Figure AI was valued at $39 billion in September 2025 after raising more than $1 billion, yet the company still does not publish a price list for its humanoid robots or disclose what BMW pays to deploy them. At almost the opposite end of the industry, China’s Unitree will sell you a G1 humanoid directly from its online store for $13,500, and its IPO filings showed that humanoid robots generated 867.8 million yuan of revenue in 2025.
Both companies are participating in what investors routinely describe as the same humanoid-robotics boom. Economically, however, they are barely playing the same game.
That contradiction explains why the central question facing humanoid robotics is no longer whether the machines can walk, lift boxes or fold laundry. It is whether those capabilities can be converted into a repeatable business with attractive margins after manufacturing, deployment, maintenance, software, data collection and human support are properly accounted for.
There is no single answer yet. In fact, the most revealing thing about the sector in 2026 is how many different answers companies are testing.
Why Humanoid Robotics Doesn’t Have One Business Model Yet
Industrial automation traditionally has relatively understandable economics. A company buys an articulated robot arm, conveyor system or autonomous mobile robot because it expects the machine to increase throughput or reduce labour costs sufficiently to justify its purchase, installation and maintenance.
Humanoids complicate that calculation because companies are trying to sell two things simultaneously: a physical machine and an intelligence platform whose capabilities are supposed to improve over time.
That distinction creates several possible revenue models. A manufacturer can sell the robot outright, as Unitree increasingly does. It can retain ownership and charge recurring fees, as Agility Robotics does through Robots-as-a-Service agreements.
It can sell a consumer robot through subscriptions, as 1X plans to do with NEO. Or it can spend years subsidising hardware deployments because the data collected in factories and homes could eventually make its underlying AI system more valuable. The maturity of these approaches varies dramatically.
Agility Robotics disclosed in June 2026 that it had secured more than $300 million of multi-year orders for its next-generation Digit v5, although the company explicitly cautioned that these orders are subject to contractual milestones and are not current-period revenue.
Tesla, meanwhile, told investors in its 2025 annual filing that it had not yet commercialised Optimus and could not predict how demand would develop across commercial or consumer applications.
That gap between order books, pilots, revenue and future expectations is where much of the industry’s hype gets blurred.
To understand how these businesses might actually make money, it helps to start with the model investors already know best.
Unitree is Selling the Hardware Directly
$13,500 is an unusually important number in humanoid robotics. That is the advertised starting price of Unitree’s G1, a 1.32-metre humanoid available through the company’s official online store.
Unitree also lists its newer R1 from $4,900, the H2 at $29,900, and higher-end humanoids at substantially higher prices, making its commercial model look more like conventional electronics or industrial robotics than most Western humanoid startups.
More importantly, Unitree now has actual financial evidence behind the model. Its IPO prospectus showed revenue rising from 392.8 million yuan in 2024 to roughly 1.7 billion yuan in 2025, while humanoid-robot revenue reached 867.8 million yuan and overtook quadruped robots as the company’s largest product category.
The company shipped more than 5,500 humanoids in 2025, according to figures reported in its prospectus.
This is the cleanest version of the humanoid business model: manufacture robots, sell units, recognise hardware revenue and expand the installed base.
But it also exposes the limitations of using shipments as a proxy for commercial success.
Reuters reported in August 2026 that practical factory applications for Chinese humanoids remain limited and that government programmes and subsidies are helping stimulate demand across an industry containing more than 150 competing companies. Unitree itself has acknowledged that many meaningful commercial applications remain immature, while rising R&D spending has already pressured profitability.
So Unitree has demonstrated something much of the industry has not: customers will pay real money for humanoid hardware.
It has not yet proved that every robot sold generates enough productive economic value to create durable industrial demand without research budgets, subsidies or novelty-driven purchases.
That is precisely the problem another company is trying to solve with a very different pricing model.
Agility Robotics: Robots as a Subscription, Not a Purchase
If Unitree resembles an equipment manufacturer, Agility Robotics increasingly resembles a labour-capacity provider.
In June 2024, logistics company GXO signed a multi-year agreement to use Agility’s Digit robots at a facility handling Spanx operations near Atlanta. The companies described the deal as the industry’s first formal commercial humanoid deployment under a Robots-as-a-Service, or RaaS, model. Digit performs repetitive material-moving tasks, transferring totes between autonomous mobile robots and conveyor systems.
That distinction matters commercially because GXO does not need to make a large speculative capital purchase and then discover whether the robot works. Agility can spread the cost of hardware over the lifetime of the machine while charging recurring fees for access to the robot, software, support, and fleet-management capabilities.
The most revealing numbers arrived when Agility announced plans in June 2026 to go public through a merger valuing it at $2.5 billion pre-money.
