Are humanoid robots the new bubble?

Why does humanoid robotics look bubbly right now?

Humanoid robotics looks bubbly because investment and promises have grown several times faster than the commercial market.

Bank of America estimates that annual investment in humanoid robotics rose from about $700 million in 2018 to $4.3 billion in 2025. By early 2026, more than 50 companies were developing their own humanoids and roughly 150 commercial models had been announced. Annual investment has therefore risen sixfold before the industry has found a repeatable mass-market use case.

The revenue base is still small. IDC estimates that about 18,000 humanoids shipped worldwide in 2025, producing roughly $440 million in sales. Shipments increased by more than 500% in one year, so the market is clearly moving. The awkward part is where those robots went: IDC says more than 85% were used for performances, education, data collection, guided tours and technology validation. Factory and logistics pilots were still a minority.

This mix explains the bubble talk better than any dramatic robot video. Capital is pricing an industrial workforce while most current shipments are still helping the industry learn how to build, train and demonstrate the machines.

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

What would a humanoid robot bubble actually look like?

A humanoid robot bubble would mean investors are paying today for adoption that may take another decade or longer to arrive.

Useful technology and inflated prices can exist together. Railways, telecom networks and the early internet created lasting value, yet many investors lost money because too many companies raised too much capital at the wrong prices. Humanoids could follow the same path.

We would call the market bubbly while three gaps remain wide: company values stay disconnected from current sales, factory announcements keep exceeding delivered robots, and customers continue running pilots without expanding them into large fleets.

The question goes beyond whether humanoids can move boxes, load parts or fold laundry in a controlled setup. Several already can. What counts is whether those abilities create enough reliable, affordable work to support the money now flowing into the sector.

Is investment outrunning real humanoid robot sales?

Yes, investment is currently outrunning humanoid robot sales by an extreme margin.

The $4.3 billion invested in 2025 was almost ten times IDC’s estimate of $440 million in worldwide humanoid revenue. Young hardware industries often need heavy funding before revenue arrives, but this gap shows that investors are buying a future market rather than expanding a proven one.

The same enthusiasm has spread into robot intelligence. Skild AI raised about $1.4 billion at a valuation above $14 billion for a foundation model designed to control many kinds of robots. Apptronik raised more than $935 million across its Series A and extension rounds while developing Apollo. Figure raised more than $1 billion at a $39 billion valuation.

These bets depend on each other. Robot makers need better AI to make their hardware flexible. Robotics AI companies need large robot fleets to generate training data and paying customers. For now, both sides are being valued as though the other side will scale successfully.

Are humanoid robot valuations already too high?

Yes, several leading humanoid robot valuations already assume a level of dominance that no company has earned through revenue or fleet size.

Figure’s $39 billion valuation is the clearest case. Apptronik has been valued above $5.5 billion, while Agility Robotics agreed to a proposed public-market transaction at a $2.5 billion pre-money value. Together, those three companies are worth about $47 billion.

That combined figure is roughly 107 times the entire industry’s estimated 2025 revenue. Company valuations naturally include future software, services and intellectual property, while the revenue estimate covers current robot sales. Still, the scale is hard to ignore. Investors are already pricing several eventual winners before the market has shown how many winners it can support.

Agility carries the easiest valuation to defend because it has named customers, tens of thousands of operating hours and conditional multi-year orders. Figure has stronger AI ambition and a far richer valuation. Apptronik has serious industrial partners, although it publishes less fleet-performance data. The differences in evidence are much smaller than the differences in valuation.

Company	Latest disclosed or proposed value	Evidence visible today
Figure	$39 billion	BMW production work, a growing production line and a new logistics deployment
Apptronik	More than $5.5 billion	Large funding base and partnerships with Mercedes-Benz, Google and other industrial groups
Agility Robotics	$2.5 billion pre-money	Active customer deployments and a conditional multi-year order book
Combined	About $47 billion	Roughly 107 times estimated global humanoid revenue in 2025

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

Are humanoid robots doing useful work today?

