Unitree Robotics: Should We Be Afraid, or Very Afraid?

Where this began. My AI assistant and I read Singer and Zvenyhorodskyi's essay "Chinese Robots Could Decide the AI Race," in Foreign Affairs of October 12, 2026, and sent Vista's research desk the question it raised for us: if China wins the robot's body, who gets paid for what the robot learns? The desk tested the question, and this inquiry grew from that exchange.

The idea

The way Singer and Zvenyhorodskyi tell the story, Washington and Silicon Valley bet on scaling software while Beijing fuses software with hardware into what the field calls embodied AI; citing Chinese scientists, the authors write that China's manufacturing base "can become a training ground for AI," and they close on the fear that the United States may build the best models while China reaps the benefits of "bringing the models to life."

It is a serious essay, worth reading to the end. The question it raised for me was narrower: if China wins the robot's body, who gets paid for what the robot learns? The inquiry grew from a report I developed with my AI research assistant; it tests the essay's assumption, that whoever deploys the robots keeps the learning, not its facts, which, the FCC rule aside, hold up.

I should say where I stand, because it colors the inquiry. My whole adult life in the world of experts and professors has been spent around allegations of Chinese theft of intellectual property. Here, refreshingly, at least as far as the public filings and records let me see, is a story with no theft in it: Chinese robots sell openly to American labs, and an American company trains its models partly on what those robots record. So far, everything I can see moved by sale or license; the test below will tell us whether that lasts. That, to me, is what makes America so great: we don't have to break any laws; we can win by inventing, as the American model maker in this story is doing on robots anyone may buy. I hold that view with conviction. And yet, for all that, I'm suspicious. What I worry about is that some of the learning may also move through channels no filing shows, and this inquiry returns to that in its seventh complication. So I hold this assessment of the race as a hypothesis, an honest bet that the lawful flows I can see matter more than any hidden ones I cannot, with Unitree's own plans given their full weight and the tests below that could prove it wrong.

The spectacle maker

In October 1608 the States General, the national government of the Dutch Republic, met in The Hague to consider a patent application from Hans Lipperhey, a spectacle maker, for a tube with two lenses that made distant things look near. According to the Galileo Project at Rice University, the gentlemen "found the device too easy to copy to award the patent." A device any lens grinder could make is not a device the state can reserve for one man. They paid Lipperhey handsomely to make several two-eyed versions instead.

The next summer Galileo heard a description of the thing and built his own, and by the fall of 1609 he had a twenty-power instrument that he turned to the sky. In March 1610 he published Sidereus Nuncius, the Starry Messenger, with the surface of the Moon and four moons of Jupiter in it. Lipperhey made the instrument. Galileo kept what was found with it.

I do not say that Unitree is Lipperhey. The story is the oldest clean example I know of the rule this inquiry works out: the maker of an instrument that is easy to copy earns a maker's wage, and the discoveries belong to whoever has the scarce thing the instrument needs, which in 1610 was Galileo's eye and pen, and in 2026 is the AI model and the rights to the recordings. Carl Shapiro and Hal R. Varian put the economics in one line in Information Rules (1999): "Information is costly to produce but cheap to reproduce."

Who Unitree is

Unitree's founder, Wang Xingxing, was born in 1990. He studied at Zhejiang Sci-Tech University, then earned a master's in mechanical engineering at Shanghai University, where as a student he built a four-legged robot he called XDog. He was there, by his own account, from 2013 to 2016, the years I spent much of my time at Shanghai University recruiting professors as expert witnesses and consultants while trying to open the Chinese expert witness market (this endeavor did not succeed); our paths may have crossed. After a brief spell at DJI, the drone maker, he founded Unitree in Hangzhou in August 2016. Its first product was a quadruped, the Laikago, a robot the size of a dog; the AlienGo followed in 2019, and the company made its name selling such machines to universities and to companies that use them for inspection. TIME put him on its 2025 list of the most influential people in AI and, in July 2026, on its cover beside Unitree's GD01, a manned mecha. He controls about 65% of the company's votes, through shares that carry ten votes each. Humanoid revenue begins in 2023.

In 2025 Unitree's revenue was RMB 1.70 billion, about $254 million, 4.3 times the year before; humanoids were more than half of that. Its net profit was RMB 278 million, after a one-time charge of RMB 349 million that the accounting rules require when a company sells its employees shares cheaply. Profit excluding one-time items, the measure Chinese filings use for what the business earns in an ordinary year, which I will call recurring profit, was RMB 591 million, and operating cash flow, the cash the business took in less what it paid out to run itself, was RMB 670 million. Hardware made, sold, paid for. The humanoid line's gross margin fell from 69% to 63% in 2025, as more rivals entered and Unitree cut prices to hold its position.

The company went public on August 19, 2026 on the STAR Market, a section of the Shanghai Stock Exchange for technology companies, announced in November 2018, opened on July 22, 2019 and nicknamed China's Nasdaq. The board's largest planned offering never traded: Ant Group's, set for Shanghai and Hong Kong, was halted on November 3, 2020, two days before its first day. Unitree's first day came. It sold 10% of itself at RMB 150.80 a share, about $22.53, raising about RMB 6.1 billion, about $911 million, before costs. The shares closed their first day at RMB 845, up about 460%. A buyer at the offering price paid, by the filing's own multiples, about 93 times the company's 2025 recurring profit and about 36 times its sales.

