In a workshop in Changzhou, several hundred robots are pushing wire harnesses into place.
A harness is soft.
Soft parts are the hard ones. A harness has no fixed shape: lift it and it sags, insert it and you work by feel. Push too far and you damage the wire. Stop short and the circuit never closes. The line runs at a takt time set before the robots arrived.
This is one of the few production floors anyone can point to this week.
That same week, $500 million went to Genesis AI and $500 million to Agile Robots, the second round led by SoftBank. Two checks, one direction: machines into factories and onto sites, into stations that have always depended on hands.
The same landing spot
Different companies, same target. That direction does not lack believers. It lacks repeat orders.
A second order depends on how long the first unit stayed on the station, how often it broke, how many minutes a fixture change costs. None of that appears in a funding announcement.
What follows a demo is usually video of a hand or a foot.
Video can be cut. Takt time cannot.
A round is not an acceptance form
$500 million buys equipment, people, and time. It does not buy the one thing that matters: someone already using the thing.
Sort this week's projects by stage: the shape is lopsided. Announced, piloting, planned: most of them. Confirmed volume production use: very few.
Funding discounts future revenue into today. A line puts today's cycle time on the table. One is signed once; the other runs one shift at a time. Discounting assumes the future arrives, and the line is where that assumption gets tested.
Headcount and test benches move within a quarter. Units on customer floors do not, because that number belongs to procurement. It waits for a signature, a stable cycle, a yield that holds.
None is in the announcement, which is why funding news and production news sit a year or more apart. Each gate from prototype to small batch to volume is a fresh customer acceptance. The announcement covers the first. The floor covers the last.
What 75 percent counts
Chengdu Rail Transit published a second list of 10 rail-plus scenarios on November 19. Of the first list's 16 items, 8 were signed and 4 had a partner designated.
8 plus 4 is 12. 12 over 16 is 75 percent. The second list has 10 items, 6 fewer than the first. Four items sit outside both counts.
Items on a list and units on a site are two sets of books. 75 percent is a signing rate, not a delivery rate. After a signature comes vendor selection, site entry, and commissioning. If any step stalls, the signed count never falls and the unit count never rises.
Five standards, four layers
MIIT's TC8 published 5 humanoid robot standards covering capability, components, data, and application. Five documents, four layers.
Standards solve comparability first. Two machines on one table need one ruler. Having the ruler is not the same as having the scale marked.
Capability asks what a machine can do. Components covers joints, reducers, hands. Data covers collection, labeling, interfaces. Application defines what passes in a setting. That is the hardest layer, because a bank lobby and an automotive line do not share a pass line.
On the Chengdu-Meishan line, robot dogs are testing safety patrols. The operator's hazard model covers 12 construction scenarios at above 85 percent accuracy.
Unpack 85: 15 of every 100 recognitions are wrong. A miss costs later; a false alarm costs now. A model entering a site has to clear the false-alarm rate first: a foreman buried in bad alerts turns them off, and a switched-off model collects nothing. Whether 85 holds is decided at the hundredth prompt, not in the lab.
3.6 patents a day
GRG brought a humanoid bank teller to a show floor. It has applied for more than 4,500 patents in 5 years.
4,500 over 5 is 900 a year. Across 250 working days, 3.6 a day.
It also reports more than 70 national standards it led or helped revise.
Neither converts into units. A company can hold many patents and run a handful of machines; another can hold few and have units working three shifts. Investment is what the company spends. Delivery is what the customer accepts.
A 500-millimeter box
Tieke Yingmai and Huawei released an onboard inspection unit with a 500-millimeter body that identifies 7 defect classes in real time, among them fastener defects, rail scuffs, mud pumping, and balise faults.
500 millimeters is not much longer than an adult's shin. It rides on the car; the car moves; the defects stay in the photographs.
Seven is the length of a list, not of the problem. Recognizing all 7 while the car moves is the work.
Which column
Two columns. Left: $1 billion across 2 rounds. Right: several hundred units in Changzhou, a floor of 1,000, a target of 10,000.
They do not divide: no unit price has been published. What compares is pace.
The left column fills in a week. The right moves in single steps, each at the speed of the slowest station. Capital arrives at once; capacity moves one process at a time, and the process that will not move is the bottleneck.
The measure is the column marked confirmed volume production use. It is nearly empty. Empty does not mean zero. It means not yet ready to be written down.
On the Changzhou line, harnesses go in one at a time. The hand stops where it should: early and it will not seat, late and it collides. In a video the motion takes 3 seconds.
On the line it runs all day.
The target is 10,000. What is in there now is several hundred.
*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*