Yes, a small number of humanoid robots are already doing useful, measurable work in factories and warehouses.

Figure’s 11-month deployment at BMW’s Spartanburg plant remains the clearest public example. Figure 02 robots loaded more than 90,000 sheet-metal parts, logged over 1,250 operating hours and contributed to the production of more than 30,000 BMW X3 vehicles. BMW has since brought Figure 03 into the same plant for a more complex logistics-sequencing task. The customer chose another workflow after the first project ended, which is stronger evidence than a one-off pilot.

Agility Robotics reports more than 65,000 operating hours across commitments involving nine facilities, including GXO, Schaeffler, Toyota Motor Manufacturing Canada and Mercado Libre. Digit has also moved more than 100,000 totes in a live commercial operation. Unitree offers a different kind of proof: it says more than 5,500 humanoids were sold and delivered to end customers during 2025.

We should still separate industrial work from general shipment volume. Unitree’s total includes research, education and entertainment buyers. Figure and Agility provide stronger evidence that humanoids can create value inside an operating business.

Company	Best public evidence	What we can reasonably conclude
Figure	90,000+ parts, 1,250+ hours and 30,000+ cars at BMW	A humanoid can sustain one repetitive factory workflow over many months
Agility Robotics	65,000+ operating hours and 100,000+ totes moved	Digit can perform repeated material-handling work across several customers
Unitree	Thousands of humanoids delivered in one year	Humanoid hardware can already be manufactured and sold at meaningful volume
BMW’s follow-on project	Figure 03 assigned to logistics sequencing	At least one major customer saw enough value to test a second workflow

Is that work anywhere close to mass adoption?

No. Current humanoid deployments remain tiny beside the automation market they hope to enter.

IDC’s 18,000 estimated humanoid shipments sound impressive until we compare them with mature robot categories. The International Federation of Robotics recorded 542,000 conventional industrial robot installations in 2024 and almost 200,000 professional service robots sold. Transportation and logistics alone took about 102,900 service robots, mostly wheeled systems built for moving goods.

Humanoids therefore represented only a few percent of annual industrial robot installations. Their share of productive industrial work was smaller again because most humanoid shipments went to demonstrations, education and data collection.

The strongest deployments today involve a handful of machines doing narrow jobs at selected sites. Mass adoption would look very different: hundreds of robots at one customer, repeat orders across several plants and the same fleet handling multiple workflows. No supplier has publicly shown that pattern yet.

How autonomous are humanoid robots now?

Humanoid robots now show real autonomy in controlled workflows. Broad, unsupervised autonomy remains out of reach.

Agility’s Digit can autonomously move totes inside mapped industrial environments. Figure’s Helix models can turn visual and language instructions into full-body actions. These systems have moved beyond fixed industrial programming because they can learn behaviors from demonstrations and adapt within a task.

The limits become clearer in homes and unfamiliar spaces. 1X tells early NEO buyers that the robot will arrive with basic autonomy. When NEO encounters a difficult task, a remote 1X expert can supervise its actions at a scheduled time. That approach gives customers a useful service and gives 1X valuable training data, but some of the apparent intelligence still comes from human support.

The number we need is intervention rate. Companies rarely disclose how often a person steps in, how long recovery takes or how many robots one operator can supervise. Until those figures become public, videos will keep making autonomy look broader than it really is.

Does the humanoid shape really make sense?

The humanoid shape makes sense for certain human-built spaces, although many jobs will remain cheaper with wheels, fixed arms or simpler grippers.

Factories, warehouses and homes contain stairs, doors, shelves, workstations and tools designed around human height and reach. A mobile robot with two arms can use that infrastructure without forcing the customer to rebuild the site. That compatibility is the strongest economic argument for the shape.

