, so we separate who makes money from who burns cash in our report on humanoid robots.

But a robot sold to a university for embodied-AI research and a robot working thousands of hours in a factory both count as one unit. Right now, shipment growth is running ahead of proven labor substitution.

That is why humanoid robotics can look commercially real and economically immature at the same time, and why a famous brand is not the same thing as a large revenue line.

What do humanoid robotics companies actually sell?

Humanoid robotics companies currently make money in several different ways, and putting all of them under "robot sales" hides most of what is interesting about the market.

Unitree is closest to a conventional robotics manufacturer. Customers buy standardized humanoid hardware, and the company records product revenue. AgiBot is following a similar path with a broad catalogue covering humanoids, wheeled robots, quadrupeds and dexterous systems.

UBTECH sells a heavier industrial package. Its own description is "hardware + software + service + operation", and its financial statements classify full-size humanoid products and services together. The customer is paying for more than a body coming off an assembly line.

Agility Robotics goes further toward selling capacity. Digit is offered through Robots-as-a-Service agreements, including commercial deployments with GXO and Toyota Motor Manufacturing Canada. The robot, fleet-management software and continuing service sit inside the same contract. Its SEC filings now show how small that revenue still is compared with the commercial story.

At home, 1X is trying yet another model. NEO is currently offered for $20,000 or through a $499 monthly subscription. 1X says first U.S. deliveries start in 2026, but it has not disclosed how many paying households have received a robot.

The same humanoid body can therefore sit inside businesses that look financially more like industrial equipment, system integration, leasing, software-enabled services or consumer electronics.

Business model	What the customer pays for	Current examples	Where the economics come from
Robot sales	Ownership of the hardware	Unitree, AgiBot	Manufacturing volume and hardware margin
Industrial solution	Robot plus integration and services	UBTECH, Figure	Larger project value and deployment work
Robots-as-a-Service	Recurring access to robot capacity	Agility Robotics	Multi-year recurring revenue
Consumer ownership/subscription	Access to a household robot	1X NEO	Hardware revenue or monthly payments

Why would a factory buy a humanoid instead of a normal industrial robot?

A factory should buy a humanoid when using a human-shaped machine is cheaper and more flexible than redesigning the whole workflow around traditional automation.

A fixed robotic arm will usually beat a humanoid at a repetitive task that never changes. It can be faster, simpler and cheaper because it does not need legs, balance, batteries, complex perception or humanlike hands. An autonomous mobile robot is also a better choice when the job is simply moving material across a flat warehouse.

Humanoids become interesting in the messy space between those systems. Factories already contain stairs, racks, carts, tools, workstations and containers built around the human body. If one machine can operate inside that infrastructure, companies may automate new tasks without rebuilding every workstation.

BMW's progression with Figure shows what customers are trying to learn. Figure 02 first handled sheet-metal components in Spartanburg. Figure 03 has since returned for a harder logistics-sequencing workflow involving manipulation, movement and cart handling. Hyundai is taking a similar route with Boston Dynamics' Atlas, beginning with parts sequencing before moving toward more complex assembly work.

The economics get much better if the same robot can move between jobs. A humanoid permanently assigned to one simple task has to compete with purpose-built automation. A humanoid that can handle several workflows over its working life spreads the hardware and integration cost much more widely.

That flexibility is currently the strongest economic argument for the humanoid form.

Are companies really buying humanoid robots, or are they buying robot labor?

Industrial customers increasingly want to buy productive robot hours rather than take the technology risk of owning an experimental humanoid.

Agility Robotics is the clearest example. GXO signed a multi-year Robots-as-a-Service agreement after testing Digit in its logistics operation. Digit works alongside existing warehouse automation, while Agility Arc manages workflows, facility configuration, fleets and troubleshooting.

Toyota Motor Manufacturing Canada followed the same path. After a pilot, Toyota signed a commercial RaaS agreement for Digit to support manufacturing, supply-chain and logistics work. In October 2026, GXO agreed to add two more U.S. e-commerce sites, in North Carolina and Tennessee, with deployments expected to begin in early 2027 on Digit 4 before moving to Digit 5.