Its investor presentation used an illustrative assumption of $8,500 per month per Digit under the RaaS model. The same presentation said more than $300 million of contracted Digit v5 orders represented 1,000 robots under three-year RaaS contracts, although the company stressed that the figure represents potential multi-year value rather than recognised revenue.
That gives investors something unusually rare in this sector: the beginnings of a unit-economic equation.
The customer does not need to ask whether a humanoid is intellectually impressive. It can compare approximately $8,500 per month against the fully burdened cost of labour, the hours Digit can operate, deployment costs, maintenance, downtime and productivity.
Agility has also moved beyond one customer. Toyota Motor Manufacturing Canada signed a RaaS agreement in February 2026 following a pilot, while Mercado Libre agreed to deploy Digit in fulfilment operations beginning in Texas. Agility says Digit has accumulated more than 65,000 operating hours and has active deployments with companies including GXO, Toyota, Schaeffler and Mercado Libre.
This is arguably the most conventional enterprise software logic being applied to the least conventional enterprise hardware: turn a large capital expense into a recurring operating expense and make the customer buy an outcome rather than a machine. 1X is attempting a consumer version of essentially the same idea.
Its NEO home robot can currently be reserved under two models: $20,000 for Early Access ownership or $499 per month through a subscription plan, with US deliveries scheduled to begin in 2026. That pricing allows 1X to test whether consumers think about a home humanoid more like an appliance they own or a service they continuously pay for.
The subscription model solves one problem, however, only by creating another. If the customer pays monthly, the robotics company retains far more of the hardware risk.
Figure AI: Betting on Data Before Data Becomes Revenue
Figure AI presents perhaps the clearest example of why analysing humanoid robotics purely as a hardware industry misses the deeper investment thesis.
Its Figure 02 robots completed an 11-month deployment at BMW’s Spartanburg plant, where Figure says they loaded more than 90,000 parts, operated for over 1,250 hours, and contributed to the production of more than 30,000 BMW X3 vehicles. Figure 03 returned to the plant in 2026 for more complex logistics work.
That is a meaningful deployment. What Figure does not publicly disclose is equally meaningful: the company has not published BMW contract economics, per-robot pricing, or a detailed breakdown of robot-generated revenue.
Instead, Figure increasingly talks about an entirely different asset that is data.
In August 2026, Figure unveiled Index, a crowdsourced system for collecting videos of people performing physical tasks. The company said contributors across 108 countries had already uploaded more than 16 million videos, with over 44,000 weekly active users, and that it had paid $15 million to contributors. Figure said it intends to spend more than $1 billion over the following 12 months on data and compute.
The videos train Helix, Figure’s embodied-AI system. This does not mean Figure currently has a data-licensing business. There is no disclosed standalone product through which outside companies buy access to Index, and describing it as current data revenue would be inaccurate.
The economic logic is more subtle. If more diverse training data makes Helix better at handling unfamiliar objects, homes, factories, and tasks, then every dollar spent building the dataset can potentially make every future Figure robot more useful.
More capable robots create more deployments, deployments create more real-world experience, and that experience can feed further model improvement. That is the physical-AI version of a data flywheel.
It helps explain how investors can assign Figure a $39 billion valuation even while its publicly visible commercial economics remain far less developed than its technology and deployment story. Figure’s September 2025 funding announcement said the capital would support BotQ manufacturing, Helix deployments and advanced data collection.
Investors are therefore not simply betting on how much one Figure 03 can sell for.
They are betting that the intelligence controlling millions of future machines could become substantially more valuable than the margin on any single robot.
But before that thesis works, somebody still has to manufacture those millions of machines.
Apptronik: Capital First, Pricing Later
Apptronik illustrates the enormous amount of capital required before humanoid economics become legible.
The Austin startup raised another $520 million in February 2026, taking its expanded Series A to more than $935 million and total capital raised to nearly $1 billion. Reuters reported the company at a valuation of roughly $5 billion, while TechCrunch separately reported approximately $5.3 billion; the difference is a useful reminder that private-company valuations are often reported from different sources and may not refer to exactly identical transaction measures.
Apptronik has serious industrial partners. Mercedes-Benz has worked with the company on manufacturing deployments; Jabil agreed both to manufacture Apollo robots and test them in its own production operations; GXO has partnered with Apptronik, and Google DeepMind is collaborating on embodied AI.
What Apptronik has not published is a standard Apollo selling price, subscription rate or detailed customer-revenue figure.
That makes the company’s current economics much harder to evaluate than Agility’s.
The strategy is therefore closer to capitalise first, industrialise second, optimise monetisation after deployments prove the product. Nearly $1 billion of external funding gives Apptronik the ability to improve hardware, AI, manufacturing capacity and field operations before requiring the business to finance its own expansion through robot revenue.