The unaudited first half of 2026 shows the shape of what is coming. Revenue grew 49% against the first half of 2025, to RMB 1.15 billion. Recurring profit fell 19%, to RMB 244 million, as research and selling expense rose sharply; operating cash flow fell 33%, to RMB 232 million; and capital spending was 3.7 times the year before.

Who buys a humanoid

This is one of the questions that moved me to write this inquiry. I am in the robot retail business myself, proudly a founder of Steel Collar Robotics and Aeronautics, an independent robot and drone dealer and robotics research think tank, and a dealer learns what people actually want to buy. To date not one customer has bought a humanoid or even asked about one, and we have never bothered to try to convince anyone to. Steel Collar is not a Unitree dealer: we hold no agreement with Unitree and have never applied to its distributor program. Our equipment guide does describe three Unitree models for buyers who ask: the B2, a four-legged inspection robot, always with the note that it is built in China and that its maker is on the Defense Department's list of Chinese military companies; and the G1 and H2 humanoids, which we never offer. One new dealer's ledger is an anecdote, not a market, and I offer it as nothing more; but it agrees with Unitree's own count, which follows.

The filing that answers this question is Unitree's March 2026 reply to the Shanghai Stock Exchange's questions before it went public, covering January to September 2025. Of humanoid revenue of RMB 595 million in those nine months, research and education, which in Unitree's definition includes developers who buy the robot to build on it, took 73.6%; commercial consumption, meaning robots bought for promotion, performances, exhibits, the home and rental, took 17.4%; and industry took 9%. In a second reply to the exchange the same month, it said that by a preliminary tally RMB 15.70 million of those nine months' humanoid sales, about $2.35 million, went to identified industrial work, smart manufacturing, inspection and logistics: 29.3% of the industry line and 2.64% of all humanoid revenue. The rest, it said, it understood to be mainly enterprise tours, robots that greet visitors in a company's exhibition hall; hence its estimate that tours were 50% to 70% of industry sales. It added that it does not fully know how its industrial customers use the robots. The subtraction, in other words, is the company's own.

The four-legged robots tell a different story in the same reply. Research and education fell from 68.61% of quadruped revenue in 2022 to 31.58% in the first nine months of 2025, the shift driven, the company says, mainly by consumer robots sold online. If the words robot dog bring to mind "Metalhead," the 2017 Black Mirror episode shot in black and white in which a woman is hunted across a ruined landscape by a relentless four-legged machine, I understand; I found it genuinely frightening. Unitree's working dogs, by the filing's account, lead a duller life, inspecting sites, going out with fire and emergency crews, surveying rough ground and carrying loads in factories, and somebody pays them for it. The quadrupeds have left the lab. The humanoids remain lab instruments.

In August the Financial Times found a buyer the filing's categories hide. Much of the makers' revenue, William Langley reported, comes from government-backed training centers, which buy the robots, have workers steer them through tasks, and sell the recordings back to the makers; Interact Analysis, a market research firm, counts more than 90 built or under way across China by June. This looks like the Foreign Affairs case at work, and it is the be-afraid verdict in miniature, government money, robots and cameras in one room and the learning staying home: a senior center manager said each maker's data can be used only by that maker, and one seller priced the training data for a five-minute robot dance at up to RMB 1 million, about $149,000. Three things weaken it as evidence of a real-world data advantage. The hours are lab hours: the same manager said only two or three of an eight-hour shift's hours are usable, on average, and Marco Wang of Interact Analysis told the FT the data is not fully useful because the robots are not in real-world settings. The money is largely public: at a state-funded center in Liuzhou, three staff told Reuters the subsidy-dependent project had no clear path to profit, and Reuters counted at least $230 million of government purchases of humanoids and related items in the first half of 2026. And the exposure is uneven: a relatively small share of Unitree's shipments went to the centers, analysts told the FT, against 45% of Leju Robot's Kuavo humanoid sales in 2025. So our hypothesis gains a clause: the hours these centers keep in China are, so far, lab hours bought largely with policy money, and real-world hours go to whoever deploys or buys them. The frozen test cannot measure it; the interviews below could.

One more number from the first reply matters later. More than half the money research buyers paid in those nine months, 56.6%, came from customers who had bought before, in any year back to 2022, against 59.4% in 2024 and 49.7% in 2023; the figure pools humanoids with four-legged robots and counts money, not customers. The labs come back.

What the filing admits about learning

Here is the sentence this inquiry turns on, a warning Unitree printed in its prospectus, in our translation from the Chinese: "with the global technology route and data scenarios not yet settled, during the reporting period the company had not yet conducted real-world data collection and factory-deployment training on a large scale." The passage around it says the company's research money went first to the robot's body and to motion control, and that from 2024 it began to spend more on what the filing calls the brain, the embodied large model, embodied meaning AI that controls a physical body, though that spending remained a small share through 2025. Research expense was RMB 145 million in 2025, under a tenth of revenue.