This changes the customer's decision. A warehouse does not need to guess what a five-year-old humanoid will be worth or whether today's hardware will become obsolete. It pays for an automation service and pushes more of the maintenance, technology and residual-value risk back onto the supplier.

For the robot company, the trade-off is harsher. Selling a $70,000 machine brings cash quickly. Renting the same machine for several years means building and financing it before collecting most of the revenue.

RaaS can eventually produce much better recurring revenue, but rapid growth can consume a lot of cash before the economics become attractive.

Can humanoid robots already beat human labor costs?

Humanoid robots can already beat human labor costs in a few high-utilization jobs, but there is currently no general cost advantage that works across factories and warehouses.

The latest detailed U.S. Bureau of Labor Statistics compensation data puts total employer cost for production, transportation and material-moving occupations at about $38.64 per hour. Manufacturing across all occupations averages roughly $48.27 per hour once wages and benefits are included. Human labor sets the price a robot must beat; we look at who pays for humanoids, and how much, in our humanoid robots market report.

At $38.64 per hour, 2,000 hours of work cost an employer about $77,000 a year. Four thousand hours cost roughly $155,000. A robot that can reliably cover extended shifts therefore has a large pool of labor cost to attack.

Purchase price alone tells us surprisingly little. Imagine a $60,000 humanoid lasting three years. At 4,000 productive hours per year, the hardware itself depreciates at only $5 per working hour before financing, maintenance, energy and support. At 1,000 productive hours per year, the same hardware costs $20 per hour before any of those additional expenses.

Speed matters too. If a robot completes a task at half the rate of a person, comparing its hourly cost with a human hourly wage becomes misleading.

The real economic unit is cost per successful task. Robot price, working hours, speed, failures, maintenance and human supervision all feed into that number.

Annual productive robot hours	Human labor cost at $38.64/hour	Labor value addressed
1,000	$38,640	Limited utilization
2,000	$77,280	Roughly one full-time workload
3,000	$115,920	Extended utilization
4,000	$154,560	Multi-shift potential

How much does humanoid robot uptime change the economics?

Humanoid robot uptime changes the business case enormously, which is why productive hours are much more useful than watching another impressive robot demo.

BMW gives us one of the rare public datasets from an actual factory. Figure 02 logged more than 1,250 operating hours, moved over 90,000 sheet-metal components and contributed to production of more than 30,000 BMW X3s during its Spartanburg deployment.

That works out to roughly 72 component movements per operating hour, or one every 50 seconds.

BMW also says the robot was used in ten-hour weekday shifts, although the companies have not published enough daily data for us to calculate a clean utilization rate across the full deployment period. That missing number is important. Two identical robots can have completely different economics if one produces useful work for 80% of a shift and the other manages 35%.

The useful metrics are pretty obvious: productive hours, successful tasks per hour, human interventions, mean time between failures and total maintenance cost. Agility said in September 2026 that its fleet had built up more than 65,000 hours of cumulative experience, but that is experience, not an audited uptime rate.

A robot manufacturer that cuts its hardware price by 20% gets headlines. A manufacturer that raises useful utilization from 50% to 80% may create far more economic value.

Why are humanoid robots starting with such boring factory jobs?

Humanoid robots are currently starting with tote moving, parts handling, sequencing and machine tending because those jobs give companies a clean way to measure whether the robot is worth paying for.

Agility's Digit began commercial work at GXO moving totes between existing automation systems. Toyota uses Digit around manufacturing, supply chain and logistics, and AGCO signed a commercial pilot in October 2026 for sorting and tote work at its Hesston, Kansas plant, planned for 2027. Mercedes-Benz has worked with Apptronik's Apollo on intralogistics, including moving components toward production lines. Hyundai plans to introduce Atlas first in parts sequencing and later in component assembly.

These jobs share a useful combination of traits. They are repetitive enough to automate, physical enough to benefit from a humanoid body, structured enough for today's AI, and variable enough that fixed automation can become awkward.

Customers can also measure the result immediately. How many parts moved? How many hours did the robot run? How many mistakes occurred? How often did a worker have to intervene?