It also creates expensive engineering problems. Legs consume energy and can fall. Dexterous hands add motors, sensors and failure points. A tall machine is less stable than a low wheeled platform. Human anatomy offers flexibility; industrial machines are usually optimized for speed, strength and reliability.

Agility’s Digit shows how practical designs may evolve. It has birdlike legs and gripper-style hands rather than a realistic human body. The company designed around the work environment instead of chasing visual imitation. More robots will probably follow that logic: humanoid where compatibility helps and deliberately machine-like everywhere else.

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

Are humanoid robots becoming cheap enough to pay for themselves?

Humanoid robots are getting cheaper quickly, but the industry still lacks enough public operating data to prove broad payback.

Bank of America estimates that pilot-stage humanoids have often cost around $90,000 to $100,000. It also estimates that the bill of materials for a Chinese humanoid fell to about $35,000 in 2025 and could drop below $17,000 by 2030. 1X now offers NEO for $20,000 or $499 a month, although early units include basic autonomy and access to remote expert support.

Hardware price tells only part of the story. A customer also pays for integration, software, charging, maintenance, spare parts, supervision and downtime. A $30,000 robot that works half the shift and regularly needs an operator can cost more than a $100,000 robot with high uptime.

Falling prices will expand the market while squeezing manufacturers. Unitree and other Chinese suppliers can push hardware toward commodity economics. Western startups valued like software platforms may struggle if customers treat the body as interchangeable and pay mainly for service, reliability and task software.

The first convincing payback cases will probably appear in expensive labor markets, hazardous jobs and long-shift operations. Low-wage sites with unpredictable work will be much harder.

Why use a humanoid instead of a simpler robot?

A humanoid earns its place only when one flexible machine can handle work that would otherwise require several specialized systems or a costly redesign of the site.

Conventional robot arms dominate stable production lines because they are fast, precise and dependable. Wheeled autonomous mobile robots already move pallets, shelves and totes through flat facilities. Those machines have years of field data and established maintenance networks.

Humanoids are strongest in the gaps between those systems. At BMW, Figure carried parts from human-oriented racks into a process served by traditional robot arms. Agility describes Digit as a way to connect separate islands of automation. The robot’s value comes from moving through a human workspace and handling material at human height.

This gives customers a straightforward test: how many different workflows can the same machine perform after a software update or a short retraining period? A humanoid that spends years repeating one motion will usually lose to purpose-built equipment. A robot that changes jobs across shifts has a much better case.

Can manufacturers actually build robots at the scale they promise?

Some humanoid manufacturers have crossed into real production, but promised capacity still exceeds demonstrated deliveries by a wide margin.

Unitree has the strongest volume evidence, with more than 6,500 humanoids manufactured and more than 5,500 delivered during 2025. Figure says its BotQ line produced more than 350 Figure 03 robots and raised throughput from one robot per day to one per hour in less than four months. Those are meaningful manufacturing gains.

The larger targets remain unproven. Agility designed RoboFab for up to 10,000 Digits a year. 1X says its Hayward factory can produce 10,000 NEOs annually and wants capacity above 100,000 by the end of 2027. It also says hundreds of NEO hands have already left a scalable line, with capacity for 10,000 hands this year. XPeng is preparing a full-chain production base and says it wants mass production of IRON before year-end.

Capacity, output and customer deliveries are three different numbers. A working factory solves only the first problem. Suppliers still need stable components, acceptable failure rates and buyers willing to take the machines in volume.

Is China building a real humanoid market or another oversupply problem?

China is building a real humanoid industry, and the same push could create a severe oversupply cycle.

The evidence for a serious manufacturing base is strong. IDC says Chinese vendors dominated global humanoid shipments in 2025. Unitree and AgiBot each shipped roughly 5,000 units, while several other Chinese suppliers reached the thousand-unit range. China’s advantage comes from its existing motors, batteries, electronics, machining and industrial automation supply chains.

