Last month, Beijing hosted the second World Humanoid Robot Games, and one of the videos that went viral in the days before showed a humanoid robot on a practice sprint reaching full speed, failing to stop in time and hitting the padded wall at the end of the track as sparks flew on impact.
When a speaker at AI Summit Barcelona brought the games up on Tuesday, that was the first image that came to my mind. It looked like entertainment more than engineering, and I suspect the organisers did not mind, because a crash with sparks gives a sporting event its highlight while a clean stop goes unnoticed. Similarly, the sparks thrown when a Formula 1 car’s floor scrapes the track make the best television, although they say nothing about who is fastest. Performative versus performance?
But showbiz is a serious matter, and so is the business of robots.
A London robotics company called Humanoid (an apt name, if not a very imaginative one) says it has received more than 34,000 preorders for its robot, which it values at about $2.4bn; its chief executive Artem Sokolov has acknowledged that these are non-binding, not committed orders. For comparison, global humanoid robot shipments were somewhere between 13,000 and 20,000 in 2025, depending on who is counting, which means the UK start-up’s reservation list is equal to roughly two years of recent global output. Mr Sokolov told Reuters in May 2026 that deliveries were planned over three years, which on our arithmetic means about 11,000 robots a year… more than half of last year’s global volume, from a company that plans to manufacture several hundred robots with Bosch in 2027.
Pushing further down the road, Bank of America expects about 1.2m humanoids to be shipped every year by 2030. Getting there from 2025’s roughly 20,000 means the yearly volume more than doubles, for five years in a row. That is not all… the International Federation of Robotics said today that factories installed more than 600,000 industrial robots in 2025 and forecasts 806,000 in 2029, so the humanoid forecast alone would overtake a total market that has been building up for 50 years.
This week we look at the gap between buying a robot and putting it to actual work… read on!
Buy now, work later
Unitree, the Chinese robot manufacturer, cut the price of its R1 humanoid in June 2026 from 39,900 yuan to 29,900 yuan (about $4,000), while its G1 model is priced at 99,000 yuan (about $13,600) in China and $16,000 in the US. But the G1 is sold for research, teaching and demo purposes: it can carry 3kg in each arm and operates for about two hours on a full charge.
Specifications move the price up quickly: Unitree’s larger H1 listed at around $90,000 in 2024, more than five times the G1. The cost of a humanoid that does useful work at a particular site is a further step up, because it includes everything needed to make the machine productive there: integration with the systems around it, safety work and support. Most of that is human labour paid at local rates, so theoretically, the same robot should cost more to put to work in Stuttgart than in Shenzhen.
Humanoid’s own numbers hint at the customer’s business case, although the company has described its $2.4bn in two ways. In May, Mr Sokolov told Reuters it represented “future annual recurring revenue”; in September, Forbes called it a pipeline and reported that Humanoid offers both direct sales and robots as a service, aiming for a one-year payback for the customer. Divided across 34,000 robots, the figure comes to about $70,000 each on our arithmetic, which is closer to a Unitree H1’s list price if it is an outright sale and a steep yearly fee if it is a subscription. The one-year payback is the more telling number, because it is the promise the customers are more likely to test.
Two measures will show whether that promise holds: the hours each robot works in a day, allowing for the time lost to charging, which docking stations and battery swaps can shorten but not remove, and the cost of getting the robots working at each new site. For Humanoid itself, a third number matters more than either: how many of its 34,000 preorders turn into actual paid contracts.
The factory precedent
Industrial robots are the nearest precedent. The operational stock reached 5m in 2025, up 9 per cent year on year, and annual installations have stayed above 500,000 since 2021; that is a healthy market, although it is not an exponential one. The statistics typically count manipulator arms, bolted to the floor or mounted on mobile bases, so humanoids are not yet a meaningful part of them.
Our premise is that demand for robots is a function of their total cost of ownership: the machine, plus the engineering needed to make it work on a particular line, plus maintenance over its life, set against the labour it replaces. Hardware keeps getting cheaper, but engineering a site is bespoke work that gains little from economies of scale, and is incurred again at every production line.
