The biggest customer for robot companies is local government data-collection centers
Last year the industry's single largest source of revenue was not factories and not households, but data-collection centers with local government participation — robot companies sell robots to local state capital, local state capital hires people to collect data and sells it back to the robot companies, and that closed loop props up the revenue threshold for an IPO.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Everyone raised a lot of money, but spent very little
The most off thing about the industry is that it is a financing race, not a technology race: everyone raised a lot of money, but spends very little. The two reporters went to ask compute providers and data providers and found that the industry invests little in the two places where it most needs to spend. One interviewee said only eight companies in the whole industry have clusters of a thousand cards or more; after the article came out, someone fed back that ‘eight is already too many’. A founder of Jusheng also said the whole industry has never run a model once on more than a hundred thousand hours of data, which contradicts many companies' own public statements.
— Zi NanProgress is hard to observe, so over- and under-estimation swing back and forth
The biggest difference between embodied intelligence and large language models is that progress is hard to observe: large models have public benchmarks, open weights, and ordinary people can try them; embodied intelligence is 2B, a costly physical entity, with no recognized public benchmark, and the same model runs differently on different hardware, so reproduction itself is harder and more expensive. The result is that this industry is either greatly overestimated or underestimated, or swings between those two states, and there is a lot of room for storytelling and bubble-blowing.
— Zi NanR&D spend of 30 million, financing of over 5 billion
A robot company with a valuation that still counts as top-tier spent roughly 30 million yuan on R&D last year, but has raised more than 5 billion yuan to date. The reporters calculated that putting that money in Yu'ebao would earn 50 million a year in interest, more than the R&D cost. By comparison, Unitree spent 145 million yuan on R&D last year, Mixue Bingcheng spent 105 million yuan on R&D last year, and AgiBot's R&D spend last year may have been around 500 million. The core problem is that some of China's most leading companies lack the judgment to decide what direction to bet on, and without that judgment, R&D is not easy to invest in either.
— Zi NanData-collection centers were last year's biggest customer
A robot company CEO said that last year the industry's biggest customer was data-collection centers in various localities. The model is that local government and a robot company form a joint venture data-collection company, the government puts up money and land, the joint venture operates it, and it hires university students or young people as teleoperators to collect data; one robot is paired with at least two people, so a hundred robots create two hundred jobs. Where local government really puts up money is buying robots; a robot sold to a data-collection center costs roughly four to five hundred thousand yuan, so a center with a hundred robots is about 50 million. Put together four such local government data-collection centers and the company clears the Hong Kong 18C revenue threshold of over 200 million yuan.
— Zi NanData-collection centers also idle and break up
After data-collection centers land, they do not all run well: there are utilization problems, data collected cannot be sold, and they still have to carry labor costs, so they simply do not start up; there are problems of long payback periods and uncertain payback; and some data-collection centers resell the robots to other units. A provincial capital in central China bought 80 million yuan of robots, and early this year the data-collection center broke up and the robots were sold to nearby colleges as teaching aids. Technology route changes are also hitting this model: now more people do EGO data and WMI data, demand for data collected by Zengji declines, and some robot companies voluntarily lower the revenue share from this part.
— Zi NanManufacturing customers are actually each other's suppliers and customers
Last year a Jusheng Intelligent startup got nearly forty percent of revenue from manufacturing scenarios, which is surprising, because humanoid robots' ability to work directly on production lines is still relatively limited. The mechanism is that robot companies sell robots to manufacturing parts companies, and the parts companies sell their parts back to the robot companies, so each is the other's supplier plus customer. Manufacturing companies also invest in robot companies: spending a billion to build a factory is worse than spending five hundred million to invest in a robot company, and incidentally enter a new field, set up a robot division, and transform into a robot supply chain. For a 110 company, robot revenue may be only 50 million, but everyone expects it to double next year and double again the year after, and is willing to give a higher valuation — in essence, spending money to buy an expectation.