Read it with care. The reporting period is 2023 through 2025, and the sentence says "not yet on a large scale," not "not at all": the words of a company that means to collect, which is where the be-afraid verdict begins. The company has tested its general model in its own factory. Of the money raised, the largest named use, RMB 2.02 billion over 36 months, about $302 million, goes to a model research project, about fourteen times the whole of 2025's research budget. And in its reply to the exchange the company described the data flywheel it intends to build: more robots at work make more recordings, which make the software better, which sells more robots, with a mechanism on deployed robots that records each failure, sends it back to be labeled and feeds it into training.

The strongest counterweight is a single sentence in the second reply to the exchange. It says that at present the company tests in industrial, household and other settings and at the same time "collects several hundred thousand to several million hours" of embodied training data a year. It does not say whose hours they are; it does not say whether the figure is a count or a target; and it does not reconcile with the prospectus's statement that large-scale collection had not yet begun through 2025. If those hours become a business, a line of income from data, models or software will appear in the accounts, which the second measure watches; that is the be-afraid verdict, Unitree in the data business, and the measure that would reveal it.

What is not in any of Unitree's offering documents is a customer contract term that gives Unitree rights to the recordings its customers' robots make; we searched them for every Chinese phrase for data ownership, telemetry, consent and data security near every description of a contract, and found none. (My wife, who speaks fluent Mandarin, declined to help, so I had to rely on AI for the translations.) Absence from a searched text is not proof, and the test treats it so. But it means the flywheel is, today, a plan, written in the future tense and paid for with offering money. Nor is a contract the only way a recording could travel; the seventh complication takes up the way no filing shows.

The maker is not idle on the open side: Unitree has released most of its AI models under open licenses, and its latest model, released in September 2026 and trained on about 2,500 hours of recordings from real robots, may be used commercially by anyone. That builds the ecosystem around the robot itself, as an instrument maker should. It does not, by itself, earn what economists call a rent: profit above the ordinary, earned because others cannot copy what you have.

Teece's framework: who keeps the learning

In 1986 David Teece published the paper that economists still use to answer the question this inquiry asks: why the people who make a breakthrough so often fail to profit from it. His answer had two parts. The first he called the appropriability regime: how hard the innovation is to copy, because of patents, secrecy or the sheer difficulty of the thing. "When imitation is easy, markets don't work well," he wrote, and the profit "may accrue to the owners of certain complementary assets." The second part was those complementary assets, the things an innovation needs to reach a customer. When imitation is easy and the complement is scarce, the owner of the complement takes the larger share.

I should say again, as I did in Inquiries No. 12 and No. 16, that David has been a friend for more than thirty years, one of the first scholars to join the firm I founded in 1994; I do not cite him by design, his work keeps turning out to be the right tool, and he has had no part in this inquiry.

Vista's Library writes Teece's rule as a bargain in the manner of John Nash, the mathematician who showed how two parties split a gain. When a model maker learns from a body maker's robots, a gain is created that neither could have made alone; each side keeps what it could earn without the other, its fallback, and the two split what is left. Scarcity sets the fallbacks. While few robot bodies suit a model maker's software, the body maker's fallback is high, because it can sell the same robots to the next model maker, and the model maker's is low, because its models would have to be retrained on another body; once bodies are interchangeable and models move easily between them, the fallbacks reverse. In the illustration we built, with a learning gain worth 100 units and equal bargaining skill, the body maker keeps about two thirds of the gain today and almost nothing once bodies are generic and models portable. Every input is an assumption; the shape is the point. The illustration also assumes the learning moves only by sale and license; if some moves by a hidden channel, the split is drawn from the visible part alone.

I know the weak appropriability regime from the inside. Years after I founded my first company, an expert network that located scholars for the lawyers and companies that needed them, our best clients began sharing emails from new competitors offering the very service we had pioneered, in language eerily like our own. As I wrote in The Scholars' Treasure, it was a harsh reminder of the volatile nature of entrepreneurship, "where the fruits of our labor can be so easily appropriated by others" (ch. 12, p. 148). Unitree is in that position with a product that can be photographed, measured and copied, and it says so: once the supply chain matures, it told the Shanghai Stock Exchange, the hardware differences between robots should narrow, and algorithms, data, ecosystem and services will become the points of competition. Its complements today are its low cost of making robots, the community of developers who write software for them, and its dealers. The question this inquiry asks is whether the learning is one of them.

Inside the box: where the profit went

The plainest test of Teece's rule I know is a teardown. In 2008 Jason Dedrick, Kenneth Kraemer and Greg Linden took apart Apple's iPod and an HP notebook computer and traced where the profit on each went, using Teece's framework as their lens. In the HP notebook of the mid-2000s, Microsoft's Windows cost HP $100 and Microsoft kept $85 of it, a margin of 85%; Intel's chips earned about $121; HP, which specified the machine, put its name on it and sold it, kept a gross margin of 28% on the whole box. In dollars HP earned the most; the two parts it bought in earned the richest margins, because nearly every notebook maker had to buy them.