General-purpose capability still matters, but mostly because it could make the next task cheaper to add. A factory does not need a philosophical proof that its humanoid understands the physical world. It needs the robot to move the right component to the right place thousands of times without stopping production.

Commercial humanoids are becoming general-purpose platforms one narrow job at a time.

Can selling humanoid robots actually become a profitable hardware business?

Yes, Unitree now gives the strongest public evidence that robot hardware can generate real revenue and profit, but no company has yet shown a standalone, repeatably profitable full-size industrial humanoid business. UBTECH shows how different the numbers become when the model includes heavy integration.

Unitree's Shanghai prospectus showed the economics behind the shipment headlines: more than 5,500 humanoids shipped in 2025, RMB 868 million of humanoid revenue and humanoids at 51.8% of total revenue. Its first public quarterly report, published in late October 2026, shows revenue of RMB 1.82 billion for January to September 2026, up 55.7%, and RMB 665 million in the third quarter, up about 70%.

Nine-month net profit reached RMB 386.7 million. The quality of that profit is less clean than the headline: excluding non-recurring items, profit fell 18.8% to RMB 349.0 million, and third-quarter net profit dropped 17.4% as R&D spending jumped. The figures also cover quadrupeds and other products, so they are not a humanoid-only margin. Unitree reported a core-business gross margin above 60% for 2025, which stands out in hardware; we break down where the margins really are in our report covering humanoid robots.

UBTECH shows another version of the business. Full-size humanoid products and services jumped from RMB 35.6 million in 2024 to RMB 820.6 million in 2025. In the first half of 2026, that line reached RMB 590.3 million, up 1,445%, or 46.5% of group revenue of RMB 1.27 billion. Gross margin rose to 44.7%.

Yet UBTECH still lost RMB 789.8 million in 2025 and RMB 338.8 million in the first half of 2026, a narrower loss but still a loss.

The difference is hard to ignore. Standardized hardware can already support attractive company economics. Large industrial projects generate much more revenue per robot, roughly RMB 640,000 per UBTECH full-size unit in the first half on our calculation, but still carry major R&D, integration, sales and operating costs.

Latest reported metric	Unitree	UBTECH
Total revenue	RMB 1.82B (Jan–Sep 2026)	RMB 1.27B (H1 2026)
Humanoid revenue	RMB 868M (2025); not broken out for 2026	RMB 590.3M (H1 2026)
Humanoid share of revenue	51.8% (2025)	46.5% (H1 2026)
Company gross margin	Above 60% core-business margin (2025)	44.7% (H1 2026)
Net result	RMB 386.7M profit (Jan–Sep 2026)	RMB 338.8M loss (H1 2026)

How much human work is still hiding behind humanoid robots?

Current humanoid deployments still require plenty of human integration, maintenance and sometimes remote assistance, so apparent robot productivity can contain more human work than a demo suggests.

BMW's factory deployment required production IT, occupational-safety teams, process managers and logistics staff. BMW also added barriers and partitions and improved 5G coverage around the work area. Getting the robot through the factory door was only the beginning.

Industrial suppliers are building service networks for the same reason. Agility Robotics partnered with Ricoh for field support, and in October 2026 announced a 50,000 sq ft Pittsburgh hub for customer support and software. Fleet software such as Agility Arc exists partly because deploying, monitoring and troubleshooting robots remains a real operational job.

The consumer version is even more explicit. 1X tells NEO customers that when the robot encounters a difficult task, an expert can remotely supervise its actions at a scheduled time. The task gets completed, while 1X also gains a new training example.

That can be perfectly workable during an early rollout. The problem starts if human support fails to shrink as fleets grow.

If ten robots need one remote operator, automation can still create considerable leverage. If every robot needs one person watching it, the company has mostly relocated the worker.

That ratio between robot hours and human-support hours deserves much more attention than it currently gets.

Will Robots-as-a-Service become the main business model for industrial humanoids?

Robots-as-a-Service is likely to become a major industrial humanoid model because it solves two of the customer's biggest worries at once: high upfront cost and rapid technological obsolescence.