— Zi NanAn 8 million order levers up the stock fivefold
A very clear example is Changsheng Axle: after Unitree appeared on the Spring Festival Gala, people went to dig through Changsheng Axle's financials and found that Unitree and other robot companies had bought only about 8 million of parts from Changsheng, yet Changsheng Axle rose fivefold from the bottom. The revenue robots contributed to it was not that high, but the market cap increase the capital market gave it was very exaggerated. There are also companies whose board secretary, after selling parts to robot companies, will tell analysts ‘I am no longer an auto industry company, I am now a robot company’, demanding to be valued at a robot supplier's multiple rather than 15x.
— Zi Nan20 million becomes the threshold for AgiBot VAP
AgiBot has a tiered sales system, graded by the sales performance you complete for AgiBot in a year: 20 million is VAP, 10 million is gold, 5 million is silver, 2 million is the lowest certification. The higher the tier, the more supply you get and the more support AgiBot gives. Among the highest-tier VAPs are companies like Junsheng Electronics, Ningbo Huaxiang, and Wolong Electric, basically the stocks that rose best in the robot sector last year. These distributors also actively help AgiBot expand scenarios, for example wondering whether a robot could also move batteries in a battery-making factory, which the AgiBot team had not thought of before.
— Zi NanMusk pulled expectations too high; Deng Taihua is more accurate than Musk
The robot progress itself fell short of expectations, and the biggest responsibility lies with Musk: he previously said 100,000 units in 2026, everyone started punching calculators, and by August only 300 had been built, annualized at 3,600, a 30x gap from 100,000 — a super big miss. By contrast, the guidance Deng gave was 16,000 units, more than Elon's low thousands, and on the domestic chain everyone considers this guidance relatively reliable and achievable. Moreover, Deng's credit record is better than Musk's, his commitment fulfillment rate is higher, and 16,000 units is not a particularly aggressive increase over their output of four to five thousand last year.
— Zi NanClub Deal: permutations and combinations of T0 industrial investors and T0 VCs
Jusheng's financing syndicate model is not like traditional venture capital: there is a syndicate organizer, possibly an FA or someone at an active institution, who sets up the camp from the first round, finds several industrial investors and several VCs, and also splits T0 industrial investors and T0 VCs. T0 industrial investors are roughly 6 to 8, including AgiBot, Zibianliang, Leju, Galaxy, and so on; T0 VCs are roughly 10 to 15, including Hongshan, Gaoling, Shunwei, Chunhua, Yuanma, and so on. On hot projects these institutions permute and combine, each round may have a three-plus-three structure, and pulling them in for the first few rounds can endorse later financing. Some institutions were asked to commit to leading multiple consecutive rounds, and some did not invest because they felt the demand made no sense.
— Xu YumengSecondary-market investors start looking for verifiable metrics
The bubble continues, but people in the secondary market have started actively looking for verifiable, quantifiable things, such as ROI and the conversion efficiency of R&D spending. When they research a company they explicitly ask how much it wants to spend on R&D this year, how much next year, what the revenue expectation is this year, where revenue comes from, what the order book looks like, and what the repurchase probability in this field is. A friend asks, and only buys if they can answer; if the answers lack detail, they do not buy. Wang Xixing said publicly that the industry may see some applications land within two to three years; that expectation is lower than what industry investors are pricing in, but may also be more reliable.
— Zi NanThe most important validations are data scaling and post-IPO pricing
There are two truly important issues for the industry: one is the technology itself — companies like AgiBot say that with ten million hours of data they might be able to verify whether the model actually scales, and the more top-tier companies in the industry may already be close to that goal, with some answers possibly coming this year; the other is the price trend of Unitree in the six months after its IPO — entering the secondary market means entering the public market, where it will be fully priced, the people who pumped the price up early will gradually exit, and the value left at the end will in theory hew closer to its intrinsic value. Investors fall into two types: those who want to make theme money and hope to exit as soon as possible, and those who firmly believe in the industry's value and whose core is confirming that the founding team is seriously doing technology.