Government policy is accelerating the buildout. The Ministry of Industry and Information Technology has called for mass production, competitive national champions and industrial clusters. XPeng, a carmaker with its own AI chips and manufacturing experience, is building a dedicated humanoid production base. The push now extends far beyond a collection of robotics startups.

The risk is familiar. Subsidies and local competition can fill factories before profitable demand exists. IDC’s finding that more than 85% of 2025 deployments went to demonstrations, education, data collection and guided tours suggests shipment growth is still ahead of labor replacement.

China will probably lower global robot prices and improve the supply chain. It will also produce plenty of companies that never earn attractive returns.

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

Is humanoid robot intelligence improving fast enough?

Humanoid robot intelligence is improving fast enough to unlock more industrial tasks, but the path to a general-purpose worker still looks long.

Vision-language-action models have changed what robots can learn. Figure’s Helix system links visual perception, language instructions and continuous movement. Its newer full-body models can reposition the robot while it manipulates objects. 1X trains generalist models with simulation, robot experience and remote human demonstrations. Skild AI is trying to build one foundation model that can control different robot bodies.

Physical data remains the bottleneck. Language models learned from enormous stores of text that already existed online. Robots need video, joint positions, force, contact, failures and successful recovery attempts. Gathering that data requires machines operating in the real world, often with humans supervising them.

This creates a compounding advantage between deployment and intelligence. Companies with larger fleets can collect more useful data, improve their models and make the next fleet more capable. The race may eventually concentrate around a few companies with enough hardware, customers and compute to maintain that loop.

Broader skills inside structured factories should arrive well before anything close to open-ended household labor.

Are humanoids safe and reliable enough for crowded workplaces?

Humanoids still lack the safety record and standards needed for unrestricted work beside people.

Sustained deployments are beginning to expose real hardware problems. During Figure’s BMW project, the company identified the forearm as its most frequent hardware failure point. Agility has passed an OSHA-recognized field inspection at a customer site, which is meaningful progress, while its upcoming Digit V5 is still being developed around cooperative safety and onboard human detection.

The standards are also catching up. ISO 25785-1, covering safety requirements for dynamically stable industrial mobile robots such as legged machines, remains under development. Factories can manage the gap with controlled zones, conservative speeds and tightly defined workflows. Homes, hospitals and public spaces are a much harder problem.

Reliability could become the industry’s least glamorous bottleneck. A robot with dozens of joints, cameras, cables and force-sensitive parts may work beautifully in a demonstration and still become expensive to maintain across thousands of daily cycles. We need fleet uptime, repair frequency and incident data before broad safety claims deserve confidence.

Are customers placing real orders or just running pilots?

Customers are moving beyond curiosity, although most humanoid demand still comes with conditions attached.

Agility offers the clearest order evidence. The company reports more than $300 million in multi-year orders for Digit V5, subject to contractual milestones, and a pipeline of more than 30 customers. It has also moved forward with a proposed $2.5 billion SPAC merger and filed a draft registration statement. Public-market scrutiny should eventually reveal far more about revenue, margins and order conversion.

Figure has partnerships with BMW, Brookfield and Catalyst Brands. BMW’s decision to assign Figure 03 a new logistics task is particularly useful because it follows an 11-month production deployment. Apptronik has agreements with Mercedes-Benz, GXO and other major companies, though detailed fleet results remain limited.

The decisive step is expansion. We want to see customers move from one workflow to several, from one building to multiple sites and from an innovation budget to normal capital spending. Milestone-based orders show serious interest. Delivered fleets and renewals show a functioning market.

Can we trust the huge humanoid robot market forecasts?

The largest humanoid robot forecasts are plausible as distant scenarios and weak as evidence for today’s prices.

Goldman Sachs projects a $38 billion market by 2035. Bank of America expects annual shipments to reach 10 million by that point. Morgan Stanley sees a market above $5 trillion by 2050, while Citi has modeled about 650 million humanoids and $7 trillion in value by the same year.