The evidence is getting stronger, but it is not yet market-wide. Agility moved from its original GXO agreement to a RaaS contract with Toyota Motor Manufacturing Canada, and GXO is adding two more U.S. sites in early 2027. Hyundai has also put Robotics-as-a-Service inside its commercialization strategy. Meanwhile, Unitree and UBTECH show that outright hardware sales in China are already far larger than any disclosed RaaS revenue.

For a factory, a monthly or multi-year service contract is easier to compare with the labor cost it replaces. If the robot fails, needs maintenance or becomes obsolete, the supplier remains economically involved instead of leaving the customer with an expensive machine.

RaaS also forces the robot maker to care about the right things. A company that sells hardware can recognize revenue when the machine ships. A company selling robot capacity keeps earning only if the machine remains useful.

The difficulty moves onto the balance sheet. Someone still has to pay for thousands of robots before several years of service revenue arrive. Agility's filing shows the problem: tiny recognized sales next to heavy losses, while it seeks public-market cash. Robot makers may eventually work with leasing companies, banks or fleet-financing vehicles rather than funding every machine themselves.

That would make the humanoid industry financially look surprisingly familiar. The technology may be new, but the model starts resembling aircraft leasing, forklifts, industrial equipment and other expensive assets sold through long-term usage contracts.

Can humanoid robot software become a real SaaS business?

Humanoid robot software can become a valuable recurring-revenue business, but today customers still buy most of that software as part of the robot or deployment.

The pieces are already appearing. Agility Arc manages Digit fleets and workflows. AgiBot is building a Skill Store around downloadable robot capabilities. Figure continuously updates its Helix AI models. Robot manufacturers can eventually charge for orchestration, fleet analytics, security, new skills, AI upgrades and enterprise management. Charging for skills and fleet tools could add revenue; we look at where the biggest opportunities are in our report on humanoid robots.

The installed base determines how interesting this becomes. Software developed for 100 robots has limited revenue leverage. The same software distributed across 100,000 machines has very different economics because another copy costs almost nothing to deliver compared with building another humanoid.

That is one reason hardware scale matters even for companies that eventually want software-like margins.

For now, calling humanoid robotics a SaaS market is too early. Customers are primarily paying to get physical machines working in the real world. Software becomes a much stronger standalone business after the installed base exists.

Can humanoid robotics companies actually sell training data?

Training data is currently more useful as a competitive advantage than as a standalone humanoid robotics business, although China is beginning to blur that distinction.

Every deployed robot can potentially create demonstrations of how the physical world works: grasping an object, recovering from an error, navigating around people or coordinating two hands. Those examples can improve the next generation of models.

IDC says China has already accumulated tens of thousands of hours of embodied-AI data and datasets approaching petabyte scale. Data collection itself was one of the important humanoid deployment categories during 2025. Figure shows the internal value: in September 2026 it said Index pretraining lifted whole-task success in 30 unseen homes from 9% to 56% in its own evaluation. No company has disclosed material external data sales.

AgiBot's manufacturing ramp gives this strategy a much larger physical base. The company announced that its 15,000th embodied-AI robot had rolled off the production line in June 2026. AgiBot says the move from 5,000 to 10,000 robots took only three months, after roughly a year was required to move from the first 1,000 to 5,000.

The compounding effect is more interesting than selling datasets directly. More deployed robots can create more physical interaction data; better data can improve the models; better models should make the robots easier to sell into harder jobs.

For most leading humanoid companies today, data is closer to fuel for the core business than a separate product line.

Why is China selling so many more humanoid robots than the US?

China currently dominates humanoid robot volume because Chinese companies have built a much broader hardware market around research, education, data collection, entertainment and early industrial use.

The size of the gap is striking. IDC estimates Chinese vendors represented about 95% of global humanoid shipments in 2025 under its market definition. Omdia's narrower dataset still puts Chinese manufacturers overwhelmingly ahead, with AgiBot, Unitree and UBTECH occupying the top three shipment positions.

