On June 29, 2026, a company that had existed for two and a half years and had not yet mass-produced a single robot announced that it was worth 2.8 billion dollars.

X Square Robot, a Shenzhen embodied-AI startup, said it had closed four consecutive financing rounds ending in a Series C, lifting its valuation past 20 billion yuan. What made the milestone unusual was not the number. Chinese robotics valuations had been climbing all year. It was the backer list. According to the company, X Square had become the only embodied-AI firm in China to secure lead-round investment, at different stages, from all four of the country’s internet giants: Meituan, Alibaba, ByteDance, and Xiaomi. The four companies do not agree on much. They compete across food delivery, cloud, short video, and consumer hardware. They had all decided to back the same robot company.

The robots themselves are still prototypes. The flagship, a wheeled humanoid called Quanta X2, exists in limited numbers and has not shipped at scale. The company’s revenue is negligible against its valuation. By the ordinary logic of a manufacturing business, none of this adds up.

It adds up differently once you see the money moving around it. In the first half of 2026 alone, China’s embodied-intelligence and robotics sector recorded 288 financing events and disclosed more than 46 billion yuan in funding, already exceeding the total for all of the prior year. Capital was concentrating at the top, and X Square sat at that top alongside a handful of rivals, several of them also valued near or above 2.8 billion dollars. The frenzy was real and it was specific: investors were not spreading bets evenly across the field, they were piling into a small number of companies they believed could own the future.

But X Square is not a manufacturing business, and the four giants were not pricing robots. They were pricing a bet on the hardest and least proven layer of the entire embodied-AI stack: the brain. This is the story of that bet, why it found X Square specifically, and what it reveals about where China’s robotics money is actually going. The short version: the body is nearly solved, and China is now paying its highest prices for the part that isn’t.

The Independent Variable

The company’s name is a statement of intent. In mathematics, the independent variable is the one you change to drive an outcome, the input that everything else responds to. The Chinese name, 自变量, carries that meaning and adds another: 自 means self, or autonomous. X Square wants to be the variable that changes the world, and it wants the change to be self-generated rather than borrowed. For a company that spent its first year unable to raise money, the name was more aspiration than description.

Its founder is not a typical robotics entrepreneur. Wang Qian earned his bachelor’s and master’s degrees at Tsinghua University, then a doctorate at the University of Southern California, where he worked on robot learning and human-robot interaction. Along the way he did something that, in hindsight, reads as remarkable: he was among the earliest researchers to bring the attention mechanism into neural networks, publishing at the same conference as Google’s early attention work in 2014, three years before the Transformer architecture would reorganize the entire field of AI. He was present at the conceptual origin of the technology that now underpins every large model.

And then he left. Wang moved into quantitative finance, founding a quant fund in the United States that by his own account did well. It should have been a satisfying outcome, the kind of career pivot that ends with a comfortable answer to the question of what to do with a doctorate. Instead he described lying awake at night with a single recurring thought: he should have stayed in robotics. The field he had left at its conceptual beginning had, in his absence, started to become the thing he had glimpsed early. In late 2023 he acted on the insomnia, walked away from the fund, and returned to the field, founding X Square in Shenzhen. His co-founder, Wang Hao, holds a doctorate in computational physics from Peking University and had led the development of one of China’s first ten-billion-parameter large models. The pairing is deliberate: a robot-learning researcher who understood the physical world and a large-model builder who understood scale. Their bet was that the future of robotics would look less like mechanical engineering and more like a foundation model, that the discipline was about to shift from building bodies to training brains.

When X Square launched, that bet was contrarian and, worse, unfundable. The embodied-intelligence landscape in China was already crowded. Galbot and AgiBot had both launched the same year with larger teams and more initial capital, and both had built their early identity around visible, demonstrable machines. X Square operated with almost no public profile, no flashy humanoid to show, and a thesis, that the invisible model mattered more than the visible robot, that was hard to sell to investors who wanted to see something move. It struggled to close early rounds. “The biggest difficulty was that nobody trusted us,” Wang has recalled. Two and a half years later, four of China’s most powerful technology companies would compete to fund it. The thing that made X Square unfundable in 2023, its insistence on building the brain before the body, is precisely the thing that made it the most sought-after embodied-AI investment in China by 2026.

The Brain, Not the Body

To understand what changed, you have to understand what X Square refused to build.

Most of the attention in humanoid robotics has gone to the body: the actuators, the joints, the bipedal locomotion, the choreographed demonstrations. Companies like Unitree turned the body into a real business, shipping thousands of capable machines at prices that collapsed the market. When twenty-five Unitree humanoids performed a kung fu routine on China’s Spring Festival Gala, hundreds of millions of people saw what a Chinese-built body could do. But a capable body is not an intelligent one. That Gala performance took months of rehearsal and pre-programmed choreography; even the stumbles were scripted. A robot that can backflip on command is still following a script. The moment the environment changes, the object shifts, the lighting differs, the scripted machine fails.

Wang’s framing for this is blunt. Traditional industrial robots, he argues, are precise, fast, and useless outside the exact conditions they were programmed for. He has described them as precise waste: machines that execute a pre-programmed instruction thousands of times with perfect repeatability and have no ability to observe the world, reason about it, and adapt. Change one variable and they break. The problem was never the body. The body works. The problem is that the body has no brain capable of handling the open, unstructured, endlessly variable real world.

X Square’s answer is what the industry calls an embodied intelligence foundation model, and specifically a Vision-Language-Action model, or VLA: a single system that takes in sensory input, video, language, tactile signals, and outputs physical action. The goal is a robot brain that parallels what large language models did for text, a model general enough that it does not need to be reprogrammed for every new task. Where Unitree bet that whoever built the cheapest capable body would win, X Square bet that the body is nearly solved and the real prize is the brain. It chose the harder half of the problem, and it chose to build it from scratch, in-house, from the first day. The two companies are not really competitors. They are betting on opposite ends of the same robot, and the market has now decided that the end X Square chose, the harder and less visible end, is worth more.

What WALL Actually Is

The brain has a name: WALL. It has arrived in three stages, and the progression tells you how the company thinks.

WALL-A came first, the company’s initial foundation model, built to integrate the Vision-Language-Action approach with what are called World Models, systems that learn to predict how the physical world will respond to an action. The combination matters. A pure VLA model maps perception to action: it sees, and it moves. Adding a World Model lets the system predict the consequence of an action before taking it, and use causal inference to understand feedback, which the company says sharply improves a robot’s ability to generalize to tasks it was never explicitly trained on. In practical terms, a robot with a World Model can reason about what will happen if it grasps an object a certain way, rather than only reacting after the grasp succeeds or fails. In September 2025 X Square released WALL-OSS, an open-source version of the model family, published on GitHub and Hugging Face. Open-sourcing a foundation model is a familiar competitive move: it seeds an ecosystem, invites outside developers to build on your architecture, and positions your model as a potential standard before anyone has agreed on one, the same logic that let earlier software platforms establish dominance by giving the core away.

In April 2026 came WALL-B, and with it the architectural claim that defines the company’s technical identity. WALL-B is built on what X Square calls a World Unified Model architecture. Where a modular system stitches together separate networks, one for vision, one for language, one for action, and passes information between them, WALL-B trains perception, language, action, and physical prediction inside a single unified network. The pitch is that unification produces stronger multimodal understanding and spatial reasoning than a pipeline of bolted-together parts, because the model learns the relationships between seeing, understanding, and acting rather than having them hard-wired at the seams. A modular system knows only what its designers told it to pass between modules; a unified system can, in principle, discover connections its designers never anticipated. Alongside it the company released WALL-OSS-0.5 and a world-modeling component called WALL-WM, which introduces event-level prediction by aligning language, vision, and action data around specific moments rather than treating them as separate streams.

Whether the unified architecture is decisively better than the modular approach is not yet settled. It is a research bet, not a proven result, and serious labs disagree about which path scales. Modular systems are easier to debug and improve piece by piece; unified systems are harder to train but may capture the world more coherently. But X Square’s bet is coherent, made by people who understand the tradeoff, and it is the thing the company is actually selling to investors: not a robot, but a thesis about how robot brains should be built.

The Benchmark Problem

Here is the number that keeps the story honest.

WALL-OSS-0.5, the model X Square released alongside WALL-B, achieved more than 80 percent autonomous task completion on four out of seventeen real-robot tasks, without post-training. That figure comes from the company itself, and read carefully, it is a study in what embodied AI can and cannot yet do. On four tasks, the model performs well straight out of the box, which is a real achievement for a general model with no task-specific tuning. On the other thirteen, it does not clear that bar. Four out of seventeen is not general intelligence. It is early, honest, promising, and nowhere near the reliability a commercial deployment demands.

This is the gap between the valuation and the capability. The body layer of the stack can be measured in units shipped and prices charged; Unitree moved thousands of robots and turned a profit. The brain layer is measured in success rates on manipulation benchmarks, and the leading Chinese brain, by its own reported numbers, clears a high bar on well under half of a seventeen-task suite. The reason this matters is not that four-of-seventeen is bad. For a two-and-a-half-year-old company, it is impressive. It matters because the market is pricing X Square as though the brain is close to solved, and the brain’s own benchmarks say it is not.

There is also a structural caution here. These numbers are self-reported. There is no audited filing, no independent third-party verification of the kind a public company’s financials would carry. A company reporting its own benchmark chooses which tasks to run, which to report, and how to define success, and every one of those choices can flatter the result. This is not an accusation against X Square specifically; it is the default condition of a privately held company announcing its own capabilities to justify its own valuation. The broader embodied-AI field has begun building shared evaluation platforms, including a real-robot benchmark called RoboChallenge that X Square itself helped organize, precisely because self-reported capability claims are hard to compare across companies and easy to frame favorably. The existence of that effort is a tacit admission by the whole industry that the current numbers cannot be taken at face value. When reading any embodied-AI company’s performance figures, including these, the right posture is interest tempered by the knowledge that the company chose which tasks to report and had every incentive to choose well.

None of this means X Square is overstating its progress. Four-of-seventeen without task-specific tuning is, if anything, an honest number, the kind a company inflating its story would have quietly rounded up or reframed. The distinction is narrower and more important: the capability is real, early, and unfinished, and the valuation assumes it will become real, mature, and finished. The distance between those two states is measured in research breakthroughs that have not happened yet, and research breakthroughs do not arrive on a fundraising schedule.

The Data Flywheel

If the model is the product, data is the moat, and X Square talks about data more than almost anything else.