That approach is not unusual for frontier technology. Semiconductor fabs, launch companies and autonomous-vehicle developers have followed similar capital-intensive paths.
The risk is that scale has to arrive before the capital runs out.
Tesla is approaching the same problem from the opposite financial position. Rather than raising venture capital specifically for Optimus, it can finance humanoid development from a corporation that ended June 2026 with more than $43 billion in cash and short-term investments. Tesla’s own filings still say Optimus has not yet been commercialised, while the company is spending heavily on manufacturing capacity and AI infrastructure.
For Tesla, the first customer can effectively be Tesla itself. For a startup, somebody outside the company eventually has to pay.
What the Deployments Already Happening Tell Us
The most important humanoid deployments today share a striking characteristic: they are narrow.
Figure’s BMW robots loaded sheet metal and are moving into logistics sequencing. Agility’s Digit moves totes and materials in structured warehouses and factories. Apptronik’s partnerships centre on manufacturing and logistics rather than the fantasy of a machine that performs every human task.
That is commercially significant. The first viable humanoid business may not sell “general-purpose labour” at all. It may sell a succession of tightly defined workflows whose economics can be measured individually.
Agility’s progress illustrates the point particularly well. A warehouse operator can calculate how many totes Digit moves, how many hours it works, what labour costs it substitutes, how frequently it requires intervention, and whether the subscription produces an acceptable return. That is a much easier procurement decision than asking a CFO to approve a humanoid because artificial general intelligence might eventually make it useful.
Unitree represents the other end of the adoption curve: relatively accessible hardware gets robots into research institutions, developer teams, and commercial experiments quickly, generating sales before perfect autonomy exists.
1X is making an even more aggressive bet. By taking NEO into homes, it gains exposure to the messy, unstructured environments where general-purpose robots ultimately need to function. But its own product information acknowledges that difficult tasks can still involve scheduled remote supervision by a 1X expert.
That detail may be more important than another dancing-robot demonstration. It tells us that today’s humanoid business can contain hidden human labour inside what appears to be automation.
The Open Questions Nobody Has Answered Yet
The first unresolved question is whether the robot itself becomes cheap enough.
Unitree has already pushed headline hardware prices dramatically lower, while Agility’s investor model assumes recurring economics that can be comparable to human labour. But purchase price is only one part of total cost. Batteries, replacement components, technicians, remote operators, software infrastructure, integration work and downtime can turn an inexpensive robot into expensive labour.
The second question is utilisation. A robot costing $8,500 a month looks very different if it performs valuable work for 20 hours a day than if it works intermittently and requires frequent supervision.
Agility’s public investor materials model economics around high utilisation, but the wider industry does not yet have years of public fleet-level reliability data comparable with mature industrial automation.
Then comes liability. A conventional industrial robot usually operates inside a controlled environment. A humanoid is valuable precisely because it can operate in spaces designed for people.
That also means working near people, equipment and unpredictable objects, raising questions around safety certification, insurance responsibility and fault when an AI-controlled machine makes the wrong physical decision.
The hardest question, however, may be demand.
Reuters’ August 2026 examination of China’s humanoid industry found impressive progress in hardware but persistent difficulty performing useful factory work reliably. Government support has accelerated production, yet conventional automation frequently remains cheaper and better at specialised tasks.
Humanoids therefore have to justify the humanoid shape. If a wheeled robot or fixed arm can perform the same job more cheaply and reliably, looking human becomes a cost rather than an advantage.
The strongest economic argument for humanoids appears where businesses have thousands of facilities, tools and workflows designed around the human body and cannot afford to rebuild all of them for conventional automation. A machine that can climb steps, reach human-height shelves, move through existing aisles, and use existing workstations can potentially automate infrastructure that would otherwise remain stubbornly manual.
That is where the business model becomes genuinely interesting.
The company that wins humanoid robotics may not be the one that produces the most impressive robot demonstration or even the lowest-cost machine. It may be the one that discovers the right combination of hardware cost, recurring software, field service, proprietary training data, and measurable labour economics before everyone else.
Unitree has shown that humanoid hardware can produce meaningful revenue. Agility has shown that enterprises will sign multi-year service contracts. 1X is testing whether households will accept a subscription.
Figure is spending at an extraordinary scale to make physical-world data a moat. Apptronik is using nearly $1 billion of capital to reach industrial scale, while Tesla can subsidise the same race from an existing balance sheet.
None of those models has won yet.
But the direction is becoming clearer: humanoid robotics will stop being a technology story when customers stop asking what the robot can do and start asking how quickly the robot pays for itself.
Stay connected with Business Herald for the latest business news, insights, and updates.
Follow us on Facebook, Instagram, LinkedIn, and YouTube.
Join our growing community on WhatsApp and Telegram for real-time updates delivered directly to you.