Production and sales have kept accelerating. AgiBot passed 15,000 cumulative embodied-AI robots produced in June 2026. Unitree's revenue rose 55.7% in the first nine months of 2026, and its implied first-half revenue of about RMB 1.15 billion, on our calculation, came in above its earlier guidance of RMB 1.05 billion to 1.13 billion. UBTECH sold 921 full-size humanoids in the first half alone.

The American model is different. Figure, Agility, Apptronik and Boston Dynamics have spent more time around a smaller number of large industrial customers where deployment quality matters more than maximizing shipments immediately. Chinese buyers still span labs, schools and demonstrations as well as factories, so the unit gap overstates the gap in productive work.

Neither route proves who will ultimately win industrial automation. China currently has the stronger evidence for manufacturing scale and hardware economics. American companies have stronger evidence in some highly integrated factory and warehouse deployments.

Over the next few years, those two models are likely to collide. Chinese manufacturers are moving deeper into industrial work while American companies are finally building factories capable of producing robots in the thousands.

Do Tesla and Boston Dynamics really have an easier path to customers?

Boston Dynamics has a huge built-in commercialization advantage through Hyundai, while Tesla has the same structural advantage in theory but has not yet reached comparable Optimus deployment.

Hyundai can use Atlas across Hyundai Motor, Kia, Hyundai Mobis and Hyundai Glovis before Boston Dynamics has to build a large independent customer base. Hyundai currently plans annual robot-production capacity of 30,000 units by 2028 and has discussed deploying more than 25,000 Atlas robots across Hyundai and Kia manufacturing operations. A parent supplying its own demand may have a head start, and we weigh who is likely to win in our humanoid robots market report.

The first Atlas deployments are already fully committed for 2026, with robots going to Hyundai's Robotics Metaplant Application Center and Google DeepMind. Full-scale manufacturing deployment at Hyundai Motor Group Metaplant America is planned from 2028, beginning with sequencing.

Tesla has an equally powerful captive environment on paper. Its factories give Optimus access to real production tasks and massive internal demand if the robot becomes useful.

But Tesla's 2026 disclosures are more cautious than earlier expectations. Its second-quarter update said its first Optimus production lines are still being installed in Fremont, replacing the old Model S and X lines. Tesla says the initial robots will go into an "Optimus Academy" for training-data collection and capability development. The company also removed earlier language about reaching Optimus volume production in 2026.

Captive demand can remove a difficult sales problem. It cannot remove the engineering problem. Hyundai currently has the more concrete deployment roadmap.

Does fast humanoid robot obsolescence make subscriptions more attractive?

Rapid humanoid robot upgrades make subscriptions unusually attractive because customers risk owning hardware that becomes technologically old long before it physically wears out.

Figure offers a striking example. After Figure 02 accumulated real factory experience at BMW, Figure said at the end of September 2026 that it had retired most Figure 02 robots as its Figure 03 fleet grew. The new robot changed the manufacturing architecture, hands, sensing, battery system and other core components.

That type of improvement is normal in an immature technology. It creates an awkward situation for a factory accustomed to buying equipment expected to remain productive for many years.

A five-year-old industrial robot arm can still be perfectly competitive at the task it was designed to perform. A five-year-old humanoid may have much weaker perception, hands, autonomy and operating time than a new generation.

RaaS shifts that risk toward the vendor. The customer can keep buying performance while the supplier decides whether a software update, component replacement or completely new robot is the cheapest way to deliver it.

The faster humanoid technology improves, the harder it becomes to convince customers that outright ownership is always the sensible option.

Can home humanoid robots become a bigger business than factory robots?

Home humanoid robots could eventually address a much larger customer base than factory robots, but the consumer business model currently asks far more from the technology.

1X is running the clearest commercial experiment. NEO costs $20,000 to own or $499 per month by subscription. According to 1X, customers reserved its entire first-year production capacity of 10,000 NEOs within five days of pre-orders opening. 1X says U.S. deliveries begin in 2026, but delivered units have not been independently confirmed.

That is real willingness to pay, even if a refundable deposit is much weaker evidence than 10,000 completed purchases.

Consumer economics are also different from factory economics. BMW can measure parts handled per hour. A family buying NEO is paying for convenience, time, novelty and eventually the ability to delegate household chores. There is no simple payroll line against which the robot can be compared.