The company organizes itself around three pillars: models, data pipelines, and hardware, and it insists the three form a loop rather than a line. To feed the model, X Square built its own data-capture tools, teleoperation rigs, exoskeletons, and a system it borrows from academic work called a Universal Manipulation Interface, all designed to record how humans actually perform physical tasks. It then runs what it calls a model-driven data pipeline, using the foundation model itself to decide what data is worth collecting and to generate more of it at scale. The model improves the data collection; the better data improves the model. Wang calls this a flywheel, and his framing of the competition is explicit: the next phase of embodied intelligence, he argues, is a battle of foundation models built on data closed-loops and their capacity to keep evolving.

The hardware exists to serve this loop. X Square has released wheeled robots, the Quanta X1, the X1 Pro, and the more humanoid Quanta X2, and it develops core components in-house, including robotic arms, joint modules, and controllers. But the hardware is downstream of the model, not the point of it. The company made a deliberate choice to build on wheeled bases rather than chase bipedal locomotion, on the argument that legs are an expensive distraction from the real problem, which is manipulation and reasoning, not walking. In one demonstration the company likes to cite, a Quanta X1 completed an autonomous food delivery through an open outdoor environment, a setting with none of the controlled conditions a lab demo relies on. It handled strong winds, a deformed package, and objects it could only partly see, using the model’s causal inference to fill in what was hidden and self-correcting when it stalled, completing the delivery loop without human intervention. In another, facing a disordered pile of parcels, the robot used generalization to pick out irregular items it had not been trained on specifically.

Whether those demonstrations generalize beyond the demonstration is exactly the four-of-seventeen question. A single successful outdoor delivery is not a reliability statistic; it is an existence proof that the approach can work at all, which is different from proof that it works consistently. But the ambition is coherent: a robot that handles the unscripted world by understanding it, not by being told about it in advance. The wager is that this understanding, once the model is good enough, transfers across tasks the way a language model’s fluency transfers across topics. That transfer is the whole thesis, and it is the thing not yet demonstrated at the reliability a business would require.

Why Four Giants Bet

The most revealing fact about X Square is the identity of its backers, because it explains what they think they are buying.

Meituan, Alibaba, ByteDance, and Xiaomi are not natural allies. They are pouring money into X Square anyway, and the logic is not robot hardware, which any of them could commission. The logic is the operating system. If embodied AI develops the way personal computing and mobile did, the durable value will not sit in the machines but in the software layer that every machine runs on, the platform that becomes the default brain for a generation of robots. Betting on X Square is a bet on owning, or at least sitting close to, that layer. It is an option on the robot operating system, bought early, while the price of the option is still a startup valuation rather than a platform monopoly.

Each giant has its own reason to want the seat. Meituan runs the largest delivery logistics network in China, a natural deployment ground for autonomous physical labor, and has led rounds accordingly. Alibaba, through its cloud arm, has a strategic interest in the compute and model infrastructure embodied AI will consume, and co-led an early round through Alibaba Cloud. ByteDance is a foundation-model competitor in its own right with an appetite for the next platform. Xiaomi builds a consumer-hardware and IoT ecosystem into which a household robot would slot directly, and led the Series B. Four different strategic motives converging on one company, and converging specifically because X Square built the foundation model in-house from the beginning rather than licensing a brain or bolting one onto someone else’s hardware. In a field where most companies bet on the body, X Square was the cleanest available bet on the brain, and the giants paid for cleanliness.

The pattern rhymes with earlier platform transitions. In the PC era, the lasting fortune went not to the companies that made the boxes but to the one that made the operating system every box ran. In mobile, the same. The hardware became a commodity; the platform captured the value. If embodied AI follows that arc, the humanoid body will commoditize the way PCs and phones did, driven down a cost curve by manufacturers like Unitree, while the brain, the model that every robot licenses or runs, becomes the layer that compounds. The four giants are each hedging against the possibility that they wake up in a robot-filled world where someone else owns the operating system. Backing X Square is cheaper than that risk. It does not require any of them to believe X Square will certainly win, only that the layer is worth owning a piece of, and that X Square is the purest available claim on it.

The Valuation Gap

Wang Qian is unusually candid about the economics, including the parts that should give an investor pause.

He has argued publicly that China’s embodied-AI companies are valued roughly an order of magnitude below their overseas counterparts, and that the sector needs more capital, more willingness to inflate, to drive the innovation required. He has said he is in no hurry to commercialize, that he allocates about two-thirds of spending to improving the model’s raw capability rather than chasing near-term revenue, and that he expects the first scenarios with positive return on investment to appear around 2026. He has predicted early consumer products in three to four years at a price somewhere between ten and twenty thousand dollars. These are the statements of a founder building for a horizon well beyond the current product, and they are refreshingly free of the pretense that the robots are about to generate meaningful revenue.

They also describe the risk precisely. A 2.8-billion-dollar valuation on negligible revenue is not a claim about today. It is a claim about a future in which X Square’s brain becomes something close to the default operating system for physical AI, deployed across delivery, logistics, elder care, and eventually the home. The valuation prices that future as substantially likely. The four-of-seventeen benchmark, the prototype-stage hardware, and the founder’s own three-to-four-year horizon all say the future is real but far, and contingent on a research bet, the unified-model architecture, that has not yet been proven to win. The gap between what the model can do today and what the valuation assumes it will do is the entire investment. Everything rests on whether the brain, in the end, generalizes.

The Bet on the Brain

Earlier in this series, Cambricon offered a distinction worth borrowing: capability versus capture. Cambricon had proven capability, a good-enough chip that ran at scale, while the market priced a capture it had not yet secured. X Square inverts the pairing. Here the capture has arrived first: four giants, 2.8 billion dollars, the pole position in Chinese embodied AI. It is the capability that remains unproven, sitting at four reliable tasks out of seventeen.

That inversion is the whole point of where China’s robotics money now flows. The body has been substantially solved. Unitree and its peers ship capable machines by the thousand, at prices no Western maker can match, and turn a profit doing it. The body is a manufacturing problem, and China is very good at manufacturing problems. The brain is a different kind of problem. It cannot be driven down a cost curve or scaled through a factory. It has to be discovered, and no amount of capital guarantees the discovery arrives on schedule.

This is why the most expensive bet in Chinese embodied AI is not the company that makes the best robot. It is the company that might, someday, make the best robot brain. X Square is that bet in its purest form: a foundation-model thesis wrapped in prototype hardware, funded by four rivals who each concluded that the software layer is where the lasting value lives. They may be right. The brain is the correct thing to want. It is also the least proven layer of the entire stack, the one part that manufacturing prowess cannot deliver, and the one China has chosen to price the highest. The body is real. The brain is the wager. The next few years will show whether the independent variable actually changes the outcome.

Sources

Company and founder: Pandaily (”X Square Robot’s Wang Qian: Robots will eventually reach Mars,” April 2026, exclusive interview); Baidu Baike (Wang Qian profile); Chinadaily (”X Square Robot expands development of general-purpose robots,” August 2025); aifun.cc company profile. Founding (December 2023, Shenzhen), Wang Qian’s background (Tsinghua bachelor’s/master’s, USC PhD in robot learning, early attention-mechanism researcher publishing concurrently with Google in 2014, quant-fund founder before returning to robotics), and co-founder Wang Hao (Peking University computational physics PhD, led one of China’s first ten-billion-parameter models) are reported across these sources. The “nobody trusted us” and “precise waste” characterizations are Wang Qian’s own words per Pandaily and Chinese-language interviews.

Funding: PRNewswire / Morningstar / Yahoo Finance / The Robot Report / Pandaily / finsmes (”X Square Robot Secures Four Consecutive Financing Rounds,” June 29, 2026). Series C valuation of over US$2.8 billion (RMB 20 billion), the four-consecutive-rounds structure, and the claim of being the only embodied-AI company in China with lead-round backing from all four internet giants (Meituan, Alibaba, ByteDance, Xiaomi) are per the company’s own announcement, carried by these outlets. IDG participated in Series C; HongShan and Xiaomi in prior rounds. Earlier rounds (Series A+ ~1 billion yuan / $140M co-led by Alibaba Cloud and CAS Investment, September 2025; Series A++ ~$140M January 2026; Series B ~$276M led by Xiaomi, April 2026) are per The Robot Report, Frontier Enterprise, and RoboticsTomorrow. Total funding across rounds is estimated at $500-700M per Cryptobriefing. Valuation is stated per company announcement; there is no audited financial disclosure.

Models and technology: The Robot Report (”X Square Robot debuts foundation model,” September 2025; “secures $140M,” January 2026); RoboticsTomorrow; Frontier Enterprise; citybiz; Pandaily. WALL-A (VLA integrated with World Models), WALL-OSS (open-sourced September 2025 on GitHub and Hugging Face), and WALL-B (April 2026, “World Unified Model” architecture training perception, language, action, and physical prediction in a single network) are per company announcements carried by these outlets. The WALL-OSS-0.5 benchmark result (over 80% autonomous completion on four of seventeen real-robot tasks without post-training) is a company-reported figure; there is no independent third-party verification. RoboChallenge, the shared real-robot benchmark X Square helped organize, is referenced per the RoboChallenge organizing committee and arXiv benchmark documentation.

Products: Robozaps and Humanoid.guide (Quanta X2 specifications: ~172cm, 95kg, 62 total DoF, 6-DoF omnidirectional chassis, dual 7-DoF arms, optional 20-DoF hands, LiDAR/IMU/ultrasonic sensing, WALL-A model); The Robot Report (Quanta X1 autonomous food-delivery demonstration). Product status is prototype per manufacturer listings.

Industry context: BigGo Finance and IT Juzi (China embodied-intelligence funding H1 2026: 288 financing events, over 46 billion yuan, exceeding the prior full year); SiliconANGLE and Bloomberg (AI² Robotics and X Square each surpassing ~$2.8B valuation, June 2026). The “hardware camp versus brain camp” framing of the sector is per BigGo’s industry analysis.

Classification: Founding facts, funding rounds, investor identities, and model release dates are reported across multiple sources and attributed to company announcements where they originate there. All valuation and benchmark figures are company-reported, as X Square is privately held with no audited disclosure, and are attributed accordingly. Wang Qian’s forward-looking statements (pricing, timeline, capital-allocation, valuation-gap views) are his own predictions and characterizations, attributed as such. The unified-model architecture’s superiority over modular VLA is a research claim, not a settled result, and is presented as a bet rather than a fact.

A few hundred precision manufacturing firms in the Yangtze and Pearl River deltas solved the humanoid body. The robot companies bought the result and put their names on it.

Watch what that looks like on television. At the 2026 Spring Festival Gala, a few hundred million people saw humanoid robots throw punches and turn backflips on the most-watched broadcast on Earth. Machines from four Chinese companies appeared across the night’s programme. Unitree’s performed the martial arts routine. Galbot’s appeared elsewhere in the show.

