On 28 October the Hong Kong Exchange was supposed to ring its bell for Changzhou Weiyi Zhizao. The price range was set. The maximum raise was about HKD 800 million. Three days before the listing, the company told the market it was postponing the global offering and returning every subscription payment in full, without interest.
It was the most visible crack in a wall that had been under pressure for weeks. Within the same stretch of days, AgiBot was reported to have shelved its Hong Kong plan and returned to private money, raising RMB 1 to 1.5 billion at a valuation of roughly RMB 50 billion. A month earlier, Huandong Technology, a maker of RV reducers for robot joints, withdrew its application for Shanghai’s STAR Market.
The profitable one left first
Huandong is the case that breaks the usual explanation. The company booked RMB 438 million of revenue and RMB 93.76 million of net profit in 2025. It was not a loss-making narrative stock looking for a lifeboat. Its parent, Shuanghuan Driveline, said market conditions had changed substantially since the spin-off was planned.
What the exchange filings show is narrower and more uncomfortable. Five customers accounted for 78.34 per cent of Huandong’s revenue, and Estun alone took close to half of it. Two rounds of questions kept returning to the same point: without the parent and the single largest buyer, is this an independent business?
That is a technical question, and technical questions can be answered with time. The same cannot be said for the other two categories of candidate now sitting in the queue.
Three different businesses wearing one label
Wind counts more than 50 embodied AI companies waiting on the Hong Kong Exchange and more than 20 on the mainland, including Leju, DEEP Robotics, Galaxea, LimX Dynamics, EngineAI, Robbyant, Ziyuan Robot and Galbot. Putting them in one queue hides the fact that they are priced on completely different logic.
Component suppliers have delivery records and gross margins. Project-based industrial robot makers have real revenue and real repeat orders, but pay for them in working capital. Weiyi holds roughly 20 per cent of China’s industrial embodied AI robot market by 2025 revenue, and its cash balance more than halved in six months. DEEP Robotics sells its X series at RMB 287,500 a unit, moved 681 of them in 2025 for close to 60 per cent of company revenue, and reports a 28 per cent repeat-purchase rate that contributes 68 per cent of that revenue. The revenue is genuine. The collection cycle is long.
Then there are the humanoid builders, where valuations ran hardest and the questions are least comfortable. How much of revenue comes from government projects and research procurement? Where is the paying industrial customer?
By the humanoid application alliance’s count, Chinese public tenders in the first half of 2026 came to 218 projects worth RMB 1.723 billion. Educational and research institutions took 55.5 per cent. Government and state-owned platforms took 20.6 per cent. Industrial companies placed 21.1 per cent, and most of those orders stopped at testing and pilot runs.
The secondary market changed the ruler first
Unitree’s stock traded as low as RMB 404 intraday in mid-October, more than 60 per cent below its peak, wiping out close to RMB 280 billion of market value. UBTech and Dobot both touched year lows in Hong Kong. The rebound since 19 October has been sharp, with Unitree back above RMB 281 billion and the humanoid index up almost 5 per cent on 28 October, but the standard has moved.
IDC puts global humanoid shipments in 2025 at about 18,000 units. Entertainment performance, research and education, data collection, and guided tours together account for nearly 90 per cent of them. Manufacturing accounts for about 9 per cent.
The revenue question became public in September. Shao Tianlan, founder of Mech-Mind, used his personal social feed less than ten days after his own listing to attack what he called assembled companies: high valuations, high profiles, and revenue manufactured by selling robots to data collection centres that then sell the data back. By industry counts there were 106 centralised data collection centres built or under construction nationwide as of August, 84 of them operating. The centres are not the problem. A model needs real-world data. The problem is how much of the purchase money comes from local state capital and how much of the data is bought back by the seller.
Why waiting is the cheaper option
First-half funding in Chinese embodied AI reached RMB 93.5 billion across 322 deals, up 137 per cent year on year on IT Juzi’s count. Demand has not kept pace. Frost and Sullivan puts global humanoid sales in the same half at 36,000 units and USD 540 million of revenue. The money went in at one speed and the market is arriving at another.
For a company that has burned cash on research and is waiting for volume, a listing priced into that gap is expensive in a way that does not show up on the day. It shows up over the following eight quarters, in a share price that cannot recover.
Delaying is not free either. It hands the slot to a competitor, and Hong Kong has more than 50 candidates standing in line. But the arithmetic favours the wait: fix the order book, prove the second sale came from the value the first machine delivered, and strip the subsidies and related-party transactions out of the revenue line before facing an examiner who now asks for all three.
The window has not closed. It has simply started pricing the difference between a robot that ships and a robot that works.
Editor’s note: This is an adapted translation of the original OFweek report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://robot.ofweek.com/2026-10/ART-898890-12003-30709229.html.