The notebook's economics were set a quarter century earlier, when IBM built its personal computer in a year from outside parts, Intel's processor and Microsoft's operating system, and published its design. Teece wrote the IBM PC into his 1986 paper as a winner, a follower that won through the complementary assets it had or quickly assembled, and his framework says what came next as well: once rival makers could buy the same chip and the same operating system, those two became the scarce complements, the profit followed them out of the box maker's hands, and in December 2004 IBM announced the sale of the business to Lenovo, a Chinese company, for about $1.75 billion.

On a personal note, I was 13 in 1983, two years after IBM's machine arrived. I already had an Atari 800, spent hours programming it (and playing video games like Frogger), and even attended Atari Computer Camp in 1984. I was among the first generation of avid computer users and programmers, but too young to make any damn money off the PC revolution. The money, it turned out, was never in being early to the machine; it was in owning the parts the machine could not do without, which is what the teardown shows in margins.

The same teardown holds the counterexample, and it matters below: Apple's video iPod of 2005, sold to the trade for about $224 against a factory cost of about $145, kept $80, a gross margin of 36%, because Apple controlled the core software and the iTunes software and store, and so kept the profit HP had to hand to Microsoft and Intel. A humanoid robot maker that owns the brain and the store is a different animal from one that buys them in.

NVIDIA plays both parts

In December 2025 NVIDIA's GEAR lab released GR00T N1.6, an AI model meant to be adapted to many humanoid bodies, and its page says the pretraining data include several thousand hours of teleoperated recordings from, among others, AgiBot's Genie-1 robot and a Unitree G1; the page gives no hours per robot and says nothing about rights. N1.7, released in July 2026, carries the first commercial grant, and NVIDIA's developer page still lists the Unitree G1 among the bodies it learned from. On June 1, 2026 NVIDIA announced a reference humanoid robot for academic research, built on a Unitree H2 Plus with an NVIDIA Jetson Thor computer inside; the release adds that researchers keep control of their robot data, training data, telemetry and logs, though NVIDIA Research will use the same design to advance its GR00T models. Jetson Thor, launched in August 2025 at $3,499, is the robot brain NVIDIA sells to everyone.

In the humanoid, NVIDIA sells the chip and trains the AI model, in part on the Chinese bodies that make the research cheap: the two complements of the IBM story in one pair of hands. This is our assessment of public documents, not a finding about any contract; NVIDIA does not disclose what it pays for recordings or earns from robot models, and Unitree discloses no payment from NVIDIA. The documents describe, so far, an American model maker learning from Chinese bodies, on terms nobody has published.

The complications

An inquiry worth the name faces the arguments against it, and there are quite a few; we found seven.

First, Unitree may not be the leading body: industry trackers relayed in the trade press put AgiBot ahead in humanoid units, though their counts disagree with each other and with Unitree's own figures. If the bodies are becoming generic, the lead among body makers is exactly the thing that should be unstable.

Second, the closed loop. Nomura, a Japanese investment bank, in a report relayed by a Chinese industry site, describes Unitree's business as a loop in which low cost drives volume and recordings from that volume feed the company's own models; whose recordings matters: the filings describe Unitree's own collection, collection with partners, outsourced collection and open-source data, and failure recordings uploaded from deployed robots without saying whose, and give it no contract right to its customers' recordings, American labs' included. Unitree's second reply to the exchange makes the same case: within three years, a general humanoid model that closes the loop of training in the cloud, inference on the robot and collection in the field. Drawn as a circle, the loop has four turns, cheap robots sold, more robots at work, more recordings, a better model that sells more robots, and learning crosses it in three places: out by sale, out by license, and in or out by a channel no one can see. Bloomberg's Saritha Rai reported in July that in China models are increasingly trained on recordings from humanoids working in the places they will serve. The training centers above add lab hours, and in August Xiaomi open-sourced a robot model pretrained, it says, on more than 100,000 hours of real-world demonstrations recorded with a handheld tool, not by a robot. We do not dismiss it; we date it, and the second test says what the annual report would have to show for the loop to be visible in revenue.

Third, the contracts on the public record point both ways. In May 2026 Hyperscale Data, a publicly traded American company, filed with the SEC its subsidiary's agreement with AgiBot: the buyer may take up to 143 robots for up to about $13.4 million, owns every record the robots make and may sell it, and AgiBot may not see the data or train on it. One negotiated contract is not a market. Boston Dynamics' published terms of sale for its Spot robot point the other way: the customer agrees that the maker "may retain indefinitely" the terrain and image data its robots record, and "may access and use such data in perpetuity." Negotiated, the body maker is left with the sale and no data; on its own standard terms, it keeps a license to learn.

Fourth, the recordings may not be scarce in the way the thesis assumes. Psi0, a University of Southern California paper this year, reports training a humanoid model for the Unitree G1 with about 800 hours of video of people and 30 hours of robot recordings, and beating standard models trained on more than ten times as much; NVIDIA itself pretrained N1.7 on 20,854 hours of video of people from its EgoScale dataset, beside the robot recordings. If video of people substitutes for robot hours, the scarce complement is the skill of whoever designs the training, which moves the rent toward the model maker and the integrator and away from both the robot and the raw recordings. And real-world hours can be bought far from China. On October 19 The Wall Street Journal reported that CNTXT AI, a data company from the United Arab Emirates that claims $300 million of committed revenues, mostly from the biggest American cloud computing companies, AI developers and large enterprises, is opening a Silicon Valley office; it also sells robot-training data recorded, it says, from workers in the Gulf's hotels, restaurants and warehouses. An independent collector that charges for its recordings may keep part of the value this inquiry assigns to the model makers.