The joints came from the same factory in Suzhou.

That company is called Leaderdrive, known in Chinese as 绿的谐波, and it trades on Shanghai’s STAR Market under the code 688017. In 2025 it booked 571 million yuan of revenue, roughly 80 million dollars. That is less than Zhipu earned in the same year, and Zhipu is a company this series has described as losing six yuan for every one it takes in.

Hold those two facts next to each other. The company that supplies the joints for most of China’s humanoid robots is smaller, by revenue, than a loss-making AI lab. And if it stopped shipping tomorrow, a large fraction of China’s humanoid programmes would stop with it.

This series has spent four articles on companies the capital markets have already found. Cambricon, priced as a proxy for the entire domestic chip industry. X Square, valued at 2.8 billion dollars on a model that clears its own benchmark four times out of seventeen. Zhipu, whose share price went up twenty-five-fold on a float under four percent. Moonshot, worth thirty billion by the last reported round.

This article is about the layer underneath all of them, where the numbers are small, the companies are unglamorous, and the actual constraint on China’s robot industry has been sitting for the last fifteen years. The story starts with a man who flew to Tokyo to look at a part.

Tokyo, 2003

Zuo Yuyu studied physics at Nanjing University and went into manufacturing. By the early 2000s he was running a metal-parts business that supplied components to GE and ABB, which is a respectable place to be and not a famous one.

Somewhere in that work he learned something that stuck. ABB, one of the largest robotics companies in the world, was unhappy with its Japanese suppliers. Not unhappy enough to switch, because there was nothing to switch to. Just unhappy, in the way a customer is unhappy when it has no alternative and everyone in the room knows it.

At the end of 2003 he flew to Japan to see the part.

In Tokyo he learned the name of the company that made it: Harmonic Drive Systems. By then it had held the global market for harmonic reducers for more than forty years, with a share that at times exceeded ninety percent. Its technology was in the drive wheels of the Apollo lunar rover. Every robot manufacturer on Earth, regardless of size, bought on its terms.

The account of that trip comes from a Chinese-language reconstruction published years later by 36Kr, so the interior detail should be read as the company’s own telling. What is reported is that the visit stung. And that Zuo’s response was a physicist’s rather than an entrepreneur’s.

Isn’t this just a more complicated physics problem? Why can’t we do it?

He went home and convinced his team to try.

Under a Millimetre

To understand why that question was harder than it sounds, you have to understand what a harmonic reducer is, because it is one of the strangest devices in mechanical engineering.

Every robot joint has the same problem. The electric motor spins fast and weakly. The joint needs to move slowly and strongly. Something has to convert one into the other, and in a humanoid robot that something has to be small, light, and precise enough that a hand arrives where the software said it would.

A harmonic reducer solves this with three parts. There is a rigid outer ring with teeth on the inside. There is a thin steel cup, called a flexspline, with teeth on the outside and slightly fewer of them. And there is an elliptical bearing, the wave generator, which sits inside the cup and squashes it into an oval so that its teeth engage the outer ring at two points. Turn the wave generator, and the point of contact travels around the ring. Because the cup has fewer teeth, it creeps backwards by that difference with each revolution. Enormous reduction ratios come out of a device barely thicker than a hockey puck.

The cup is bent out of shape, continuously, for the entire life of the machine.

That is the elegance and it is also the problem. The flexspline is a structural component designed to be deformed millions of times without failing. Its wall is typically under a millimetre thick. Its teeth must be cut to ISO Class 4, which means pitch tolerances in single-digit micrometres, and its runout has to stay within about five micrometres. It is made of high-strength steel, which resists being machined, and it must be clamped for machining without distorting, which is difficult precisely because it is thin enough to distort.

Get the mesh slightly loose and the joint has backlash, which means the robot’s hand does not arrive where the software said. Get it slightly tight and the reducer cannot be assembled at all. And the failure mode that matters most is not immediate. It is fatigue, cracks that begin at the roots of the teeth after millions of cycles, seeded by a fillet radius that was slightly wrong or a tool mark left behind during machining.

This is the reason the monopoly lasted forty years, and it is not the reason people usually assume. The principle is public. It was patented in the 1950s and the patents expired long ago. Any competent engineer can read how a harmonic drive works in an afternoon.

What could not be read was the process. How to hold a sub-millimetre steel cup so it does not flex under the cutter. Which heat treatment gives fatigue life without warping the geometry. How to finish a tooth root so it does not seed a crack. That knowledge lives in machine operators and process engineers and years of scrapped parts, and it does not transfer by reading.

There is a particular cruelty in the feedback loop. A flexspline machined slightly wrong does not announce itself. It assembles correctly, turns smoothly, passes inspection, and works for months before a crack opens at a tooth root and a joint fails inside a customer’s machine. The interval between making the error and learning of it can be a year, and there is no way to run that loop faster. A barrier of this kind is not measured in cleverness. It is measured in how many times you have been allowed to be wrong.

Zuo was right that it was a physics problem. He was wrong about how long the physics would take.

Six Years

The company was not founded in 2003, or 2004, or 2007.

Leaderdrive was incorporated in 2011, more than six years after the Tokyo trip. Six years passed between deciding to make the part and having a company that made it. Those years went into the part of the work that does not photograph well: metallurgy, grinding, heat treatment, measurement, and the slow accumulation of knowledge about how a thin steel cup behaves when you try to make it perfectly.

In March 2011 the company sent its first reducer prototype to the Jiangsu provincial inspection centre for reducer product quality. The part passed. It was the first harmonic reducer in China to reach industrial production and large-scale application, which is a sentence that means a Chinese factory could finally buy a domestic one.

The company that resulted is not large even now. Its revenue is a fraction of any of the AI labs this series has covered. It listed on the STAR Market in 2020 and was the first company from its district of Suzhou to do so, which tells you something about the neighbourhood it grew up in: an industrial town where a listed company was still an event.

The Share Shift

For most of the following decade, the numbers moved slowly. Then the humanoid programmes arrived, and the market moved faster than it had in forty years.

Precision matters in describing this, because share figures for harmonic reducers get quoted with several different denominators and they are not interchangeable.

Within China’s harmonic reducer market, domestic producers as a group held roughly fifteen percent in 2022. By the first quarter of 2024 that had risen to around thirty-eight percent. As of 2022, Harmonic Drive Systems still led the Chinese market at approximately thirty-eight percent with Leaderdrive at roughly twenty-six percent, and JPMorgan has more recently put Leaderdrive at thirty to forty percent of that same market.

Inside the humanoid segment specifically, the picture is different and more lopsided. Leaderdrive is reported to hold about sixty percent of Chinese service and humanoid robot harmonic reducers. Morgan Stanley assumes it will take about forty percent of the global humanoid harmonic reducer market in 2026, settling to around twenty-five percent over the longer term.

Those measurements run on different denominators and this article is not going to average them. What they agree on is direction. A market that barely moved for four decades reorganised itself in about three years, and it reorganised fastest exactly where the new demand was.

The financial result showed up in 2025. Revenue of 571 million yuan, up 47.31 percent. Net profit attributable to shareholders of 124 million yuan, up 121.42 percent. The fourth quarter turned a small prior-year loss into a 31 million yuan profit. In the first quarter of 2026 revenue rose 42.96 percent and profit 61.17 percent. Gross margin held near 37 percent for the full year.

The equity market noticed in its own way. Leaderdrive shares rose about forty percent over the year to April 2026, and Zuo Yuyu and his brother Zuo Jing became billionaires on paper.

Twenty-three years after a trip to Tokyo to look at a part nobody in China could make.

Investors looking for humanoid exposure had begun stepping past the robot companies and buying the joints instead, which is a rational thing to do when the robot companies are numerous and the joint suppliers are not.

The Delta

Leaderdrive is the cleanest example, not the whole story. Around it sits a supply chain that has been quietly assembling for two decades, and the aggregate effect is larger than any single firm in it.

Start with the multiplier. A humanoid robot needs somewhere between twenty and forty harmonic reducers, depending on design. A six-axis industrial arm typically uses harmonic gears only on its three wrist axes, with sturdier cycloidal units carrying the base and shoulder. Every humanoid that ships is therefore worth roughly seven to thirteen industrial arms to a harmonic reducer maker, which is why a market that grew arithmetically for forty years is now being asked to grow geometrically.

The multiplier explains why the joint dominates the bill of materials. Joint actuators account for more than thirty percent of a humanoid’s component cost by industry estimates, reaching around half in simpler configurations. Dexterous hands add another fifteen to twenty percent. Taken together, somewhere between roughly half and two-thirds of the physical cost of a humanoid robot is the machinery that makes it move. The price of the robot is largely a question about the price of motion, and the price of motion is set in a handful of industrial districts.

The rest of the joint has followed the same path. Leadshine reported delivering more than 120,000 frameless torque motors in 2025, more than twenty times the prior year. Shuanglin has developed sixty-three ball and planetary screw products for three customers. Sanhua’s robotics revenue grew 320 percent year on year in the first half of 2025, a figure the company disclosed while denying a widely circulated report about the size of a specific order.

The most dramatic collapse happened in sensing. Five years ago a tactile sensor of the kind a robot hand needs had to be imported and cost more than a hundred thousand yuan. After domestic breakthroughs, prices have fallen as low as 199 yuan. The cost of one imported sensor five years ago now equips an entire domestically made dexterous hand.

Dexterous hands themselves ran an average of about 7,960 dollars globally in 2024. Some Chinese models are now under a thousand. Lingqiao’s DexHand021 Pro offers twenty-two degrees of freedom at a fifth of the price of comparable international products.

None of this happened in laboratories.

It happened in the industrial belts of the Yangtze and Pearl River deltas, in towns like Mudu, the district of Suzhou where Leaderdrive built its plant, and in several hundred others that look much the same: low buildings, a rail spur, a canteen, a car park filling with the vehicles of people who machine things for a living. These were factories already making precision parts for cars and appliances and industrial automation. Around 2024 they discovered that a new customer had appeared, one that wanted the same skills pointed at smaller and stranger parts.

What followed is legible in what they bought. Screw factories bought grinding machines. Reducer factories bought inspection equipment. Dexterous hand makers went looking for micro motors and flexible materials, which meant orders for their suppliers, which meant orders further up again. A purchase order for a grinding machine is a slower and more reliable signal than a funding round, because nobody buys one to be seen buying it.

Analysis by Gasgoo’s automotive research institute puts the cost advantage of localised core components at fifty to seventy percent against foreign equivalents. That is the number underneath Unitree’s price destruction, described earlier in this series. Unitree did not invent a cheaper robot. It bought from a chain that had spent a decade learning to make the expensive parts cheaply, and then passed the saving on faster than anyone expected.