The home is also a brutal technical environment. Clothes deform. Kitchens differ. Children move unpredictably. Pets get in the way. A consumer expects one machine to handle many of those situations. Figure's 56% whole-task success in unseen homes, from its own September 2026 test, is real progress but far from what a paying family expects.

A factory can justify an expensive humanoid that performs one job extremely well. A household paying $499 every month will eventually expect much broader usefulness.

That makes the home potentially enormous, but probably harder to monetize reliably in the near term.

Can Agility's $300 million in commitments hide a weak business?

Yes. Large commitments, a working robot and a weak business can all exist at once, and Agility's filings show why conditional commitments should never be read as firm orders or sales.

Agility says it has more than $300 million of multi-year Digit 5 commitments. The Robot Report, citing the filing, says they come from one customer, and Agility itself says the figure depends on milestones and does not represent current-period revenue. Against that, Agility booked $1.78 million of net sales in 2025, about $140 million of operating loss and roughly $100 million of cash burn, according to The Robot Report. Its first-half 2026 sales fell to $174,000 from $625,000 a year earlier. Churchill Capital Corp XI shareholders vote on the proposed listing on 2 November 2026, and approval and completion are separate steps.

Manufacturing exposes the same gap. Figure says BotQ moved Figure 03 from roughly one robot per day to demonstrating a one-robot-per-hour cycle in less than 120 days, with more than 350 units delivered internally and externally by the end of April 2026. Yet first-pass yield was still just above 80%.

Tesla has been unusually explicit about the same issue. In a 2026 earnings discussion, Elon Musk described Optimus as the hardest manufacturing scale-up Tesla has attempted because almost every component is new and there is no mature supply chain comparable with automotive parts.

Deployment cost can be equally dangerous. If each new warehouse requires months of custom engineering and continuous expert support, revenue may grow while margins remain poor. Customer concentration makes it worse: one delayed buyer can move a whole year.

Price competition adds another pressure. China's hardware manufacturers are already showing that sophisticated humanoids can be sold for tens of thousands of dollars rather than permanent six-figure prices. Western companies cannot build business plans around hardware staying expensive forever.

The successful companies will have to improve four things at the same time: robot capability, manufacturing cost, deployment cost and productive utilization. Being excellent at only one of them will probably not be enough.

So how do humanoid robotics business models actually work?

Humanoid robotics currently works as a hybrid business: companies make money from hardware and deployment today, while trying to turn those installed robots into recurring labor, software and service revenue over time.

The clearest current hardware model is already real. Unitree reported RMB 1.82 billion of revenue and RMB 386.7 million of net profit in the first nine months of 2026, although recurring profit fell. UBTECH sold 921 full-size humanoids in six months but still lost money. AgiBot's ramp shows how quickly volume grows once buyers extend beyond factories.

The industrial model is developing differently. UBTECH sells integrated robot solutions. Figure works closely with factories to prove particular jobs. Agility sells robot capacity through recurring contracts. Boston Dynamics can use Hyundai's own manufacturing network as an enormous first market. Each company takes a different route to revenue, so we sort out who is ahead and who is behind in our report covering humanoid robots.

The strongest long-term model is probably a combination of these approaches. Build the hardware cheaply enough to deploy in volume, charge customers around useful work rather than novelty, keep recurring revenue through software and service, and use the deployed fleet to improve the next generation of AI.

The crucial economic test remains very simple. A humanoid has to complete enough useful tasks, for enough hours, with little enough human help, that the total cost falls below the alternative.

Today, only a limited number of deployments clearly pass that test. But humanoid robotics revenue is no longer hypothetical. There are profitable hardware vendors, hundreds of millions of dollars of disclosed humanoid revenue, expanding RaaS agreements and real factory workloads. There are also U.S. leaders whose commitments run far ahead of recognized sales.

What we do not have yet is evidence that one general-purpose humanoid can economically replace human labor across dozens of unrelated jobs.

For now, the companies with the strongest business models are the ones getting paid for useful work while steadily making each new deployment cheaper than the last.