Fifth, the door to the American market. On July 28, 2026 the Federal Communications Commission added foreign-produced advanced robotic devices to its Covered List of equipment judged a risk to national security. A new model on the list cannot get the FCC authorization it needs to be imported, marketed or sold here; models already authorized may still be sold, with software updates allowed under a waiver to at least January 1, 2029. The rule names no country, though the evidence behind it cites Unitree, and the notice has no research exemption. An older import rule, 47 CFR 2.1204, lets up to 4,000 unauthorized units of a model in for testing and product development; the FCC calls that a loophole and has proposed a cap of 40. Singer and Zvenyhorodskyi call that an exemption for robots imported for American product development; more exactly, the notice is silent, the older allowance may still apply, and the commission proposes to shrink it a hundredfold. Unitree's six current main models hold FCC certification, so the action freezes the American catalogue at the 2026 models. A pending bill, the GUARD Act introduced in the House on June 3, 2026, would put humanoid and four-legged robots from adversary countries through a security review and add those found risky, or not reviewed within a year of enactment, to the Covered List.

Sixth, DeepSeek. China's best known AI model maker bought Unitree shares in the offering under a 36-month promise not to sell, and holds, by Unitree's account at its roadshow, a strategic cooperation memorandum: joint research toward general AI, including AI that controls a physical body, and a preference, on equal terms, for Unitree's robots when DeepSeek builds such AI. Whether any data arrangement sits behind it is not disclosed. If the learning from Chinese bodies flows to a Chinese model maker as readily as to an American one, the essay's authors are right about the direction after all, though the investor's version does not change: the rent still sits with the model maker, and Unitree still discloses no income from it.

Seventh, what no filing can show. I said at the start that I am suspicious; here is why. Every channel above is one anyone can see: a sale, a license, a contract, a line in an annual report; what I suspect is a channel nobody can see, and it splits this inquiry's thesis in three. The company question, whether Unitree earns from the learning, largely survives it: if there is devilry afoot, if the state quietly takes the recordings by some insanely surreptitious technology, that is value Unitree does not book, and the test below still reads the accounts rightly. The national question, whether the learning flows to America, is where the verdicts part. So long as the learning moves by sale and license, or stays home through training centers, a memorandum and a data business the accounts will eventually show, the answer is either not afraid or be afraid, and the six measures can tell which. But if Chinese robots in American labs, or anything in the chain that sells and services them, send recordings home, the learning flows the other way unseen; a Chinese maker learns from what American labs do with its robots without owning a single recording; the Chinese state gains intelligence that no one books as income; and no filing or test here can see it. Under that scenario China wins big, and this inquiry's assessment of the race is wrong. This, dear reader, is the be-very-afraid scenario.

There are four things that make me lean toward taking that scenario seriously. None of them proves it; together they make a hidden channel possible, not shown. China's National Intelligence Law of 2017 says, in Article 7, that "any organization or citizen shall support, assist, and cooperate with state intelligence work according to law," in the English rendering Murray Scot Tanner gave in Lawfare that July; the law, he noted, nowhere explicitly lets those asked refuse. One technical report describes data leaving a Unitree humanoid. In September 2025 Víctor Mayoral-Vilches of Alias Robotics, a robot-security firm, published a study of a Unitree G1, one author's work and not peer reviewed; it found two services holding a persistent connection to Unitree's servers from seconds after boot, servers the report places in China, and sending every 300 seconds a message of the robot's state: battery voltages and temperatures, its tilt and heading, the positions of more than twenty joint motors, the services running. Its summary says audio and video are sent, its body that they are gathered for potential transmission; the traffic it captured was that state telemetry, not sound or pictures. On September 29 Unitree told its users that its robots are built for offline use and connect only with a customer's authorization, that basic product information such as serial number and health status may then go to servers in Singapore or local servers, and that without authorization it collects no private or sensitive data. Washington has already acted: in June 2026 the Defense Department added Unitree to its Section 1260H list of Chinese military companies, on grounds of indirect state ownership and state support, not data; the FCC's July action follows a national-security determination that foreign-produced robots present "unacceptable supply chain and cybersecurity vulnerabilities," with evidence that cites Unitree, and the GUARD Act is pending. And the support for Unitree that is on the books is real: Singer and Zvenyhorodskyi note that embodied AI is named an industry of the future in China's current Five-Year Plan; China's national social security fund and arms of state-owned oil, power and telecom groups took shares in the offering; government-backed training centers buy the robots; and DeepSeek holds a stake and a memorandum. Whether state favors also pass off the books, letting Unitree earn from the learning unseen, can be neither ruled out nor confirmed from filings. I do not allege that any of this is happening. I say that the United States government has acted as if it might, that one security firm's report documents a channel on one model, and that the lawful flows this inquiry can see are the only flows it can test.