— Zi NanIn their own words · checked verbatim
Everyone raised a lot of money, but everyone spends very little, and what we feel is probably still that everyone is in a financing race, or an IPO race — we compare who has the higher valuation, and we compare who lists earlier.
大家拿了很多钱 但是大家花钱很少 然后我们感受到的可能还是 大家处在一个融资竞赛的阶段 或者说上市竞赛的阶段 我们比谁估值更高 然后我们比谁更早上市
Zi Nan2:04
This industry has another problem, which is that its progress is very hard to observe, and this leads to it either possibly being greatly overestimated, or possibly being underestimated, or just swinging between those two states.
这行业还有一个问题 就是它的进展很不好观测 这就导致 它要么有可能会被极大的高估 要么有可能会被低估 或者就在这两种状态之间在摇摆
Zi Nan5:08
Last year the industry's biggest customer was actually data-collection centers in various localities.
去年整个行业其实最大的客户是地方各地的数据采集中心
Zi Nan19:14
Robot companies sell robots to manufacturing parts companies, and then the parts companies sell their parts to the robot companies, so each is the other's supplier plus customer.
机器人公司把机器人卖给制造业零部件公司 然后零部件公司又把他的零部件卖给机器人公司 那这样就是互为对方的供应商加客户
Zi Nan29:24
It pulled everyone's expectations too high — it came right out and previously said 100,000 units in 2026, so everyone started punching calculators.
它把大家的预期 拉得太高了 它上来就 之前就说2026年10万台 那大家就 计算器就开始摁了
Zi Nan45:37
The first four parts are: question the bubble, understand the bubble, embrace the bubble, enjoy the bubble. What is the fifth part?
前面四部是 质疑泡沫 理解泡沫 拥抱泡沫 享受泡沫 第五部是什么
Zi Nan54:38
Figures
| Number of companies in the whole industry with clusters of a thousand cards or more | 8 (someone fed back that 8 is already too many) | 3:06 |
| Cumulative financing of a top-tier robot company | over 5 billion yuan | 13:14 |
| Unitree's R&D spend last year | 145 million yuan | 14:14 |
| Mixue Bingcheng's R&D spend last year | 105 million yuan | 14:14 |
| AgiBot's R&D spend last year | about 500 million yuan | 14:14 |
| Hong Kong 18C revenue threshold | over 250 million Hong Kong dollars | 18:14 |
| Price of a robot sold to a data-collection center | about four to five hundred thousand yuan each | 21:16 |
| Scale of a data-collection center | generally a hundred robots, about 50 million | 21:16 |
| Number of publicly findable data-collection centers | more than 80 | 22:16 |
| Changsheng Axle's robot parts order | about 8 million yuan | 32:24 |
| Domestic humanoid robot sales at the end of last year | about 8,000 units | 34:29 |
| Tesla's robot output in August | 300 units | 45:37 |
| Robot guidance given by Deng Taihua | 16,000 units | 45:37 |
Glossary
- 18C / Hong Kong Listing Rules Chapter 18C
- Listing rules that took effect at HKEX in 2023, allowing specialist technology companies to list with a lower revenue threshold.
- VAP / VAP distributor
- The highest-tier distributor in AgiBot's sales system, completing over 20 million in sales performance for AgiBot in a year.
- Club Deal
- Multiple institutions investing jointly in the same round, spreading risk and getting a seat at the table together.
- T0 / T0-tier institution
- The most top-tier, most active industrial investors or VCs in the industry, often pulled into the financing camp of hot projects.
- EGO data / first-person-view data
- Operational data collected from the robot's own perspective, as distinct from data collected by Zengji.
How to listen
Investors and founders watching embodied intelligence and the humanoid robot sector, and practitioners who want to understand where this industry's revenue comes from and how the IPO race is being run.
If you only want technical progress, you can skip the data-collection-center business-model section from 19:14 to 28:22.