Two numbers crossed paths this month, and together they say more about the state of humanoid robotics than any demo video released this year.
The first number is 34,000. That is the count of preorders claimed by Humanoid, the London startup that became Europe’s first pure-play humanoid unicorn in July, for its wheeled HMND 01 industrial robot. Founder and CEO Artem Sokolov confirmed the figure to Forbes in early September, alongside a commercial pipeline he values at approximately $2.4 billion. The second number is 16,000. That is Berg Insight’s estimate for total humanoid robot shipments, worldwide, across every vendor, for all of 2026.
Sit with the arithmetic. One two-year-old company’s reservation list is more than double the delivered output of the entire global industry. If every preorder converted, Humanoid alone would need to multiply planet-wide production capacity by a factor of two just to clear its own backlog. It will not, of course, and Sokolov says so himself: “While these are not committed orders, they are strong indicators of customer interest and reflect the strongest pipeline in the market at this stage of development.”
That sentence, and the gap it politely describes, is the subject of this piece. The humanoid industry in late 2026 runs on two parallel ledgers: a demand ledger written in reservations, and a supply ledger written in shipments. Understanding how each is constructed, and where they are likely to converge, is the difference between reading this market clearly and getting swept up in it.
Inside the $2.4 Billion Pipeline
Start with what Humanoid has actually built, because the company is more substantive than the preorder headline suggests.
Founded in May 2024 by Sokolov, who has said his motivation came from personal experience watching repetitive industrial work grind people down, the company now employs more than 250 people and has raised roughly $270 million. The July Series A of $152 million, led by Prime Movers Lab, landed at a $1.35 billion post-money valuation. The commercial architecture around the machine is unusually mature for a company that has not delivered a single revenue unit: Robert Bosch Robotics is the contract manufacturer for Europe, with several hundred robots planned for 2027, and Schaeffler occupies three roles at once as investor, actuator supplier, and largest publicly announced customer, with an offtake agreement targeting a four-digit number of wheeled robots deployed across its facilities by 2032. Pilots have run with Bosch, Siemens, SAP, and Ford.
The product itself is a deliberate bet against bipedal romanticism. The HMND 01 pairs a torso, head, and two arms with a wheeled base, on the argument that this configuration addresses roughly 90% of industrial use cases while sidestepping the cost and reliability burden of legs. Current Alpha units carry about 10 kilograms total, five per arm. The Beta model, targeted for Q4 2026, is designed to double that. The company says it built the wheeled platform in seven months and a bipedal variant in five, which tells you the mechanical iteration loop is fast even if the commercial one is not.
On the software side, Humanoid develops its own stack, a four-layer framework called KinetIQ, including a reinforcement learning system dubbed KinetIQ Ascend. The company claims some trained tasks run at roughly 1.5 times human speed with success rates as high as 99%. Two caveats belong next to those numbers: they are vendor-supplied, they apply to selected tasks, and there is no independent benchmark or sustained fleet data behind them. Claims like these are currency in this industry, and everyone quotes them the same way, with an asterisk that is sometimes visible.
The pricing model spans both purchase and Robot-as-a-Service contracts, with a stated customer payback target of about one year and remote teleoperators available to intervene when autonomy runs out of road. The 2030 production target is 150,000 units annually across multiple manufacturing partners.
Now run the sanity check. Several hundred units in 2027, growing to 150,000 in 2030, is not a ramp, it is a vertical line. It requires roughly a 100x scale-up in three years across a supply chain that, per Berg Insight, does not yet produce 100,000 humanoids for all vendors combined. The preorder number is doing a lot of narrative work in that plan.
How a Preorder Pipeline Is Built
It is worth being precise about what a preorder is, because the term sounds more binding than it is.
