By Sarah Xiao 
Two numbers, nine days apart, describe 2026 in embodied AI better than any forecast.
On one side: XPeng’s robotics unit closed a $900 million first round, the largest single private placement in Chinese embodied AI to date. TianGong Ultra cut the 100 m sprint for humanoids from 21.50 s to 8.64 s, inside the human world record.
On the other side: Nanjing Wujing Future Power Technology stopped paying salaries in August. Around thirty to forty people are owed close to 8 million yuan. The company has halted operations.
The same nine months produced 437 funding events worth roughly 203.2 billion yuan, according to GaoGong Humanoid Robot’s count. That is 146% more deals and 317% more money than the same period in 2025. And it also produced a wage dispute that emptied an office in Nanjing.
The money is real, and it is narrowing
The count of companies is the loudest number. GGII puts the humanoid supply chain at 692 companies today, against 332 a year ago. The roster doubled in twelve months.
New entrants keep arriving from outside robotics. In September alone, seven companies founded in 2026 closed rounds, about 13.5% of that month’s deals. Feimo Yi and Xieyue Intelligence were founded by autonomous-driving executives. Xirang Kaiwu came out of Huawei’s foundation-model organisation. Xiyuan Robot and Zhongyi Embodied were spun out of the embodied divisions of RoboSense and Zhongyi Group.
Then the money curve bent. Monthly funding peaked at 35 billion yuan in June and fell every month after, reaching 13.4 billion in September. Several IPO processes that had been treated as certainties ran into delays.
The counters started running
What separates 2026 from the demo years is that some companies now publish production numbers instead of video views.
Qiyuan Robot announced its 5,000th personal robot off the line on 25 October. The product went on sale on 20 September. That is 5,000 units in thirty-five days.
Wang Xingxing said at the 2026 World Physical AI Conference that Unitree has built roughly 10,000 to 20,000 G1 units since the model launched in May 2024.
Field data arrived alongside the unit counts:
- Figure AI livestreamed parcel sorting continuously for over 200 hours, handling about 249,600 packages.
- Zibianliang sorted 1,816 parcels in one hour.
- Agibot’s Lingxi G2 ran a full tablet quality-inspection line on camera: 64,828 items processed, 99.99% task success.
Prices went the other way. Unitree pushed its general-purpose humanoid to the 29,900 yuan level. Songyan Power’s Xiaobumi sells below 10,000 yuan, less than a decent laptop.
The floor fell out of the rental market
The first business model to break was the one that arrived first: renting robots for events.
Daily rental rates for humanoid performance work ran at ten-thousand-yuan levels in early 2025. They are now at thousand-yuan levels, an 80 to 90% decline. On Xianyu, Unitree units start at a hundred yuan a day.
Rental was never going to fund an industry, but it was the earliest evidence that anyone outside a lab would pay for a humanoid. When that price collapses by an order of magnitude in eighteen months, it says the supply of robots grew faster than the supply of events worth renting them for.
Where paid work actually exists
Production deployments remain single-station or small-batch pilots, concentrated in low-complexity, highly structured environments.
China Southern Power Grid is running embodied-model robots at substations in Shenzhen and Guangzhou. Beijing Anzhen Hospital has brought in several classes of medical robot covering outpatient, pharmacy and ward tasks. Both are closed sites with repeatable routes, known obstacles and a supervisor within earshot — the conditions under which today’s models behave.
The model gap is now measured in data, not parameters
Two technical arguments moved to the front this autumn.
The first is whether foundation models will swallow embodied models outright. GPT-6 Astra’s performance on embodied tasks restarted that debate, and the industry has not settled it.
The second is measurable and worse. Most Chinese teams are still fine-tuning on π0, π0.5 or OpenVLA — baselines that were open-sourced a while ago — and sit roughly two generations behind the leading closed models.
Data tells the same story with harder numbers. Dyna-2 was trained at million-hour scale. Going from 100,000 to one million hours of pre-training data moved average success across 14 self-built task sets from 45% to 53%. Eight points for ten times the data. Most Chinese models are still under 100,000 hours, a full order of magnitude short.
That curve is the real constraint. If a tenfold increase in data buys eight points of success, then closing a data gap is not a matter of buying more GPUs. It is a matter of who owns the environments that generate hours.
Policy picked a side
On 23 October the State Council General Office issued its opinion on adapting to AI development in ways that support employment-friendly growth. The document says AI applications should be prioritised in sectors short of labour and in high-risk environments.
Read as a procurement signal, that points away from showroom humanoids and toward substations, mines, hospital logistics, construction and hazardous inspection — precisely the closed, structured sites where pilots are already running. For robot companies it raises the cost of choosing the wrong scenario, and it makes the scenario choice a commercial decision rather than a demo decision.
The honest summary
Wang Xingxing’s line at the same conference is the closest thing to a balanced verdict anyone has offered publicly: the industry is very hot, the cake is very large, the prospects are very broad, and the competition is very brutal.
All four clauses are true at once, and that is what makes 2026 hard to read. The funding totals, the unit counters, the field hour logs and the rental collapse are not contradictory signals. They are the same signal measured at different points in the value chain. Capital arrived first, hardware volume followed, paid work has not caught up, and the companies that cannot reach paid work are now running out of runway while the headline numbers keep setting records.
The question for 2027 is not whether embodied AI is real. It is how many of the 692 companies can name the customer who pays twice.
Theodore Li