Optimus Runs on Chinese Joints

The awkward part of this is not Chinese.

By multiple accounts, something like seventy percent of the components in Tesla’s third-generation Optimus are sourced from Chinese suppliers. Tuopu, a Ningbo company that has supplied Tesla vehicle chassis since 2016, became an actuator supplier for Optimus and delivered batches in the second quarter of 2025 sufficient for roughly seven hundred robots. Its reducers are priced thirty to forty percent below Japanese equivalents. Leaderdrive has passed Tesla’s supplier validation.

Musk’s stated target is an Optimus that costs under twenty thousand dollars, because his own internal arithmetic says the market only becomes enormous below that line. The arithmetic works if the joints are Chinese. It is considerably harder if they are Japanese, and it has not been demonstrated at all if they have to be American, because the American supply chain for this specific category of precision component largely does not exist.

This is the same structure the earlier articles in this series found in silicon, arriving from the opposite direction. There, export controls made the best chips unavailable to China, and a domestic industry grew in the space that was left. Here, no controls were needed. The capability simply migrated, one machine tool and one process engineer at a time, to the place that was already making everything else.

A humanoid robot is a machine assembled from a few thousand precision parts. Whoever makes those parts most cheaply and most reliably will end up inside everyone’s robot, including the robots built by companies that would prefer otherwise.

The obvious response is to build the capability somewhere else, and the obvious response is harder than it sounds. A chip fabrication plant can be bought. It costs a great deal of money, the equipment list is known, and a government that wants one badly enough can write the cheque, as several have. A precision components industry cannot be bought, because what makes it work is not the machines. It is several thousand people who have spent fifteen years learning what a particular alloy does when you grind it at a particular speed, distributed across hundreds of firms that supply and poach from one another inside a few hundred kilometres.

That is why the joints are a more awkward dependency than the chips. Chips are a chokepoint that can be attacked with capital. Joints are a chokepoint that has to be attacked with time, and the clock started in the deltas around 2005.

What the Capacity Decision Looks Like

A chokepoint only stays one if it can supply. For Leaderdrive the live decision is not whether demand is coming. It is how much capacity to build before it arrives.

Get it wrong in one direction and the orders go to someone else, permanently, because a validated supplier relationship in a safety-critical component is not easily reversed. Get it wrong in the other and a company with 571 million yuan of revenue is sitting on a factory built for a market that did not show up on schedule.

Leaderdrive chose to build. Monthly harmonic reducer capacity went from about fifty thousand units in the first quarter of 2026 to roughly seventy thousand by mid-year, with a stated plan to reach one hundred to one hundred and twenty thousand by year end. That is roughly a doubling across 2026, funded by a company whose entire annual revenue would cover about three months of Zhipu’s compute bill.

The demand forecast underneath it comes from the sell side and should be read as such. Morgan Stanley raised its 2026 shipment forecast for Chinese humanoids from twenty-eight thousand units to fifty thousand, and projects four hundred and forty-six thousand by 2030, a compound annual growth rate of 106 percent from 2025. It expects humanoid-related sales to reach thirty-five percent of Leaderdrive’s revenue in 2026 and half by 2027.

Set against that, the current reality is modest. Soochow Securities expected global humanoid sales below thirty thousand units in 2025. UBTech’s total order value for the year approached 1.4 billion yuan and Unitree’s approached 1.2 billion. These are real businesses and they are not yet large ones.

Which means the capacity being built today is a bet on a forecast, made by a company that cannot afford to be wrong twice.

The Force That Eats Its Parents

There is a problem with being the company that made things cheap, and it is visible in the numbers already.

Look again at what happened to tactile sensors. An imported unit cost more than a hundred thousand yuan five years ago. Domestic versions now sell for as little as 199 yuan. That is not a discount. It is a category destroyed and rebuilt at roughly a five-hundredth of the price, and the destruction reached the domestic sensor makers who performed it, not only the foreign incumbents they displaced.

Nothing about harmonic reducers exempts them from the same process. The barrier described earlier, process knowledge held in machine operators and scrapped parts, is durable against a foreign competitor starting from nothing. It is much less durable against the fiftieth Chinese firm to enter, hiring from the first, buying the same grinding machines from the same vendors, in the same industrial district.

The share data hints at this. Domestic producers collectively went from roughly fifteen percent of the Chinese market in 2022 to roughly thirty-eight percent by early 2024. Leaderdrive is the largest of them and did not capture all of that gain. The rest went to companies most readers have never heard of, and there are more of them every quarter, because a market growing at the rate the sell side is forecasting attracts entrants the way an open window attracts weather.

Leaderdrive’s gross margin was 36.91 percent for 2025 and 33.62 percent in the first quarter of 2026. Those are healthy numbers for a components manufacturer, and they are also the target that every new entrant is aiming at. The company’s own answer is to move up: a stated strategy of pairing reducers with integrated electromechanical actuator modules, which is the standard defence of a component maker trying not to become a commodity supplier, and which works for exactly as long as the module is harder to copy than the part.

The uncomfortable symmetry is this. The thing that made Chinese humanoids possible was a decade of relentless cost reduction in precision components. The thing most likely to damage the companies that performed it is another decade of the same. Unitree could destroy the price of a robot because its suppliers destroyed the price of a joint. Someone is now preparing to destroy the price of a joint again.

And this time Zuo Yuyu is the incumbent.

He spent six years learning to make a part that a Japanese company had owned since the 1960s, and roughly fifteen more turning that into a business with a national share. Somewhere in a similar town, someone who once worked for him is currently explaining to an investor that the technology is not that hard, that the patents expired decades ago, and that the process knowledge can be acquired by hiring the right eleven people.

That argument was correct when Zuo made it in 2003. It is not obvious why it stopped being correct.

That is the ordinary fate of a supply chain that works. The current market narrative, in which the joint makers are the scarce and defensible layer of the humanoid trade, describes a moment rather than a structure.

The Layer Nobody Priced

Every article in this series has found a constraint.

Cambricon exists because export controls made the best chips unavailable, and good enough and available beat best and unavailable. X Square is expensive because the brain is the unsolved layer of embodied AI. Zhipu cannot yet make the arithmetic work because a model is built once and paid for again every time somebody uses it. Moonshot ran out of compute three days after shipping the model that made the market question whether the world needed so much of it.

The joints are the constraint nobody wrote about, because they were solved before anyone was watching.

While capital was pricing chips and models and brains, a few hundred manufacturing companies were working out how to machine a sub-millimetre steel cup to five-micrometre precision, and how to do it at a price that would let a robot cost less than a car. That work did not produce a valuation story until 2025. It produced something more durable: a physical capability that sits inside almost every humanoid robot on Earth, including the ones assembled in Texas.

There is a version of this story that credits national strategy, and it would not be wrong. China’s humanoid action plan sets a target of a hundred thousand robots deployed by 2027, and calls explicitly for a complete domestic supply chain in actuators, dexterous hands, and perception. The policy is real and the money behind it is real.

But the physics was done first, and it was done by people who were not waiting for a plan.

In 2003 a man who had studied physics flew to Tokyo because he wanted to look at a part that his customers could not buy anywhere else. It took him more than six years to make one, another decade to make it well, and about three years after that for the rest of the world to notice that the joints inside its robots had quietly changed nationality.

The bodies were never the hard part, in the end. They were just the part that somebody had to learn to make.

Sources

Leaderdrive company and founder: 36Kr (Chinese-language profile of Zuo Yuyu and the founding of Leaderdrive); Kunihiro Koreeda’s Chinese Corporate Encyclopedia via Nikkei-affiliated commentary (April 2026); Leaderdrive corporate website and STAR Market disclosures (688017.SH). Zuo Yuyu’s physics background at Nanjing University, his metal-parts business supplying GE and ABB, ABB’s dissatisfaction with Japanese suppliers, the 2003 Tokyo visit, the “more complicated physics problem” framing, the six-year development period, and the 2011 founding in Suzhou are drawn from these accounts. The interior detail of the Tokyo trip rests on a single Chinese-language reconstruction published years after the fact and is presented as the company’s own telling rather than as independently verified reporting. The March 2011 submission of the first prototype to the Jiangsu provincial reducer inspection centre, and Leaderdrive’s status as the first Chinese firm to achieve industrial production and large-scale application of harmonic reducers, are reported in Chinese industry coverage. Leaderdrive’s position as the first STAR Market listing from its district of Suzhou is per the company’s own account of its 2020 listing.

Harmonic Drive Systems and the historical monopoly: 36Kr; Nikkei-affiliated commentary. The forty-year duration of the monopoly, the peak global share above ninety percent, and the use of the technology in Apollo lunar rover drive wheels are as reported in these accounts. Harmonic drive principles were patented in the 1950s and those patents have long expired.

Harmonic reducer engineering: EMAG (flexspline machining specifications); MDPI Actuators, “A Novel Strain Wave Gear Reducer with Double Flexsplines” (2023); Springer, International Journal of Precision Engineering and Manufacturing (2026); EVS International reducer comparison (May 2026); USPTO patent documentation on double-flexspline designs. Flexspline wall thicknesses typically under one millimetre, ISO 1328 Class 4 gear quality with single-digit micrometre pitch tolerances, runout within approximately five micrometres, typical workpiece diameters of 25 to 140 millimetres, and high-cycle fatigue at tooth roots as the life-limiting failure mode are drawn from these engineering sources. The characterisation of process knowledge rather than intellectual property as the durable barrier is this article’s reading of that material, not a claim made by any single source.

Leaderdrive financials: Company annual results and quarterly disclosures as summarised by Soochow Securities via Sina Finance (May 2026). FY2025 revenue of RMB 571 million (+47.31%), net profit attributable to shareholders of RMB 124 million (+121.42%), Q4 2025 revenue of RMB 164 million with net profit of RMB 31 million against a prior-year loss of RMB 3 million, Q1 2026 revenue of RMB 140 million (+42.96%) and net profit of RMB 33 million (+61.17%), FY2025 gross margin of 36.91% and net margin of 21.79%, and Q1 2026 gross margin of 33.62% are per those disclosures. H1 2025 revenue of RMB 251 million with 34.77% gross margin is per 36Kr citing company reporting.

Market share: Figures in this section carry different denominators and are labelled accordingly. Domestic Chinese producers’ collective share of the Chinese harmonic reducer market rising from approximately 15% in 2022 to approximately 38% in Q1 2024 is per industry analysis circulated in June 2026. Harmonic Drive Systems at approximately 38% and Leaderdrive at approximately 26% of the Chinese market as of 2022, and Leaderdrive at approximately 60% of the Chinese service and humanoid robot segment, are per Nikkei-affiliated commentary. JPMorgan’s estimate of Leaderdrive at 30-40% of China’s harmonic reducer market is as reported by humanoid.guide. Morgan Stanley’s assumptions of 40% of the global humanoid harmonic reducer market in 2026 and approximately 25% long term are per its research as reported by BigGo Finance. No average or composite of these figures is calculated here.