A reservation is an option, not an obligation. The customer, typically an industrial firm with a labor problem and a innovation budget, signs a nonbinding expression of interest, sometimes with a nominal deposit, and secures a place in a delivery queue that does not yet exist. The customer pays almost nothing, risks almost nothing, and gains early access to a technology that might matter. The vendor books the notional contract value in a pipeline figure, which then anchors valuation conversations with investors. Both sides get something. Neither side has bought or sold a robot.
This is the same instrument Tesla used for the Cybertruck and dozens of EV startups used before folding with five-figure reservation lists. In consumer hardware it is a marketing tactic. In industrial robotics it has one genuine virtue: a reservation list with names like Schaeffler attached is a real signal about which way sophisticated industrial buyers are leaning, even if it is not a bankable order book. Schaeffler, a tier-one supplier that makes the actuators going into these machines, effectively hedged its own production line by reserving robots from a customer it also finances. That is not nothing. It is just not revenue.
The industry’s reliance on this metric is understandable, because the alternatives are thin. Counterpoint Research data reported this month shows that large-scale repeat orders remain rare across the sector. Deployments to date are a mix of purchases, preorders, and pilots. When audited repeat revenue does not exist at scale, pipeline is the best number available, and everyone has learned to quote it with the qualifier Sokolov used: strong indicator, not committed.
The Other Ledger: Who Actually Ships Robots
The supply-side picture is dominated by a fact that Western coverage keeps relearning: five Chinese companies shipped 86% of the world’s humanoid robots in the first half of 2026, according to Counterpoint Research.
The drivers are structural rather than mysterious. The electric vehicle supply chain provides batteries, sensors, harmonic drives, and actuators locally, letting Chinese robot makers iterate faster and cut prices aggressively. Capital is abundant, with XPeng’s robotics unit raising more than $900 million in August at a post-money valuation above $6.3 billion, and Galbot collecting RMB 2.5 billion, about $362 million, in March after a late-2025 round that valued it near $3 billion. And the deployment surface is enormous: automotive plants, electronics lines, logistics hubs, aerospace, and energy facilities all serving as testing grounds. Selina Xu, China and AI policy lead in Eric Schmidt’s office, put it plainly: Chinese firms are moving faster and at greater scale than U.S. rivals at this stage.
But the same reporting carries a counterweight worth engraving. Commercialization remains early everywhere. Deployments skew toward trials rather than repeat purchases. Jiang Han, a senior researcher at the Pangoal Institution, argues that repeat orders and customer payback periods are the real tests of whether robots solve problems or merely attract visits from the innovation team. And Fu Sheng, chairman of Cheetah Mobile and OrionStar, frames the reliability barrier as “the last 1%”: a 99% success rate still means one failure in every hundred attempts, and an industrial customer who budgets around a human worker does not tolerate a robot that fumbles one box in every hundred.
That is the ledger nobody’s preorder number addresses. The 86% shipment share is real output, but it is output measured in the low thousands of units globally, delivered mostly into pilot programs. The West’s answer, in the form of companies like Humanoid, Figure, Agility, and Apptronik, is to concede early shipment volume and compete on deployment quality, software, and integration depth in high-value facilities. Both strategies quote numbers that flatter them. China quotes shipments. The West quotes pipelines.
The Long Forecast: 26 Million Units and $554 Billion
Berg Insight’s new report supplies the third number every roadmap in this industry quietly depends on. The firm forecasts annual humanoid shipments rising from 16,000 units in 2026 to 26 million by 2040, a compound annual growth rate of 63.7%, with market value climbing from $890 million to $554 billion over the same span.
Some texture inside the projection: more than 100 companies are now developing full-size humanoids, with Figure AI, Tesla, Agility Robotics, Boston Dynamics, Apptronik, 1X Technologies, NEURA Robotics, Unitree, UBTECH, AgiBot, Leju, EngineAI, and RobotEra named as the current market shapers. Investment accelerated through 2025 and 2026, with NVIDIA, Microsoft, Google, Amazon, Hyundai, and OpenAI all taking strategic positions. And in the detail most relevant to the industrial buyers filling out preorder forms, Berg Insight expects cellular connections tied to humanoid robots to grow from 5,000 in 2025 to 50.6 million in 2040, an 84.9% CAGR, because a deployed fleet is not just an autonomous machine but a managed endpoint requiring remote monitoring, diagnostics, software updates, and lifecycle services.
