Two numbers from the same company, published within months of each other, tell opposite stories.
The first: Citic Securities is guiding Unitree toward a Shanghai listing valued at more than 50 billion yuan—about $7.4 billion—with a raise of roughly 4.2 billion yuan earmarked for research, development, and production.
The second, buried in the company's own disclosures: through the first three quarters of 2025, more than 70% of Unitree's humanoid revenue came from scientific research and education, and industrial applications accounted for just 9%.
So here is the question public investors should be asking before they wire the money: at that valuation, what exactly are they buying?
Last week's thread on this beat tracked the overcapacity problem in the abstract—too many production lines chasing commercialization claims nobody could verify. This week the abstraction has a price tag, a prospectus, and a second candidate: Agility Robotics is going public via SPAC, the first pure-play humanoid firm to do so. The question is no longer whether these companies say they've commercialized. It's whether the public-market documents will contain a number that proves it.
Start with what the market is actually being sold. The dominant metric in every headline is production volume.
Agibot announced its 15,000th robot rolling off the line in June. Unitree cleared roughly 11,000 units in March and plans to expand annual capacity to 30,000. Owen Chen, the company's Asia-Pacific director, reportedly says shipments this year are expected to at least double against 2025's 6,500-plus units and could nearly triple.
These are real manufacturing feats. Building sophisticated hardware at scale is genuinely hard, and China's two humanoid oligarchs are doing it faster than anyone predicted.
But production capacity is being priced as if it were deployment—as if a robot leaving the factory were equivalent to a robot doing paid work. Unitree's own revenue mix says it isn't.
A robot that ships to a university lab or a classroom demo counts toward the production figure. It does not count as a machine that has displaced a wage or beaten an existing automation line. It is inventory that happens to be sitting somewhere impressive.
The 36Kr analysts covering WAIC 2026 put it precisely: a large number of shipped robots "have not really entered the factory to work but stayed in laboratories and classrooms." Production growth, they note, "is not directly equivalent to the in-depth advancement of industrialization."
The tell shows up in the gap between production and shipment rankings. UBTech ranks third in production but slipped to fourth in shipments—a spread that "usually means inventory backlog or delayed delivery rhythm."
UBTech's consumer story compounds the point. Its U1 humanoid logged over 13,361 cumulative orders. But those orders came largely from demonstration halls and government programs, not private buyers.
That is the entire problem in one line item. An order from a demonstration hall is an order to demonstrate. It is not evidence that the robot pays for itself doing anything.
An order from a demonstration hall is an order to demonstrate. It is not evidence that the robot pays for itself doing anything.
Which brings us to the argument that reframes the whole debate. Shaun Edwards, co-founder and CTO of Plus One Robotics, makes the case that the industry has been looking at the wrong wall.
The wall is not AI. Vision systems have matured, foundation models keep improving, actuators and hands are close enough that plenty of use cases already work. As Edwards frames it, the conversation has moved from "Can they do the work?" to "Can we justify the cost?"
The wall is unit economics. And he offers a killer proof from his own domain: parcel sorting.
The best humanoid demonstrations hit around 1,000 picks per hour in controlled conditions. Specialized parcel-sorting robots run between 1,300 and nearly 3,000 picks per hour.
That is not a rounding error. It is a machine losing to the incumbent it was supposed to replace, on throughput, before you even discuss price.
And the price is the second blade. Edwards cites one leading humanoid developer pegging cost at several hundred thousand dollars per robot.
Spend that, he argues, and you need a multi-shift operation running the machine around the clock before the math works. But if you already run that kind of operation, you can put the same money into specialized automation that gives you more throughput and pays back faster.
This is the sentence that should haunt every humanoid prospectus: in high-volume production, specialization beats the general-purpose machine. The flexibility humanoids sell as their core advantage is precisely what makes them slower and dearer than the fixed system optimized end-to-end for one workflow.
So the survival metric is not units produced. It is cost-per-deployed-robot measured against the specialized automation it must beat—and, downstream of that, an auditable cost-per-task or utilization figure.
Notice which number that is. It is the one none of the IPO-stage firms will publish.
They will publish production milestones. They will publish valuations. Agibot will livestream its G2 robots running a factory for six days and report a success rate that, per the demo's headline, reached 99.99% across over 64,000 tasks—though the coverage body describes it as nearly 100%.
That livestream is worth examining, because it is the archetype of what gets shown instead of what matters. A near-perfect success rate is a reliability figure. It arguably says the robot rarely fails at the task.
It arguably says nothing about picks per hour versus the specialized line, nothing about the fully loaded cost per successful task, nothing about how many human supervisors the deployment required. Edwards' own model assumes people stay in the loop as high-level supervisors—one person overseeing 50 or more robots, with intervention that drops over time but never hits zero. That labor is a cost. It belongs in any honest unit-economics disclosure. A 99.99% highlight reel routes around it.
Agility, to its credit, is positioned differently. CEO Peggy Johnson declined to offer financial forecasts, emphasized industrial safety certification over lab demos, and gave a refreshingly unglamorous timeline of ten-plus years before humanoids are common in homes.
That measured posture is the right instinct. But instinct is not disclosure. A SPAC merger is being framed as a rare window into the sector's real financials—so the test is whether the window actually opens onto a deployment number, or onto another slide of vetted-customer pipeline and manufacturing ambition.
Because the manufacturing-scale thesis is real and it cuts both ways. Edwards is right that whoever can build hundreds of thousands of units efficiently drives component costs down and moves the economics—the Tesla-and-Optimus argument. Scale is the path to viable prices.
But scale is also the thing that lets a company report enormous production figures long before a single one of those units clears the ROI bar against a conveyor and a fixed-cell arm. The same 30,000-unit line that will eventually lower cost is, today, the line producing inventory the market is being asked to value as deployment.
There is one disclosed number that at least reads honestly. Unitree warned in its IPO filing that the FCC's addition of foreign-produced advanced robotic devices to its Covered List could affect future access to the US market, which has been a significant share of its overseas revenue.
That is a real risk factor, disclosed as a risk factor—which is exactly the standard the deployment numbers should meet. If a company can quantify a regulatory threat to a revenue line, it can quantify how much of that revenue line is paid deployment versus classroom placement.
So here is the disclosure test, and it is simple. The first S-1 or prospectus that publishes an auditable cost-per-task or utilization figure—not a production milestone, not a viral demo, not a 99.99% success rate—is the only one that has earned a public-market premium.
Everyone else is asking you to underwrite a factory and call it a business. Unitree's own 9% tells you those are two different things.