For months, this paper has recorded the same contest on a single scoreboard: China delivers the tonnage, the West sells the story. On that axis, the West always loses. Unitree Robotics, based in Hangzhou, delivered more than 5,500 humanoid robots in 2025, the world's leading volume, ahead of all its American competitors. AgiBot followed with some 5,168 units, while Tesla Optimus, Figure AI and Agility Robotics each shipped around 150 robots that same year. All told, Chinese companies accounted for between 85 and 90% of global deliveries in 2025, with six of the top ten vendors based in China. Capital is following suit: in the first quarter of 2026, China's sector raised 68.1 billion yuan, more than across the whole of 2025, with fifty funding rounds in the month of March alone.
This week, the West changed axis. Three developments, read separately, look anecdotal; cross-referenced, they trace out a single bet: stop chasing the number of bodies, and crack the hand instead.
Genesis AI: the hand before the body
On 6 May 2026, Genesis AI announced GENE-26.5, billed as the first foundation model specialised in human-level robotic manipulation, capable of driving a proprietary dexterous robotic hand. The demonstration looks nothing like a tonnage show: the model solves a Rubik's Cube by manipulating it in mid-air with a single hand, and plays the piano at a human level. The task catalogue is deliberately hostile to cheating: cooking a meal in twenty steps with chopping, a one-handed egg crack and bimanual coordination, smoothies prepared in mid-air, high-precision laboratory experiments — pipetting, liquid transfer — and cable-harness assembly.
The hardware follows the same logic: the Genesis Hand 1.0 has twenty active degrees of freedom and reproduces the anatomy and function of the human hand at a 1:1 scale. The centrepiece, though, is not the hand but the data. The collection glove developed by Genesis, fitted with a tactile-sensor electronic skin, achieves a 1:1:1 mapping between the human hand, the glove and the robotic hand, at a cost said to be a hundred times lower than existing alternatives. This is where the shift in unit of account plays out: not one more robot on the line, but a dexterity-learning loop that is finally affordable.
The company itself runs counter to the mega-factory folklore. Founded in December 2024 by Zhou Xian (CEO, a former Carnegie Mellon researcher) and Théophile Gervet (president, a former Mistral AI researcher), it has lined up on its cap table Eric Schmidt, the former Google CEO, the entrepreneur Xavier Niel, and the researchers Daniela Rus and Vladlen Koltun. Six months after its creation, the demonstration drops — a tight timeline that invites caution, but that also says where Western money is going: towards the gesture, not the gross frame.
Figure and BMW: paid work, not a show
On 9 May 2026, Figure AI confirmed the commercial deployment of the Figure 03 at BMW's Spartanburg plant in South Carolina, with an initial fleet of forty units, presented as the first large-scale commercial deployment of a general-purpose humanoid in an industrial setting. The contract covers concrete tasks — parts handling, sub-assembly placement, quality inspection — billed at around 25 dollars per operating hour. This is no longer a viral video: it is a line of cost, something you can argue against, set against the hourly cost of an operator. The West, accused until now of selling nothing but emotion, is at last presenting a hard capability put to work and paid by the hour.
Tesla frees the premium line for Optimus
On that same 9 May, the last Tesla Model S and the last Tesla Model X rolled off the Fremont line in California, ending production of those two models. The line thus freed up is being converted to build the Optimus Gen 3. The gesture reads clearly: the manufacturer that made automotive tonnage its identity is sacrificing its premium range to make room for the humanoid body. Here too, Western capital is shifting from the mass-production vehicle to the machine that manipulates.
Two races, two units of account
The contrast with China is not a matter of engineering quality: it is a matter of metric. The Unitree G1, flagship model of the world's number one, stands 1.32 m tall, weighs 35 kg, lines up twenty-three degrees of freedom for the whole body, lasts around two hours on a 9,000 mAh battery and lifts 3 kg per arm. It is a locomotion platform that walks, sturdy and delivered by the thousand. At Fourier Intelligence, the GR-2 pushes to fifty-three degrees of freedom for 63 kg and two hours of battery life, after more than a hundred GR-1 units already delivered to companies across varied sectors.
These figures describe bodies that move about and carry a load — not hands that crack an egg without crushing it. That is precisely the distinction the week brings to light: a volume of locomotion platforms is not a dexterous-manipulation capability. Twenty-three degrees of freedom spread across a whole body are not worth twenty concentrated in a single hand at human scale. As the thesis defended by Genesis essentially puts it, the brain and the hand are the two most complex parts of the robot, and they are the ones the West has chosen to take on.
This piece owns its continuity with the analyses of 30 March and 4 May: "Washington sells the tender humanoid, Shanghai ships 10,000 of them", "the West sells the factory, the service and the OS". Those texts faulted the West for showing no hard capability, nothing but story and emotion. This week, the thesis flips: the West is at last displaying a measurable capability — manipulation — and real paid work. The conclusion swings, but stays falsifiable: let a Chinese champion bring out a human-level dexterous hand, and the Western bet falls. In the meantime, none of the thousands of robots delivered in China in 2025 can do what GENE-26.5 demonstrates this week. Two races, two units of account — and an implicit bet: whoever cracks the hand makes the bodies that walk a commodity.