That cost is not fixed. It drops when the robot arrives ready to run out of the box, through pre-engineered work cells, robots that learn a task by being shown how to do it and machines safe enough to work without fencing. If our premise holds for humanoids, the 2030 forecast depends less on how many robots Chinese factories can build than on how close to plug and play the robots become, as that can validate the business case at almost any site, while robots that need a whole setup only make sense in the largest sites.
Two legs or four wheels
Going back to those racing robots, whether working robots end up looking human, with two legs and two arms, remains an open question, and Tesla is betting on both answers at once. Elon Musk told TED’s Chris Anderson in 2022 that humanity had designed the world for “a bipedal humanoid with two arms and ten fingers”, so a robot that fits in, such as Tesla’s Optimus, must be roughly the same size and shape; the same company’s Cybercab, its driverless taxi, has no steering wheel or pedals, and it does not look like a horse either.
The case for the human shape starts with the buildings. Architects have measured their work against the human body at least since Vitruvius, who set the proportions of temples by those of a well-formed man, and Le Corbusier built his Modulor system of proportions around a man 1.83m tall; stairs, door handles, worktops and shelves are all sized for us. A robot with legs, hands and a human reach could in principle work anywhere a person can, without anyone rebuilding the site.
The argument also runs the other way. The human body was shaped over millions of years of evolution for a world with no stairs in it, and the stairs came later, built to fit the body. Many factories and warehouses have since been rebuilt once already, around forklifts, pallets and conveyors, with flat floors and wide aisles, and on those floors wheels are cheaper and more stable than legs.
Both are probably true at once: homes and hospitals were built for people, while many factories were rebuilt for machines. The humanoid may prove the right shape for some settings, one shape among several, or a demonstration and entertainment format that stays in showrooms and never reaches production; in any case, too few of either design have been deployed for a solid comparison.
Thinking on board
AI inference, the step in which a trained model turns what it sees into an answer or an action, faces its tightest deadline in a robot. In the 100-metre semi-final of the Robot Olympics, a humanoid called Tiangong Ultra finished in 8.86 seconds, faster than the 9.58 seconds Usain Bolt has held as the world record since 2009. This works out at about 11 metres a second, so if a decision had to travel to a distant data centre and back, even in a tenth of a second, the robot would have covered a metre before the answer came back.
Speed is one reason to keep intelligence on board, and there are others. Warehouse and home connections drop out, and a robot cannot stand still every time the Wi-Fi does; its camera feeds are large and, when it works inside a home or a factory, full of private and sensitive material, which makes sending them away by default expensive, unwelcome and, in some jurisdictions, restricted by law.
Humanoid’s HMND 01 Alpha, for one, runs its AI models on Nvidia’s Jetson Thor, whose production module listed at $2,999 when Nvidia launched it in August 2025, three-quarters of the price of a whole Unitree R1. At that price, Bank of America’s 1.2m robots a year would need about $3.6bn of on-board AI chips a year, less than four days of Nvidia’s revenue, which reached $96.2bn in the quarter to July 26, 2026, and half of the $7.2bn its edge computing business, which includes robotics, made in that same quarter.
The chip also has to be powered. Nvidia’s developer kit for Thor runs within a power budget of up to 130 watts, drawn from the same battery that drives the motors, so part of the price of a smarter robot is paid in autonomy, the time it can work between charges.
Ends and means
Beyond the figures, Tuesday’s panel left me with an idea that I am borrowing rather liberally: robots, and AI agents, should be anchored by values, the human kind, as well as driven by goals. A system judged only on reaching its goal has every reason to follow the principle commonly attributed to Machiavelli, that the end justifies the means; the OpenAI agents that broke into Hugging Face’s systems in July, apparently to steal the answers to a test run with reduced safeguards, are the most prominent example. Values are what would rule out certain means before the work begins, and they matter even more when the system pursuing the goal has a humanoid body and shares a floor with people.
“Trust is what takes machines out of the lab and into the real world,” said Carina Kaeser, co-founder and chief executive of Ply, who sat on the same panel in Barcelona.
Ronald Reagan liked to quote a Russian proverb to Mikhail Gorbachev: trust, but verify. With machines, the verifying will have to be designed in.