The differences reveal how sensitive these forecasts are. A small change in adoption timing, useful life, robot price or service revenue becomes enormous when projected across 25 years. Morgan Stanley also expects adoption to accelerate mainly in the late 2030s. That leaves a long period in which current startups must keep raising money, improving reliability and surviving price competition.

The forecasts help illustrate the possible size of robotic labor. They do not prove that today’s manufacturers deserve tens of billions in value.

Forecaster	Headline projection	Assumption doing most of the work
Goldman Sachs	$38 billion market by 2035	Industrial costs fall enough for about 1.4 million annual units
Bank of America	10 million annual shipments by 2035	Shipment growth remains extremely high for a full decade
Morgan Stanley	More than $5 trillion by 2050	Adoption accelerates sharply in the late 2030s and 2040s
Citi	Around 650 million humanoids and $7 trillion by 2050	Robots become a common global labor platform

What could burst the humanoid robot boom?

The boom will weaken quickly if customers keep testing humanoids without buying large fleets.

The first danger is a pilot trap. Manufacturers can accumulate famous logos while deploying only a few robots at each site. That produces publicity and training data, but it cannot support large factories or premium valuations.

Reliability could cause the next shock. Frequent repairs, short battery life or slow recovery from errors would erase labor savings. A serious workplace accident could slow deployments and increase insurance costs across the whole category.

Price competition may hurt investors even while helping adoption. Chinese suppliers can drive hardware prices down faster than Western startups reduce costs. The market could grow rapidly while hardware margins collapse.

A slower AI curve would be equally damaging. If every new workflow requires months of engineering, large teleoperation teams and site-specific integration, humanoids will look like a difficult automation business. Many valuations currently assume something closer to a scalable software platform.

What would prove that the humanoid robot market is real?

The market will look real when customers publish fleet economics instead of suppliers publishing isolated demonstrations.

We would expect hundreds of robots at individual companies, followed by repeat orders across several sites. The same robot should handle multiple workflows with limited retraining. Customers should be able to compare the cost per completed task with human labor and simpler automation.

The missing operating metrics are clear: autonomous hours, intervention frequency, uptime, repair cost, energy use, task-completion rate and safety incidents. Revenue quality also matters. Recognized sales, recurring software income, renewals and positive gross margins would tell us much more than factory capacity.

BMW’s second Figure deployment and Agility’s conditional order book are early steps in that direction. They remain well short of the evidence needed to support a mature industry. Once buying a humanoid becomes an ordinary operational decision rather than an innovation announcement, the bubble argument will lose much of its force.

Are humanoid robots the new bubble?

Yes, humanoid robots are the new bubble in finance, while the underlying technology has become real enough to survive the eventual correction.

Finance has run far ahead. Annual investment reached about $4.3 billion against roughly $440 million in industry revenue. Three prominent manufacturers carry a combined value near $47 billion. Long-range forecasts now reach trillions of dollars even though productive deployments remain small and most shipments still serve demonstrations, education or data collection.

Engineering has caught up more than critics admit. Figure completed sustained production work at BMW and has returned with Figure 03. Agility has moved from trials into substantial operating hours and conditional orders. Unitree has delivered thousands of robots. Figure and 1X have also moved manufacturing from prototypes toward repeatable lines.

Humanoid robots should become an important automation category, but today’s capital assumes too much success, too soon, across too many companies. Adoption will probably take longer than promoters suggest. Hardware prices will fall faster than many business plans expect. Most manufacturers will struggle to capture the value now implied by their funding rounds.

The robots are likely to stay. A large share of the current valuations will collapse.

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

Who is the author of this content?

NEW MARKET PITCH TEAM

We track new markets so founders and investors can move faster

We build living "market pitch" documents for emerging markets: AI, synthetic biology, new proteins, and more. Instead of outdated PDFs or hallucinated LLM answers, our clients get a clean, visual, always-updated view of what's really happening: key players, deals, regulations, and signals that matter. Learn more about us.