A 63.7% CAGR sounds like analyst exuberance until you notice what it demands: the industry must roughly double shipments every 19 months for 14 consecutive years, through recessions, tariff regimes, safety incidents, and at least one promised actuator or battery shortage. There is precedent for this shape of curve, in smartphones and in EVs, and in both cases the inflection arrived years after the first credible prototypes, triggered by a combination of unit cost, reliability, and a use case that needed no persuasion. Harry Mellsop, co-founder of the simulation startup Antioch, has described physical AI as being in its “GPT-2 era,” meaning the raw capability is visible, the scaling laws are plausible, and the data and compute needed to finish the job do not exist yet at the required scale.
Forecasts like Berg’s are best read not as predictions but as consistency checks. They tell you what the world must look like if every vendor roadmap, including Humanoid’s 150,000-unit 2030 target, approximately holds. The 2040 number is a destination, not a schedule.
What Converts a Pipeline Into a Business
If preorders are options on a future, the interesting question is what makes them expire valuable rather than worthless. Four gates matter, and the next 18 months will test all of them.
First, hardware delivery on schedule. Humanoid’s Beta lands in Q4 2026 with a promised payload doubling to 20 kilograms. A slip of two quarters is normal in robotics and forgivable; a slip of four starts converting nonbinding reservations into nonbinding cancellations.
Second, Bosch’s 2027 production run. Several hundred units through a contract manufacturer is the moment the company’s engineering claims meet manufacturing yield, supply chain reality, and field failure rates simultaneously. Every humanoid company that has scaled past prototypes has discovered that the factory is a harder problem than the robot.
Third, the Schaeffler conversion. Of the 34,000 reservations, the four-digit Schaeffler offtake through 2032 is the only block backed by an investor-supplier-customer relationship with real money on three sides of it. When Schaeffler moves from pilots to purchase orders, the pipeline gains its first anchor in audited reality. If Schaeffler stalls, the $2.4 billion figure should be discounted accordingly.
Fourth, and hardest, the payback math. Humanoid targets roughly one year for customer payback, which would make the RaaS contract an easy yes for any plant manager. That target assumes uptime, task success rates, and intervention costs that no vendor has yet sustained across a fleet in production. Fu Sheng’s last-1% problem lives here: 99% task reliability at 1.5 times human speed is a compelling spreadsheet, and 97% reliability at human speed is a machine that gets unplugged and shelved.
The Verdict
The correct reading of this month’s two headline numbers is not cynicism and it is not celebration. The 34,000 preorders are not fake; they are evidence that industrial buyers, including sophisticated ones who build actuators for a living, believe the wheeled-humanoid value proposition enough to reserve a place in line at near-zero cost. The 16,000 shipments are not a failure; they are what year zero of a manufacturing industry looks like when the product category is three years into commercial life.
What the numbers expose is a market running on promissory notes in both directions: Chinese vendors shipping thousands into pilots that have not yet produced repeat orders, and Western vendors accumulating reservations against production lines that do not yet exist at scale. The convergence event, when pipelines meet factories and factories meet payback periods, is scheduled roughly between Q4 2026 and the end of 2027. That is when the two ledgers start merging into one, and when the number that finally matters, delivered units generating repeat revenue, begins to exist.
Until then, treat every pipeline figure the way Sokolov asks you to treat his own: a strong indicator of interest, in a market where interest is abundant, orders are forming, and proof is still on the truck.
Sources: Forbes, Berg Insight via IoT Business News, Counterpoint Research via Inside AI, Humanoid.guide