What the price assumes

We do not use Unitree's share price beyond its first day. The offering price's multiples above, about 93 times recurring profit and 36 times sales, already ask a buyer to believe either that the hardware business grows many times over at its present margin or that a learning business appears in the accounts. Can an American reader buy it? Not easily: Hong Kong's Stock Connect program, the usual route for foreigners, does not include Unitree as of October 16, 2026, a new company needs a year of trading to qualify through the index, with further tests for weighted votes like Unitree's, and only institutional professional investors may trade such shares. For most American investors Unitree is a thing to watch, not to own.

Where the framework points in publicly traded markets

Each hypothesis below carries the evidence for it and the evidence that would falsify it. None is a recommendation. The dates are chosen so that each can fail.

NVIDIA, the two complements. The evidence is the section above: the GR00T training data, the reference humanoid, the commercial grant. The hypothesis is that NVIDIA keeps both bottlenecks of the humanoid, the chip and the model, through 2027, and that the learning from Chinese bodies keeps flowing to it; falsified if NVIDIA's next GR00T model after N1.7, released by April 30, 2027, names its real-robot training bodies and neither Unitree nor AgiBot is among them. Untested if no such model appears by then.

Unitree, the lab bench. The evidence is the buyer mix, the margin and the cash above; the hypothesis is that Unitree is, through 2026, a profitable instrument maker whose learning business remains a plan; its five tests are the frozen test in the next section.

UBTECH and DOBOT, the integrated route: two Chinese robot makers traded in Hong Kong that sell the robot with installation and service, UBTECH's buyers including the training centers above. In the first half of 2026 UBTECH reports RMB 590 million of humanoid revenue on 921 units with a group loss of RMB 339 million and an operating cash outflow of RMB 749 million; DOBOT reports revenue of RMB 316 million, a loss of RMB 108 million and an operating cash outflow of RMB 259 million. Their robots work at customers' sites, so they could collect working recordings first; the hypothesis is that this route consumes cash before it earns any learning rent, and that a maker keeping the learning would show up as cash at one of these two before it shows up at Unitree. Falsified if either reports positive operating cash flow for fiscal 2026 with humanoid revenue growing.

Tesla, the integrated alternative. Tesla's quarterly filing for the quarter ended June 30, 2026 says it is "capitalizing on our strengths in real-world AI data to advance the development of Optimus," its humanoid. Tesla owns the body, the brain and the first customer, its own factories: Apple's iPod position. It is a watch item rather than a test, because Tesla discloses no unit counts, no Optimus revenue and no data terms: the divided outcome would be challenged if, in its annual report for 2027, Tesla reports Optimus doing paid work for customers outside Tesla, with revenue disclosed, before any model trained on bought Chinese bodies does the same. Tesla is also the one maker in this inquiry that need not share: Optimus will not be roaming Chinese labs any time soon, so whatever Tesla learns stays home; and, if I may say so, the learning a model gathers in Chinese labs is not the same as the learning gathered in America's leading ones.

The American catalogue. The hypothesis is that the lab bench in the United States shrinks in share: the American share of Unitree's revenue, 13.3% in 2025 and 19.5% the year before, falls further in 2026; falsified if the annual report shows it at or above 13.3%, which would mean the certified models and the testing allowance kept the bench open. Untested if the report gives no regional split.

The frozen test

What follows was frozen on October 16, 2026, written into a program with a digital fingerprint so the lines cannot move. It is scored from Unitree's annual report for 2026, due under the STAR Market's rules within four months of the year's end, so by April 30, 2027. Three rules, given in full in "How the figures were made," govern every measure: a measure the report does not disclose scores untested, never as absence; a range is scored at both ends; and each test is reported on its own, with no overall verdict that adds them up.

Every measure sits in the support zone today, as it should: the test asks whether the picture holds. Identified industrial work would have to rise almost fourfold to leave support and more than sevenfold to falsify. The humanoid margin has 8 percentage points of room above the support line after falling 6 points in 2025. Cash coverage is the measure moving fastest, from 1.13 in 2025 to 0.95 in the first half of 2026 as capital spending rose 3.7 times, and the company's overall gross margin in the half, 56.0% against 60.2% a year earlier, says to watch the humanoid margin too.

T1 and T5 appeared only in the March 2026 replies to the exchange; if the annual report does not repeat them they score untested.

Two signposts come before the score, and neither is scored. The third-quarter report, scheduled for October 29, gives the first nine-month figure for the cash measure. And the FCC may cut the testing allowance for covered robots from 4,000 units per model to 40; we will record which way it went beside the scores.

A second leg, unfunded. The report this inquiry grew from proposed training the same AI model on recordings gathered four ways, from customers' robots, a standardized lab, people demonstrating with handheld tools, and simulation, and testing each at customer sites it has never seen, assessed by October 14, 2027. No partner, permission or funding exists for it; if it is never run, its claims stay untested, and the public leg above carries the inquiry.