Capacity and forecasts: Morgan Stanley research as reported by BigGo Finance (June 2026). Monthly harmonic reducer capacity rising from approximately 50,000 units in Q1 2026 to approximately 70,000, with a stated plan of 100,000 to 120,000 by year end; the raised 2026 Chinese humanoid shipment forecast from 28,000 to 50,000 units; the 2030 projection of 446,000 units at a 106% compound annual growth rate from 2025; and the expectation that humanoid-related sales reach 35% of Leaderdrive revenue in 2026 and 50% in 2027 are per that research. Sell-side forecasts are estimates and are presented as such. Soochow Securities’ expectation of global humanoid sales below 30,000 units in 2025, and 2025 order totals approaching RMB 1.4 billion for UBTech and RMB 1.2 billion for Unitree, are per 36Kr.

Wider supply chain: Gasgoo Automotive Research Institute (February 2026) for the 50-70% cost advantage of localised core components and Lingqiao’s DexHand021 Pro specifications; Tianxia Gongchang Research (July 2026) for Leadshine’s 2025 delivery of more than 120,000 frameless torque motors representing more than twentyfold growth, Shuanglin’s 63 ball and planetary screw products, and the characterisation of delta-region factory capacity expansion; China Humanoid Robotics Tracker (March 2026) for tactile sensor pricing falling from above RMB 100,000 for imported units to as low as RMB 199, and for global average dexterous hand pricing of approximately $7,960 in 2024 against sub-$1,000 Chinese models. Sanhua’s 320% year-on-year robotics revenue growth in H1 2025 is per company disclosure as reported by Humanoids Daily; Sanhua publicly denied a separately circulated report regarding the size of a specific Tesla order, and no order figure is asserted here.

Tesla Optimus supply chain: optimusk.blog compilation citing iNEWS Robot Core supply chain reporting and Yicai Global; KR-Asia (January 2026). The estimate that approximately 70% of third-generation Optimus components are sourced from Chinese suppliers, Tuopu’s role as an actuator supplier and its Q2 2025 delivery volumes sufficient for approximately 700 robots, its pricing 30-40% below Japanese competitors, and Leaderdrive’s passage of Tesla supplier validation are reported across these sources rather than confirmed by Tesla. Tesla’s internal target of an Optimus unit price below $20,000 is per Musk’s public statements and Tesla’s Q3 2025 earnings call.

Policy: MIIT and five other ministries’ 2025 Humanoid Robot Action Plan, setting a national target of 100,000 humanoid robots deployed by 2027 and calling for a complete domestic supply chain in actuators, dexterous hands, and perception systems, is per Silicon Valley Robotics Center research (April 2026).

Cross-references: Figures cited for Cambricon, X Square Robot, Zhipu, Moonshot, and Unitree are drawn from the preceding articles in this series and their sources.

Classification: Leaderdrive’s financial results are Confirmed from company disclosures. Engineering specifications for flexspline manufacturing are Confirmed from published technical sources. Market share figures are Reported, carry differing denominators, and are labelled individually. Capacity plans and shipment forecasts are sell-side estimates and are attributed as such. The Tokyo founding narrative is a single-source reconstruction and is flagged in the text. Tesla supply chain composition is Reported and not confirmed by Tesla.

Memory, more than the roadmap, limits Chinese AI compute. Huawei’s 950DT reportedly passed 250,000 yuan while CXMT’s IPO left 36.8 billion yuan unassigned.

Memory, more than the roadmap, limits Chinese AI compute this week, and the evidence shows the shortage priced in one place and unassigned in another. At Huawei Connect on September 17, Huawei moved its next accelerator forward three quarters, and deputy chairman Eric Xu told reporters that China will not catch up with demand for AI hardware, including memory chips and optical components, until 2030. The chip Chinese labs are lining up for, the memory-heavy Ascend 950DT, has reportedly been repriced past 250,000 yuan, while CXMT, the DRAM maker that listed in Shanghai in July, reported its final IPO proceeds at 66.6 billion yuan and left 36.8 billion of it above the three projects it named. For an investor sizing 2027 supply, the number that matters is not a launch date but who is funded to build the memory that goes on the chip. No consensus estimate of either company’s output of high-bandwidth memory, or HBM, turned up in this pass, so this issue tests a proposition and does not claim a mispricing.

The date moved and the volume did not. Huawei’s own release puts the Ascend 960DT in the first quarter of 2027 and the 960PR in the third, three quarters and one quarter ahead of plan. It also launched a new pod, the Atlas 960E SuperPoD, a rack-scale cluster that works as one machine and scales to 4,096 chips. The earlier roadmap keynote described an Atlas 960 SuperPoD of up to 15,488 chips for the fourth quarter of 2027. Neither 2026 page says whether the new pod replaces that one or sits beside it, and neither gives the 960E a ship date, so this issue does not call the pod smaller. They are different products. All three Huawei pages were read in full this session; anything marked Reported below was not. What the pages do not contain is a count of 960DT chips that will exist in 2027. That count, not the date, sets what a lab can plan an inference fleet around.

The price sheet says what is short. Reuters reported the indicated price of the 950DT on September 10 at above 250,000 yuan, up 20 to 50 percent from quotes two months earlier, with the older 950PR above 80,000 yuan from about 60,000 at the start of the year [Reported: Reuters, anonymous sources, Huawei did not respond]. Bloomberg put the 950DT rise at about 60 percent over three months [Reported: Bloomberg, anonymous sources]. The two agree on the level and differ on the rise, which each measures over a different window, so the price is the fact and the percentage is not. Cambricon, a listed rival, repriced its unreleased next chip by 20 to 30 percent, per Reuters, so the squeeze is not only Huawei’s. Huawei cited tight component supply to customers, per Bloomberg. Reuters’ sources put the squeeze on memory: Chinese chipmakers lean on grey-market HBM that costs several times the price paid outside China, and Huawei has not said where its own HBM is made. In July this publication tied the inference ceiling to the 950DT’s fourth-quarter date and said the ceiling would stay if the chip slipped or shipped thin. The date still stands. The first evidence on volume points to scarcity: DeepSeek plans to deploy at least 160,000 of these chips, more than Huawei can yet supply [Reported: Bloomberg, no company confirmation found in this pass]. At the indicated price, 160,000 chips times 250,000 yuan is roughly 40 billion yuan for one lab, before order-size discounts and auxiliary equipment.

The memory maker has cash and no named HBM line. CXMT’s May declaration draft named three projects for 29.5 billion yuan of proceeds: a memory wafer line upgrade, a DRAM technology upgrade and DRAM forward research. None is called HBM. The raise then came in larger. The final proceeds announcement of August 27 puts gross proceeds at 66.6 billion yuan and net at 66.3 billion, and says the amount above the planned figure will be used after review and disclosure. Net proceeds of 66.3 billion less the 29.5 billion planned leaves 36.8 billion yuan not yet assigned. The excess exists because investors paid 8.66 yuan a share against about 4.41 implied by the draft’s plan, which is 29.5 billion divided by 6.69 billion planned shares, and the stock stayed above 8.66 yuan through the exercise window, so the over-allotment ran in full. Cash does not look like the constraint at CXMT. Its late-August half-year filing shows 150.3 billion yuan of revenue, 117.1 billion of profit before tax and 131.2 billion of operating cash flow, all unaudited. The May draft attributes the price surge behind such numbers to compute demand and larger makers reallocating capacity, so on this reading the same memory shortage that lifts Huawei’s price sheet is being paid to CXMT as DRAM margin. In the documents read, neither company says whether CXMT supplies Huawei, and the project names do not say whether forward research includes HBM. The filings show cash in hand. They do not show HBM output, which is where Reuters’ sources place the shortage, and no document read this week quantifies it. This item rests on the May draft and the August announcement; the July listing prospectus and the full half-year report were not checked this session.

None of this says Huawei’s roadmap is hollow or that CXMT will not build HBM. Xu named optical components alongside memory, Huawei’s manufacturing output is also limited [Reported: Bloomberg], and CXMT’s margin reflects a DRAM cycle that would run without Huawei. This issue follows memory because that is where the price reporting points. Huawei executives also said the company now holds a larger share of China’s AI chip market than Nvidia [Reported: Bloomberg], the 960DT date may hold, and CXMT’s board may yet assign the 36.8 billion yuan to HBM. The gap is between a date Huawei has announced and a memory supply chain that no document quantifies.

What would settle it. The nearest test is CXMT’s third-quarter report, expected by the end of October, together with the board’s plan for the unassigned 36.8 billion yuan. An HBM line in either would show domestic memory being funded, which is the precondition for the 950DT’s price to ease. Money parked, or spent on commodity DRAM, would make Huawei’s price sheet the durable reading: supply, not capital, sets 2027 volume. The second test is Huawei’s fourth quarter. If the 950DT ships in the quarter Huawei promised but the indicated price holds above 250,000 yuan and DeepSeek’s order stays unfilled, the July ceiling has moved from date to volume and has not lifted. If the price falls back, this issue was early. A cheaper tell is whether Huawei restates the 15,488-chip pod on its roadmap, which would say the 960E is a step and not a replacement.

ICM Weekly is research, not investment advice.

The cost of a factory robot includes the people needed to keep it working. A lower purchase price reduces one part of that cost. Better autonomy can reduce another, repeatedly, for as long as the machine remains in service. Which improvement matters more depends on how much human support the task requires and whether the factory can actually remove the associated expense.

UBTECH provides a useful place to ask the question. On 21 January 2026, Reuters reported UBTECH’s statement that Airbus had purchased Walker S2 robots. Airbus told Reuters that the collaboration was at an early concept-testing stage. That description belongs to January; it should not be read as a verified account of the project’s position today. Reuters

A purchase establishes commercial interest. A production record would establish something different: how much acceptable work the machine completes, with how much assistance, at what cost. Both matter, but only the latter can substantiate a labor-saving calculation.

The financial stakes have become more concrete. UBTECH’s unaudited interim results for the first half of 2026 report RMB590.3 million in revenue from full-size embodied-intelligence humanoid robot products and solutions, representing 46.5% of group revenue. The category includes products and solutions; it is not a Walker S2-only revenue line. UBTECH interim results, printed page 3

For this article, the pivotal relationship is human support minutes per productive robot-hour. It connects a model’s ability to recover from an awkward situation to the factory’s operating bill. It also forces a distinction that shipment and revenue figures cannot make on their own: whether each additional installation requires another substantial allocation of human effort.

