On 16 February 2026, the eve of the Year of the Horse, four Chinese humanoid robot manufacturers — Unitree Robotics, MagicLab, Galbot and Noetix Robotics — stepped onto the biggest stage of the year, CCTV's Spring Festival Gala. The press fixed on the choreography: Unitree's G1s notched up several world firsts for a humanoid there — the first continuous table-vault parkour, the first ejection backflip from over three metres, the first continuous one-legged flips, and a 7.5-rotation Airflare. Spectacular, and misleading. Because the news of the week is not the dance. It is what happens just after the curtain falls.
The stage was a test bench, not an end in itself
Wang Xingxing, Unitree's founder, was not selling a circus act. He was selling an infrastructure. In his words, the Gala's cluster control performances "lay the groundwork for the future deployment of robots in cluster operations or single-unit deployment across other scenarios". Translation: the stage was merely a demonstration of simultaneous piloting. The proof came hard on its heels. After the Chunwan, Unitree filmed 49 G1s performing a synchronised martial arts demonstration in front of the Temple of Heaven in Beijing — on the same cluster control platform as the gala's. The television set and the public square share the same technical building block. The show was not the product; it was the product's sales pitch.
The real figure is a purchase order
If you look for the hard metric that the kung-fu folklore papers over, it lies elsewhere. In the first two hours of the Gala's broadcast, JD.com recorded a 300% rise in robot searches, a 460% rise in customer-service enquiries and a 150% rise in order volume. And this demand was not confined to the big cities: the purchases came from more than a hundred Chinese cities, from first-tier conurbations down to small towns. The stage converted, in real time, viewers into buyers.
The industrial response is immediate and quantified. After the Chunwan, Unitree announced it was aiming to deliver 20,000 humanoid robots over the course of 2026, against roughly 5,500 units in 2025 — a fourfold increase in its manufacturing capacity, presented explicitly as a response to post-gala demand. The backflip is not the information. The x4 in capacity is.
On both sides of the Pacific, capital does the same sum
That same week, the money confirms the diagnosis — and it is no longer paying for the demo. On 11 February, Apptronik closed a Series A-X extension that lifts its valuation beyond five billion dollars, roughly triple its initial 2025 Series A. The detail matters more than the amount: its Apollo robot is already deployed in real-world conditions, in factories and warehouses, at Mercedes-Benz and GXO Logistics. And the use of funds is plain — to scale up Apollo's production, expand the global deployment network, and build dedicated robot-training and data-collection facilities in Texas.
On the other side of the Pacific, the same sum. Galaxea AI closed a Series B led by Jinding Capital, which lifts its post-money valuation to ten billion yuan, or roughly 1.4 billion dollars. The make-up of the round is the signal: three new industrial investors come in — BAIC Group Industrial Investment, Hone Capital and Yifeng Capital. From Beijing to Austin, capital no longer funds a choreography. It funds the factory and the data.
The hidden pattern: the human who wears the sensors
And the data, precisely, is still collected on a human. This is the most buried event of the week, and the most revealing. On the same day, or nearly, that the G1s were dancing on television, AgiBot spun off Maniformer, a subsidiary in which it retains 75% of the capital — a controlled spin-off, not a sale. Yet Maniformer's product is not a robot. It is a data platform for embodied AI, positioning itself to plug a shortfall estimated in 2026 at more than five million hours of quality data.
Its method completes the unmasking of the sector: a collection method "without a physical robot", in which human operators wear sensor devices to carry out tasks in real-world conditions and record trajectories, vision and tactile feedback. The human in the sensor harness is the robot, for the duration of the learning. And the operation is anointed by capital: Sequoia China led its seed round, and the funds are earmarked for expanding data-production capacity, international deployment and the creation of a global data alliance for embodied AI.
Two faces of the same week
The "autonomous" robot of the Gala and Maniformer's human-in-a-harness are not opposed: they are the two faces of one and the same industry. One puts on the show and generates the demand; the other quietly does the learning that will make that demand sustainable. The question, then, is not whether the Gala's robots were really autonomous. It is who picks the order placed on JD.com — and who wears the sensor harness meant to teach these machines to become so. The body dances, in the East as in the West. The data, for its part, is still collected on a human.