Wirelessly yours,
Ziad Matar
Co-founder, Tokenando
Editor-in-chief, The Compute
After Hours
On screen, I have been rewatching the first season of The Gentlemen so that I am ready for the second, which Netflix released on September 3, while clearing whatever is left on my “continue watching” tab and getting ready for the sixth season of Slow Horses, which Apple TV began releasing on September 16.
On the page, I found my copy of Asymmetry by Lisa Halliday, which I had misplaced halfway through, so it is back in my reading queue behind The Flames, Sophie Haydock’s novel about the women who sat for Egon Schiele, and La question d’Orient by Jacques Frémeaux.
In my AirPods, Caamp made his usual appearance, alongside two Souad Massi classics, “Raoui” and “Ghir Enta”, and a late discovery: Ondara, and his song “Lebanon”.
References
Interesting Engineering, “Video of humanoid crashing in half before Robot Olympics goes viral”, August 21, 2026: https://interestingengineering.com/ai-robotics/china-humanoid-robot-crashes-hard
Reuters via Al Jazeera, “Chinese robot Tiangong clocks sub-9-second 100 metres in Beijing”, August 25, 2026: https://www.aljazeera.com/sports/2026/8/25/chinese%2Drobot%2Dtiangong%2Dclocks%2Dsub%2D9%2Dsecond%2D100%2Dmetres%2Din%2Dbeijing
Forbes, “34,000 Pre-Orders, $2.4 Billion Pipeline: Inside Europe’s Newest Humanoid Robot Unicorn”, September 8, 2026: https://www.forbes.com/sites/johnkoetsier/2026/09/08/34000-pre-orders-24-billion-pipeline-inside-europes-newest-humanoid-robot-unicorn/
Reuters, “British tech company Humanoid targets U.S. IPO by 2030”, May 13, 2026, via MarketScreener: https://www.marketscreener.com/news/british-tech-company-humanoid-targets-u-s-ipo-by-2030-ce7f5bdcd88eff27
Morgan Stanley humanoid report, as reported by KrASIA: https://kr-asia.com/pulses/162404
Bank of America Institute, “Transformation: Physical AI, part 2: Humanoid robots”, March 12, 2026: https://institute.bankofamerica.com/content/dam/transformation/physical-ai-part-2.pdf
Fortune, “More people will own a humanoid robot than a car by 2060, BofA predicts”, March 13, 2026: https://fortune.com/2026/03/13/bank-of-america-humanoid-robot-forecast-3-billion-2060
International Federation of Robotics, World Robotics 2026, September 24, 2026, as reported by RoboticsTomorrow: https://www.roboticstomorrow.com/news/2026/09/24/five-million-robots-now-operate-in-factories-globally/27154
Unitree R1 price cut, Hangzhou government news portal, June 2026: https://www.ehangzhou.gov.cn/2026-06/26/c_298239.htm
Unitree G1 price and specifications, EE Times Asia: https://www.eetasia.com/videos/unitree-robotics-humanoid-robot/
Unitree G1 and H1 list prices, The Robot Report: https://www.therobotreport.com/unitree-robotics-unveils-g1-humanoid-for-16k/
TED, “Elon Musk: A future worth getting excited about”, interview with Chris Anderson recorded April 6, 2022: https://www.ted.com/talks/elon_musk_a_future_worth_getting_excited_about
The Robot Report, “Humanoid takes seven-month path to HMND 01 Alpha”, January 15, 2026: https://www.therobotreport.com/humanoid-takes-seven-month-path-to-hmnd-01-alpha/
TechRepublic, “Nvidia’s Jetson AGX Thor ‘Robot Brain’ Is Now Available”: https://www.techrepublic.com/article/news-nvidia-thor/
Nvidia, second-quarter fiscal 2027 results, August 26, 2026: https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000073/q2fy27pr.htm
OpenAI, “The Hugging Face incident and the road ahead”, August 26, 2026: https://openai.com/index/hugging-face-incident-and-the-road-ahead/
Hugging Face, technical timeline of the agent intrusion: https://huggingface.co/blog/agent-intrusion-technical-timeline
AI Summit Barcelona, Panel 6 on physical AI, September 22, 2026