What is new here, and what is not

The facts are public. Singer and Zvenyhorodskyi posed the national question; this inquiry tests the assumption inside it against the filings of the very company they cite as China's example. Marco Wang of Interact Analysis told the Financial Times in May that only about 9% of Unitree's shipments went to real-world uses; the FT's William Langley and a Reuters team examined the training centers in August. What we add: Unitree's own count set beside its repeat-buyer figures and frozen as a test; Teece's split set against the robot body and the robot model; NVIDIA read as the holder of both of the PC's complements at once; the observation that, as far as the record shows, the learning has so far moved by purchase and license; six thresholds frozen in code before the first annual report, with untested kept apart from absent; and the channel no filing can show named as the limit of these tests. The be-very-afraid scenario is not mine alone; the FCC, the GUARD Act's sponsors and one security firm have each raised it in their own terms. What I add is that I don't actually believe it. The United States government, and the companies whose models would be poisoned or copied, have every reason to look for such a channel and the means to find it.

What it means for an investor

Separate three questions the headlines run together: who profits from selling the robot, who learns from it, and who keeps the surplus of the learning. If the body stays an instrument and the learning stays with the model makers, the investable expression of Chinese robotics is not Chinese at all; it is the owner of the chip and the model, and the Chinese body maker earns a maker's margin that competition will press on. For a reader in Washington the lesson is the same one turned around: as we read the public record, so far China sells the robots and an American model maker learns from them, openly, by purchase and by license. I hold that as a hypothesis and not as a boast, and I can name the three ways it fails: if Unitree's RMB 2 billion buys a model its customers pay for, which the second test would show; if the learning flows to DeepSeek, or to the Chinese state's intelligence services, as readily as to NVIDIA; and if Washington shuts out new Chinese models, as the FCC began to in July.

Set the three verdicts beside the money. I think NVIDIA wins financially under not afraid, because the learning from cheap bodies flows to the owner of the chip and the model, and it wins even under be very afraid, because a hidden channel copies what NVIDIA has learned without taking away its chip, its model or its customers. The Chinese robot makers win under be afraid, and only there: if they go into the data business, and fast, because, as I suspect though cannot show, China has many more labs putting robots through tasks than America does. The six measures say which world we are in. Industrial work beyond tours, the first, and the repeat buyers, the fifth, rising together mean the robots are working and the recordings exist; a data, model or software line, the second, means Unitree is selling them, which is be afraid; the humanoid margin, the third, and cash against profit, the fourth, holding mean the instrument maker's business is whole, which is not afraid; and NVIDIA naming Chinese bodies in its next model, the sixth, means the learning still leaves by license, not afraid again. No measure can show be very afraid. That is why the sixth missing number and the interviews below come before the scores.

Then there is the cheapest part of the picture. Under the assumptions of the report this inquiry grew from, a humanoid at Unitree's average price pays for itself only where an hour of labor costs more than about RMB 103, about $15; cut the robot's price to RMB 20,000, about $3,000, and the hardware becomes a tenth of the yearly bill, and at RMB 50 an hour, about $7.50, even a free robot loses money. Once the humanoid robot is cheap, its hardware is not where the economics are decided; that is the whole of Teece in one sentence, and it is why who owns the learning is the question.

There is a line in the book of Joshua that I keep coming back to when I think about instrument makers, spoken by God, through Joshua, to a people already living in a land they did not build: "cities which you didn’t build, and you live in them" (Joshua 24:13, WEB). It is a verse about grace and inheritance. I do not stretch it. But the model makers of 2026 are training in cities they did not build, on bodies cheap enough for a university lab to buy, and the inheritance is going to be settled by contracts, not by who made the lenses.

The data that we're missing is...

Likely, we can get some of it from primary research, which is what Vista is all about. Six numbers would settle the most.

• The first is the share of humanoids that do a shift. Unitree's filing gives revenue by end use and its own count of identified industrial work; nobody has published what share of the robots sold to industry run a task on a schedule without a person steering, and for how many hours a week. That number is the difference between a lab bench and a worker.
• The second is the data clause. Not one of Unitree's filings shows a customer contract term over the recordings a robot makes. On the public record the terms point both ways: AgiBot's negotiated contract gives the buyer everything and bars AgiBot from training on it, while Boston Dynamics' standard terms keep a perpetual right to use its robots' performance data. Whether the vendor, as one robotics executive who has negotiated such contracts wrote this month, usually keeps the right to train its models on the customer's sites, is known to the people who sign them.
• The third is the hours by body in the model makers' training mixes: how many hours of recordings each make of robot contributes to the training of models like GR00T, and under what rights. NVIDIA's page names the bodies and not the hours; that share would say how much the learning depends on the lab bench.
• The fourth is the repeat order from an unrelated industrial customer: the second purchase, after the first robot has worked, by a factory that is not a research buyer and not a tour operator. The filing's repeat rates pool humanoids with quadrupeds and count revenue, not customers. A second sale is a business model. A first sale is a press release.
• The fifth is the price of an hour. One seller named a price to the Financial Times, a Gulf seller names its committed revenues, and the model makers name no price at all; what an hour of a working robot's day sells for, and who pays it, is known to the buyers and the sellers.
• The sixth is whether anything leaves the robot unseen: what a deployed Chinese humanoid sends home, to whom, and whether anyone in a position to know has watched recordings move without a contract. No filing will ever carry it. It is known to the people who have assessed such collection for a living, and to the engineers who have read the traffic. I doubt that, as good as Vista is at procuring primary research, we can procure this kind. But we will still try.