What this establishes

The filing establishes a reported revenue category; the product pages establish what UBTECH says its robots can do. Neither supplies a customer-validated staffing requirement. The cost comparison below is calculated from illustrative assumptions. Its economic inference depends on the weakest link: less assistance must reduce an actual expense or increase valuable output at comparable quality. Whether that happens for Walker S2 remains unresolved.

This continues ICM’s examination of what a robot-hour costs. Here the boundary is production: the human work needed to deliver an acceptable box, and the party that pays for it. The earlier article examines humans continuously driving robots to collect training data. Here, the question is how much intermittent assistance a robot needs while doing production work. A collector-to-robot ratio from the first setting cannot establish a support ratio in the second.

Define the work before counting the robots

Start with a box.

UBTECH’s Chinese industrial-solutions page lists box handling among the Walker S series’ applications. It describes carrying boxes of different sizes between pallets and production lines. This is the company’s description of an application, not customer-validated evidence of a particular installation’s productivity. It gives us a bounded task to examine without inventing an Airbus workstation. UBTECH industrial applications

For an illustrative assessment, define the job as moving an acceptable box from an agreed pickup position to an agreed destination. The payload, travel distance, permitted damage, handoff conditions, and required completion rate all belong in the definition. So does the point at which a human must intervene.

The task sounds simple until its boundary becomes explicit. Does someone straighten the boxes before pickup? Who removes damaged packaging? Does the robot recognize an obstructed destination, or does someone clear it? If the machine completes the transfer but the next process cannot use the result, should the movement count as output?

These are questions for a task study, not allegations about Walker S2. Their purpose is to make the comparison reproducible. The same payload and acceptance standard must apply to the existing process and the proposed robot cell.

A demonstration can establish that a movement is possible. A trial can expose the conditions under which it fails. Accepted production output can establish usefulness. A repeat order can show that a customer wants more. None automatically supplies the missing measurements from the other stages.

Time needs the same discipline. A powered robot can be waiting. A moving robot can be repeating a failed action. A robot completing useful transfers can still require a worker assigned to the cell throughout the shift. Count acceptable units per scheduled cell-hour alongside human support time. Otherwise, a slow machine can look impressively independent while producing too little to justify its cost.

There is also a choice before the humanoid comparison begins. The alternative might be manual handling, a conveyor, a mobile platform with a manipulator, or a redesigned process. Compare feasible systems serving the same task. Human shape is a design choice whose commercial value must survive that comparison.

The resulting question is precise: how much avoidable cost does this installation remove while maintaining the required output, quality, and operating conditions?

Autonomy has a staffing boundary

Human work enters a robot installation at different times and for different reasons. Putting all of it into a single support figure would obscure what can improve.

Integration comes first: laying out the cell, connecting equipment, teaching the task, and validating the result. For the economic model, treat that expenditure as part of installed capital and spread it across a disclosed useful life. A deployment that requires extensive initial work can still be attractive if that work is durable and inexpensive to repeat elsewhere.

Recurring support needs another ledger. It can include required supervision, remote operation, exception recovery, maintenance, calibration, and validation after software changes. Record both active work and staffing that must remain available even when nothing goes wrong. Use actual obligations and operating records to decide which categories apply.

The distinction becomes visible in battery management. UBTECH says Walker S2 can replace its own battery within three minutes and describes a system that selects swapping or charging according to task priorities. These are vendor-stated capabilities, not independently measured factory performance in this article. Walker S2 product documentation

Autonomous replenishment could remove a recurring human chore. Its economic contribution would depend on the work it displaces, the equipment it requires, and the effect on productive time. The same reasoning applies to recovering a grasp or selecting another route. Each solved problem changes a specific part of the operating process.

That is where task intelligence meets the robot’s body. In a possible failure chain, uncertain perception could produce a poor grasp; a compliant gripper could make that uncertainty tolerable; a fixture could remove it before the robot acts. Better software, better hardware, and a more structured workstation can therefore improve the same measured outcome. Their costs and transferability may differ considerably.

The useful measurement is how much human involvement remains after the whole system has been designed. It should preserve the reason for each intervention, because a maintenance stop and a planning failure call for different investments.

Even a good average cannot establish a staffing ratio on its own. Imagine, purely as an illustration, a robot requiring six active support minutes per productive hour. Dividing sixty by six produces ten. It does not establish that a worker can support ten robots.

Several machines could request help together. The worker may need to walk between them, finish an existing recovery, or remain available for another duty. Some interruptions can wait; others may stop a downstream process. The average reveals labor demand, while the distribution of interruptions helps determine staffing.

A viable staffing ratio therefore needs a defined task, a permitted delay, and an operating record. It cannot be extracted from a success-rate headline.

Price the acceptable box

The economic calculation can stay simple even when the operating process is not:

Cost per acceptable unit = annualized installed capital cost plus annual operating costs, divided by annual acceptable output.

Installed capital includes the robot, peripherals, integration, and commissioning. Annual operating costs include attributable staffing, maintenance, energy, software, and service fees. If a service fee already includes a technician’s work, do not add that labor again.

Downtime usually enters through reduced acceptable output. Additional losses, such as scrapped material or disruption elsewhere in the plant, belong in the numerator only when they have not already been counted. The calculation should make it possible to trace every cost back to a payment, resource, or stated assumption.

Here is an Estimated illustrative scenario, constructed for arithmetic rather than inferred from UBTECH or any customer. All input figures are assumptions. Currency is RMB, with no claim that the amounts reflect prevailing Chinese prices or wages.

Assume a cell’s robot purchase costs RMB500,000 and other installed capital costs RMB100,000. Spread both over five years with zero residual value and no financing charge. Annual support staffing costs RMB180,000, other operating costs RMB60,000, and annual acceptable output is 600,000 boxes.

Annualized capital is then RMB120,000. Total annual cost is RMB360,000, or RMB0.60 per acceptable box. That establishes a baseline for sensitivity analysis, not a claim of attractive payback: we have not assigned a cost to the existing process.

Estimated scenario, not measured UBTECH performance. Other annual operating costs remain RMB60,000; annual acceptable output remains 600,000. Per-box results are rounded. Support savings assume an actual reduction in recurring expense without a reduction in throughput or quality. Costs of achieving the improvements are excluded and must be added for an investment decision.

Test the assumptions. The companion cost calculator lets you change the purchase price, installed capital, useful life, support expense, and acceptable output. Its default figures reproduce this illustration; they are not measured UBTECH costs.

Under these assumptions, the hardware discount saves RMB20,000 annually after the purchase saving is spread over the assumed life. The support reduction saves RMB36,000 annually. The second is larger because the support bill exceeds the annualized robot purchase cost, not because autonomy is intrinsically more valuable than hardware.

The break-even relationship makes that limitation explicit. For equal percentage reductions, unchanged output, and the simple capital treatment above, support savings exceed purchase-price savings when annual avoidable support expense exceeds robot purchase price divided by useful life. Different financing, residual values, upgrade costs, or output effects change the comparison.

The result can reverse. If annual support expense in the same example were only RMB60,000, reducing it by 20% would save RMB12,000. The identical purchase-price reduction would still save RMB20,000 annually. Hardware would matter more.

It can also fail to appear in cash. A model update that cuts active recovery time by 20% may leave the same worker assigned to the cell. Then the support-cost row does not apply. The benefit might instead be more output, less overtime, lower strain, or capacity to handle another task. Each has to be measured on its own terms.

Nor does this table compare the return on engineering investment. A hardware discount and a reduction in support expense may cost very different amounts to achieve. A factory evaluating an upgrade must include its price and disruption. A supplier deciding where to spend research money must include development cost and the number of installations over which it can recover it.

The table identifies the variable worth investigating. Operating evidence determines whether the assumed saving exists.

Follow the bill across the contract

Once a factory can produce the same acceptable output at lower cost, a second question opens: how much of that improvement reaches the robot supplier?

UBTECH’s interim filing offers a useful accounting boundary. Its expense-by-nature note aggregates costs across cost of sales, selling, administrative, and research expenses. It separately lists an after-sales repair-service accrual. Neither that note nor the broad humanoid revenue category supplies a Walker S2 installation-level support account. This reading concerns those specific disclosures, not an assertion that the company has published no operational information elsewhere. UBTECH interim results, printed pages 3 and 40

The contract determines how operating burdens are divided. Under an outright sale, a supplier could collect the purchase price while the customer pays for routine operation. Warranty and integration obligations can still leave substantial work with the supplier. Under a service agreement, a recurring fee may cover specified support while excluding other work. Under an output-based arrangement, the supplier may bear more of the risk that the robot produces too little.

Those are possible structures, not descriptions of UBTECH’s Airbus agreement. The commercial terms needed to make that attribution have not been established here.

The distinction matters because identical customer savings can produce different supplier earnings. A fixed service fee can become more profitable if reliable operation reduces the supplier’s actual support expense. A competitive renewal could pass much of that saving back to the customer. A discount on the next robot could do the same. Technical improvement creates value; pricing and obligations determine who retains it.

Cash timing adds another dimension. Where payment depends on installation or acceptance, a longer commissioning period can leave the supplier funding equipment and engineering work before collection. Where a customer prepays, the funding burden can shift. Without the terms, revenue alone cannot tell us which situation applies.

Before assigning the benefit to model improvement, test the other explanations against the same operating record:

These are competing hypotheses, not findings about UBTECH. Several could contribute at once.

My capital judgment is conditional but consequential: delivery growth deserves a stronger earnings interpretation when repeat installations require less incremental support and the supplier retains some of the resulting benefit. If each installation still requires substantial bespoke work, the model should carry that labor and cash requirement alongside the sales forecast.

The earnings interpretation should follow the cost structure of the business being sold. A hardware company, an integration business, and a provider of guaranteed productive work can all be valuable. They earn their returns through different obligations.

Assistance can purchase a better next deployment

There is a strong reason to tolerate substantial support early: the assistance may help train the system.

Berkeley’s HIL-SERL research provides a concrete mechanism. Its published workflow combines demonstrations with human interventions during reinforcement learning. The researchers describe reducing intervention as the learned policy achieves greater task success and faster cycles. This is primary research evidence for a learning method, not evidence about UBTECH’s implementation or commercial fleet economics. HIL-SERL research project

The commercial possibility is straightforward. A correction made at one installation could help prevent similar failures later. If improvements transfer, early support expense can contribute to a capability whose value extends beyond the original customer.

But training expenditure and routine operating expenditure must remain distinguishable. Calling assistance “data collection” does not establish that it improves the next deployment. The evidence would be a comparable task performed with less intervention, less setup work, or better output after the update.

A convincing comparison would track recovery frequency, recovery duration, acceptable throughput, and required staffing together. It would also record changes to the gripper, fixtures, layout, payload, and acceptance standard. Otherwise, an easier task could be mistaken for a more capable model.

Engineering the environment is a legitimate success. A fixture that makes box pickup reliable may produce a better return than a more general policy. Its commercial limits are different: the fixture may need redesign for a new box or a new workstation, while a transferable policy could reduce work across several settings. The cost of that difference belongs in the expansion plan.

The next installation is therefore a particularly informative test. Does it inherit the earlier improvement, or does another team spend comparable time teaching and adapting the system? Does the customer need fewer paid support hours, or do engineers simply complete a more demanding task with the same resources?

Both outcomes can create value. Only the first supports a claim that support requirements are falling for equivalent work. The second requires a different calculation, based on the additional capability sold.

The strongest positive case for humanoids is not that they never need help. It is that help produces improvements that can be reused, and that reuse eventually changes the cost of deployment.

The answer belongs to an operating record

How many robots can one worker support? The evidence reviewed here does not establish a numerical answer for Walker S2. It establishes a commercial case worth investigating, a documented application to examine, and a set of costs whose allocation matters.

The decisive disclosure would pair a customer-validated operating period with the relevant service terms. It would identify the task, acceptable output, scheduled hours, staffing, interventions, and downtime. It would show whether reduced assistance changed an actual expense or increased valuable output. The contract would identify which party paid for the work that remained.

That record could strengthen the investment case. Low support requirements and repeatable commissioning would make the purchase price a more useful guide to installed economics. A measurable decline in support across comparable deployments would give the learning argument commercial weight.

It could also weaken the particular thesis advanced here. If recurring support is already inexpensive, hardware and integration costs may dominate. If staffing cannot fall, autonomy may create value through throughput or capability instead of payroll savings. The analysis should follow the measured benefit.

At the factory floor, the layers of China’s physical-AI stack meet in a practical result. The model must choose an action, the body must execute it, and the cell must deliver acceptable work under a cost structure the customer can repeat. The supplier must be able to support that repetition profitably.

The next box matters because it tests whether all of that can happen again, with less help.

Research boundary: This article develops a task-level economic framework. It does not estimate UBTECH’s actual intervention rate, factory payback, or Walker S2 margin. Company disclosures, reported statements, and illustrative estimates are distinguished below.

Inside China’s Machine is independent research, not investment advice.

Two disclosures this week resolve into one pattern, and it is not the pattern either buyer wants investors to see. Unitree’s regulatory filing shows humanoid robots crossing half the company’s revenue for the first time, at a gross margin that held, while the bottom line fell anyway. A parallel investigation into Chinese military procurement shows the next demand story, the one meant to succeed the industrial-deployment story this publication has already priced apart, running the identical first act: heavy documentation of intent to buy sensing and data systems, no confirmed record of an actual deployment. The market is being asked to pay for deployment twice, on two different ledgers, before either one has produced a single fielded unit doing the job it was bought for.

The filing sets the pivot number. Unitree’s response to the Shanghai Stock Exchange’s audit implementation letter, dated May 20 and read against the underlying figures this session, shows humanoid robots reaching 51.78 percent of 2025 revenue, the first year the category has carried the majority of a business long built on quadrupeds. Full-year 2025 revenue came to 1.699 billion yuan on a 60.44 percent gross margin, and R&D spending more than doubled year over year to 144.97 million yuan. None of that explains the first quarter. Non-GAAP net profit fell 52.55 percent year over year in Q1 2026 even as revenue grew 68.49 percent, and the company’s own interim disclosure in August showed the pattern held into the half: revenue up roughly 48 percent, non-GAAP profit still down about 19 percent [figures for the August interim disclosure via company-results reporting, not the raw filing checked this session]. The stock listed on August 19 at a 219 times issue multiple and closed its debut day up 629 percent, a gain that puts a market value near 66 billion dollars behind a business whose cost base is currently growing faster than the higher-margin product line that is supposed to fix it. Core components are more than 90 percent self-developed, per the same filing, which is the strongest argument for why the margin should eventually hold. It is not evidence that it has. The market priced the mix shift as a margin repair. The filing shows the mix shift arriving with a profit decline attached, not instead of one, which is a different and cheaper thing to have bought.

The pattern repeats one layer up. A Reuters investigation published September 7 reviewed more than 100 Chinese military procurement notices, patents, academic papers, and defense-contractor materials on humanoid robots, and found the same evidentiary shape as the industrial story: a National University of Defense Technology tender from June 2025 for an “embodied humanoid robot intelligent perception and dexterous operation system,” a roughly 300,000 dollar contract for a camera-and-radar training-data collection rig, and, at August’s World Robot Conference, defense manufacturer Norinco unveiling a teleoperated humanoid called Fuxi built for sentry, reconnaissance, and patrol roles [Reported: Reuters investigation, September 7; procurement notices themselves not independently checked this session]. What the more than 100 records do not show, per that investigation, is a single confirmed operational deployment. Chinese manufacturers already hold roughly 95 percent of global humanoid shipments, per BofA Global Research, so the hardware capacity to arm this narrative already exists. What does not yet exist is evidence that any of it has moved past a sensor-and-data acquisition phase into a unit actually standing a post. That is the same gap the industrial-deployment story carried before this publication went into the prospectus and found it: procurement volume mistaken for procurement outcome.

The competitive control group is arriving on the same terms. Leju Intelligence, a second humanoid maker, cleared the “already inquired” stage of its ChiNext listing review in late May and is working through the exchange’s questions now. Its reported financials show why the timing matters: revenue near 54 million yuan in 2023, 55.5 million in 2024, and then roughly 258 million in 2025, a near-quintupling in one year, while net losses widened each year and stood at about 71 million yuan in 2025 [Reported: company-profile and press aggregation of the ChiNext filing; not verified against Leju’s own inquiry-response document this session]. Leju is proposing to raise 2.6 billion yuan on a growth curve steeper than Unitree’s and a margin structure that, unlike Unitree’s, has not yet crossed into profitability at all. Tencent and Shenzhen Capital Group sit among its shareholders, which is the same kind of strategic-backer signal that helped set Unitree’s price, and it will invite the same read: growth this fast is treated as evidence of demand rather than as a company still subsidizing its way to scale. If Leju prices anywhere near a comparable multiple, the market is not making one bet on whether humanoid deployment is real. It is making the same bet twice, with two different companies supplying the growth number and neither supplying the deployment number yet.

None of this says the thesis behind either valuation is wrong. Revenue at both companies is real, audited, and growing faster than almost anything else in Chinese hardware. The gap is specifically between the deployment story being sold and the deployment evidence on file, in both the industrial version and now the military one, and gaps of that shape do eventually close in one direction or the other.

What would settle it. Unitree’s third-quarter report, due by China’s standard late-October filing deadline, is the nearer test: the market is pricing in a second-half inflection in the cost ratio, and a quarter that still shows expenses outrunning revenue would mean the margin story, not just the deployment story, needs re-pricing. The farther and larger test is a single confirmed instance, in a company disclosure or an official military record rather than a procurement notice, of a Chinese-made humanoid actually performing an operational role rather than being purchased to learn how to perform one. Either disclosure would move a number this publication is already tracking. Neither has happened yet.

This is research, not investment advice.

XPeng began small-batch trial production of a humanoid on Friday, and the interesting fact is where the machines are going. Not to a factory floor. To 487 car showrooms, to greet customers and explain vehicles. That is the shape of this layer right now: capacity is arriving on schedule, into places that generate no manipulation data. So the document that matters this week is not a production announcement. It is a ministry notice from June whose first physical instances went up at WAIC on Monday, and what it allocates is not money but access to workplaces. Read against the three prospectuses this publication has now read in full, it says China’s answer to the trajectory shortage is administrative rather than commercial, and that changes what a model earmark is worth.

The capacity is real and it lands in the wrong place. XPeng’s IRON entered small-batch trial production at its Guangzhou plant on 24 July, with the mass-production line in final commissioning, a target of over 1,000 units a month by year-end and roughly 5,500 units planned across 2026. First deployment, in Q1 2027, is into 487 company stores for greeting and guiding. [Reported: company statements via Guandian and Sina, 24 July; not a filing]

Hold that against the market structure. Deep Robotics’ prospectus carries IDC’s split of the roughly 18,000 humanoids shipped globally in 2025: four scenarios take close to 90%, being entertainment performance, research and education, data collection, and guided retail. Industrial manufacturing takes about 9%, so the whole world put on the order of 1,600 humanoids into manufacturing work last year. XPeng’s single-company 2026 plan is more than three times that, aimed squarely at one of the four. Shipments are compounding. The share of them touching an unstructured object is not.

The state wrote the diagnosis into the operative section of a document. MIIT and the SASAC issued a joint notice on 3 June, number 256, launching a real-scenario training programme for humanoids and embodied intelligence. Its stated purpose is to accumulate high-quality real-machine data through training in genuine environments, and its end-2026 target is to move representative products out of demonstration into what it calls operating mode, at ten-thousand-unit scale.

The allocation clause is the part that has not been widely read. The notice is addressed to ten provincial industry departments, Beijing, Tianjin, Shanghai, Jiangsu, Zhejiang, Shandong, Hubei, Hunan, Guangdong and Sichuan, and to central state-owned enterprises. Each province must select no fewer than 20 real scenario units across at least two of the industrial, service and special-operations domains. Each central SOE must select no fewer than 10. That is a floor of 200 workplaces from the provinces alone. The scenario unit is defined physically: a production workstation, a service point, an emergency response station.

And it specified the file format. The third task instructs the consortia to build high-fidelity datasets, then names the fields: whole-body motion trajectories, force-position control curves, operation execution sequences and their temporal logic, plus spatial semantics, object attributes, and the handling of anomalies, interruptions and edge conditions.

That is a trajectory-data specification written by a ministry, describing the same object the trade quotes at 500 to 1,000 yuan an hour more precisely than any commercial listing this publication has found. A state that subsidises an industry funds capacity. A state that writes the schema is specifying a product.

The instrument the state holds is the door. Collection rigs can be bought and collectors hired at the wages documented earlier in this series. What cannot be bought is permission to instrument a working substation, a hospital ward, a casting line or a fire station, because the owner of those premises has no commercial reason to let a robotics company record inside them. The notice solves that by instructing the owner to open up and making it a named party to the consortium.