Primary research

The people who deployed these robots, trained on them, covered the company, sold for it and wrote the data clauses know what the filings cannot say, and finding and asking them is what Vista is really about. We would most like to hear from these.

Every conversation would draw on past, general experience only: no current employees of the companies named, nothing confidential, and every Advisor screened for conflicts before a word is said. For a client engagement, Vista would combine this inquiry with those interviews: two or three Advisors, structured written conversations, and one brief that says where they agree, where they split, and what would change the answer.

And Tesla might even win, if Optimus arrives soon and is really good. Tesla is the wildcard of this inquiry: it owns the body, the brain and its first customer, its own factories, so none of the three verdicts applies to it, and nothing it learns passes through a Chinese lab or a Chinese contract. All bets are off if a humanoid built that way does paid work before any model trained on bought Chinese bodies does. On October 21, 2026, Elon Musk said he expects deliveries to companies outside Tesla to begin around the end of 2027. So the watch item above becomes a test with a date: if Tesla's annual report for 2027 discloses revenue from Optimus doing paid work for customers outside Tesla, however small, before any model trained on Unitree or AgiBot recordings is reported earning revenue from paid work, then the learning rent sat with the integrated maker all along, and this inquiry's division of the humanoid into body and brain was the wrong map. If that report still shows no Optimus revenue and no unit count, the wildcard stays in the deck and the thesis stands.

In October 1608 the gentlemen of The Hague found the telescope too easy to copy to award a patent, and they were right about the telescope. They could not have foreseen that the thing worth owning was the sky, and the man who looked at it. The robot body is the telescope of this story. The sky is the learning. The question this inquiry has set down for April is who will be found looking through the glass, and whether anyone was looking back.

Questions this inquiry answers

Is Unitree Robotics profitable?

Yes, from selling hardware. In 2025 Unitree's revenue was RMB 1.70 billion, about $254 million, its recurring profit, which leaves out one-time items, was RMB 591 million, and its operating cash flow was RMB 670 million; net profit was RMB 278 million after a one-time charge of RMB 349 million for shares sold cheaply to employees. It sold 5,215 humanoids at an average of about RMB 166,400, about $24,900, and kept 63% of the price as gross margin, down from 69% the year before.

Who buys Unitree's humanoid robots?

Mostly labs and developers. In Unitree's March 2026 reply to the Shanghai Stock Exchange, covering January to September 2025, research and education buyers took 73.6% of humanoid revenue, commercial uses such as promotion, performances and exhibits took 17.4%, and industry took 9%. By the company's own preliminary count only RMB 15.70 million, about $2.35 million or 2.64% of humanoid revenue, went to industrial work it could identify; the rest of the industry line it understood to be mainly enterprise tours.

Does China keep what its robots learn?

As far as the public record shows, mostly not yet. The recordings that stay in China come mostly from government-backed training centers, more than 90 of them built or under way by June 2026, where workers steer robots through practice tasks in a lab, paid for largely with public money; recordings of robots doing real work go to whoever deploys the robots or buys the recordings, Chinese or American. NVIDIA trains its GR00T robot models partly on recordings from a Unitree G1 and AgiBot's Genie-1, and none of Unitree's offering documents shows a customer contract giving it rights to its customers' recordings; a hidden channel, recordings leaving robots in American labs with no contract, cannot be ruled out from filings, and the inquiry names it as a risk.

Who profits from what a robot learns?

By David Teece's rule of 1986, when a technology is easy to copy the profit tends to go to the owners of the scarce assets it cannot do without, its complements. For robots this inquiry names those as the rights to the recordings, the AI models trained on them and the know-how of the customers who put a robot to work, which today points to the model makers, NVIDIA above all, more than to the body maker. A 2008 teardown of an HP notebook shows the pattern: Microsoft kept $85 of the $100 HP paid for Windows, while HP kept a gross margin of 28% on the whole box.

Can American investors buy Unitree stock?

Not easily. Unitree went public on August 19, 2026 on Shanghai's STAR Market at RMB 150.80 a share, about $22.53, and closed its first day at RMB 845, up about 460%; at the offering price a buyer paid about 93 times its 2025 recurring profit and about 36 times its sales. As of October 16, 2026 it was not in Hong Kong's Stock Connect program, the usual route for foreigners: a new company needs a year of trading to qualify through the index, with further tests for weighted votes like Unitree's, and only institutional professional investors may trade such shares.

What would prove this inquiry wrong, and when?

Six measures were frozen on October 16, 2026 and are scored from Unitree's annual report for 2026, due by April 30, 2027, and from NVIDIA's next GR00T model. The view fails if data, model or software revenue reaches 5% of revenue, a disclosed contract gives Unitree reuse rights over its customers' recordings, industrial work beyond tours reaches a fifth of humanoid revenue, the humanoid gross margin falls under 45%, operating cash flow falls under half of recurring profit, repeat buyers fall under 35% of research and education revenue, or NVIDIA's next model names its real-robot training bodies without Unitree or AgiBot among them. A measure the report does not disclose scores untested, and a separate test asks whether Tesla's annual report for 2027 shows Optimus doing paid work for outside customers first.