The first instances are physical. At WAIC on 20 July the National and Local Co-built Humanoid Robot Innovation Centre unveiled the country’s first embodied-intelligence training-ground demonstration site with Huawei, running on Ascend compute, alongside six scenario agreements spanning emergency rescue, power, retail, automotive manufacturing, city services and precision casting. Its Linglong robots are already inside Yanfeng, an automotive interiors manufacturer, doing frontline collection and autonomous-operation validation. [Reported: institution and partner announcements, 19 and 20 July]

Why the private build continues anyway. The notice requires each consortium to specify intellectual property ownership and division of benefits, and separately calls for the data to be opened and shared in an orderly way. Those provisions pull against each other, and the tension is the investable part. Shared data is non-exclusive, and non-exclusive data converges a model rather than differentiating it. The programme reliably produces parity, which is worth having only if the alternative is nothing.

Which is why Lejuu, whose prospectus this publication read last week, is doing something else in parallel. Its six collection joint ventures with local state capital, none majority-held, began forming in June 2025, a year before this notice. Those are staged facilities: build the scene, hire the collector, own the output. The ministry programme is the other production method, instrumenting workplaces that already exist. The first costs more per hour and yields exclusivity; the second is cheaper and yields a shared asset. A company running both is saying how little either delivers alone.

The consequence for the queue. The three companies in the listed queue have earmarked 48.13%, 46.72% and 59.84% of their raises for model development and the data to feed it, every figure from their own use-of-proceeds tables. Those earmarks assume trajectories will be available to spend the money on. Two supply mechanisms are now visible and neither is commercial: state-capitalised collection ventures, and administratively allocated premises. Both are access problems, not purchasing problems.

What follows is uncomfortable for how this sector is priced. Shipment leadership and data access have come apart. Unitree ships the most humanoids in the world, and that fact alone says nothing about its position in the trajectory supply, because the units go to researchers, performances and showrooms. Consortium membership and premises relationships are the asset, and they appear on no balance sheet, in no comparable-company table, and as a line item in none of the three prospectuses.

What to watch. The notice sets 30 November for provinces and SOEs to file outcome summaries and requires a scenario inventory throughout. The settling disclosure is any provincial publication of that inventory naming consortium members per scenario, because that shows who actually got the doors opened. Second, whether the orderly-sharing provision yields a dataset with a stated licence or stays inside the consortia. Those outcomes have opposite implications for whether this creates a public good or a private allocation.

Last week’s two open items remain open. Lejuu’s inquiry response has not appeared at Shenzhen, where the file list is still empty. Unitree’s registration took effect on 2 July with no pricing published, leaving the first public valuation of this layer unobservable four weeks on.

The MIIT and SASAC notice, 工信厅联科函〔2026〕256号, dated 3 June 2026, was read in full this week from the ministry text. Prospectus figures come from the filed applications of Unitree (STAR Market, 20 March 2026), Deep Robotics (STAR Market, 18 May 2026) and Lejuu (ChiNext, 19 May 2026), read in full in earlier sessions. Items marked Reported are company or institutional announcements, not checked against a filing.

Inside China’s Machine. China is building the machine that builds physical intelligence. Silicon, models, robots, factories. We read it one layer at a time and turn each into capital judgment.

This is investment research, not investment advice.

The Construction Closes

Four issues of this series have been assembling one number. What does it cost to manufacture one hour of robot trajectory data in China?

The equipment term came from Leju’s ChiNext filing: 577 Kuavo humanoids sold in 2025 at an average recognized price of 308,100 yuan, and 44.94 percent of that product line’s revenue came from data collection buyers. The labor term came from the wage ladder: 21 to 26 yuan per shift-hour of collector time, under a market where the resulting data changes hands at 500 to 1,000 yuan per hour. The compute term came from the card-hour market and the treasury voucher: 1.3 to 2.3 yuan per domestic card-hour after rebate, the cheapest of the three by a wide margin and the one export controls were supposed to make dear.

Three terms, all bounded. What was missing was the denominator. Every one of those costs is a cost per day or per unit, and converting them into a cost per usable trajectory-hour requires knowing how many trajectory-hours a robot and its operator actually produce in a day. That number is the hinge of the entire embodied-data investment case, and no filing in China states it.

It can now be bounded, and the bound is wide enough to change the conclusion. Using two disclosures from the same company, the operator of the largest such facility in the country by its own account, the labor cost of one usable trajectory-hour lands somewhere between 265 and 2,244 yuan. The 500-to-1,000-yuan selling band sits inside that range. Add the equipment amortization and most of the range sits above the band.

This piece is not the discovery of a hidden loss. It is the discovery that the industry’s own numbers cannot tell you whether the business is profitable at the gross level, and that the gap between its best case and its average case is a factor of eight, hiding in a variable nobody reports.

One correction to this series belongs at the top rather than buried. The labor piece bounded the labor term and said explicitly that yield was an unobservable multiplier it would not invent. That was a scoping decision and it was stated each time. This piece closes it. The same piece also said the equipment term amortizes to a small figure per hour at any plausible utilization. That was a conditional claim, the condition is now measurable, and at the utilization implied by the operator’s own cumulative disclosure the condition does not hold. The equipment term is not small. It is the largest of the three.

The Unit the Industry Reports In

Before any of this can be computed, the industry’s numbers have to be converted into a unit they are not published in.

Chinese embodied-data operators report output in trajectories, or clips. AgiBot’s data factory produces tens of thousands of clips a day per its own promotional material. Pasini’s Tianjin facility projects close to 200 million clips a year. Demand, meanwhile, is denominated in hours. The founder of the physical-AI data platform Mifeng, who is also the executive running AgiBot’s collection operation, estimates that reaching a GPT-3.5-like general capability in embodied models requires data on the order of 100 million hours, against a global effective supply he puts in the hundreds of thousands of hours, a gap of two to three orders of magnitude. That is an Estimated figure from an operator with a commercial interest in the scarcity he describes, and it is used here for its unit, not its magnitude.

Production is counted in clips. Demand is counted in hours. The conversion rate between them is published almost nowhere.

It is published in one place, and that place is a primary technical document. The AgiBot World Colosseo paper on arXiv states the dataset’s size in both units: 1,001,552 trajectories with a total duration of 2,976.4 hours. That is 10.7 seconds per trajectory, and roughly 336 trajectories to the hour. The number is corroborated by the only other disclosure this series has found that gives both units, the Jiangsu exchange listing examined earlier, at 25,000 clips of about ten seconds each. Two operators, two document types, the same answer.

Apply the conversion and the industry’s headline volumes deflate by a factor of three hundred. AgiBot World, described in its own paper as the largest trajectory dataset released to date, is 2,976 hours. Pasini’s projected 200 million clips a year, converted at AgiBot’s rate and flagged as a cross-source extrapolation because Pasini does not publish clip durations, is on the order of 590,000 hours a year. Against an operator’s own estimate of the requirement, the largest open dataset in the field is three parts in a hundred thousand.

A clip is not a small hour. It is ten seconds. The industry reports in clips because clips is where the numbers look like industrial output, and hours is where they look like a research project.

What One Robot Actually Produces

Now the denominator, from three disclosures by one company.

The staffing ratio first, because it decides everything downstream. Standing in the facility in May, AgiBot’s embodied business president told visiting press that the site runs 200 machines and that each machine is staffed by at least one collector, with some tasks adding a second person to reset the scene. Not one operator supervising a bank of robots. One operator per robot, sometimes two. A training ground in Shandong reported in January runs 33 collectors against 31 robots across 28 stations, a ratio of 1.06. Tesla’s collection seat is described the same way, one robot plus one motion-capture rig plus one operator. Three facilities on two continents converge, because teleoperation is a human driving a machine in real time and there is no version of that where one human drives five.

The yield mechanism is documented too, and it is stricter than this series assumed. At the Shandong site, collectors work about 7.5 actual hours and must hit a daily quota of effective collection time, where effective is defined by the output passing standard. A take that runs thirty seconds and conforms counts thirty seconds. A take that fails is redone and the time is not counted at all. Collectors there repeat a single action sequence more than a thousand times in one scenario.

Now the output rates, and they do not agree with each other.

The station rate. A reporter visiting AgiBot’s facility in June 2025 watched a drinks-shop station collect about 200 clips a day, with the item positions changed every take and the cup and bag styles rotated every ten. At 10.7 seconds a clip, that station produced about 36 minutes of trajectory data in a day.

The facility rate. The same executive told the same reporter that the facility, running since September 2024, had accumulated over one million clips. Across roughly nine months that averages about 3,700 clips a day, which is 11 hours of trajectory data a day across the whole building. A separate visit in February 2025, when the site ran 100 robots, reported the facility completing just over a thousand clips a day, which is under three hours.

Put the two rates side by side and the discrepancy is the finding. The station rate multiplied by a hundred robots would be 20,000 clips a day. The facility’s own reported output is between 1,000 and 3,700. The working station is running five to twenty times faster than the average robot in the same building.

That difference is utilization, and it is the number the entire sector declines to publish. It is scenario rebuilds, equipment failures, staffing gaps, robots down for maintenance, stations between tasks, and the simple fact that a facility with a hundred robots does not have a hundred robots collecting on any given day. Every projection of a data factory’s annual output that this publication has seen is built by multiplying a station rate by a station count. The operator’s own cumulative disclosure says that multiplication overstates by most of an order of magnitude.

That completes the free layer: the conversion rate established from the primary document, the staffing ratio confirmed by the operator, and the utilization gap measured against the operator’s own two numbers. The paid layer assembles the three cost terms against that denominator and prices the robot-hour.

Assembling the Robot-Hour

The construction runs at two utilizations, because the operator published two, and this piece will not pick one for them.

At the station rate, 36 minutes of trajectory data per robot-day, one usable trajectory-hour consumes about 1.7 robot-days. With one collector per robot on a 7.5 to 9.5 hour shift, that is 13 to 16 collector-hours, and at 21 to 26 yuan per shift-hour the labor term is 265 to 416 yuan.

At the facility rate, 6.6 minutes per robot-day, one usable trajectory-hour consumes about 9.1 robot-days. That is 68 to 86 collector-hours, and the labor term is 1,431 to 2,244 yuan.

The equipment term moves with the same denominator and moves harder, because a robot depreciates by the calendar whether or not it collects. At Leju’s filed 308,100 yuan and a three-year life, the instrument costs 281 yuan a day. Spread over 36 minutes of output, that is 473 yuan per usable trajectory-hour. Spread over 6.6 minutes, it is 2,555. A five-year life, generous for a machine whose recorded data does not transfer across embodiments, gives 284 and 1,533. The three-year figures are used below and the five-year figures are the sensitivity.

The compute term remains what the previous piece found, tens of yuan per usable hour at plausible intensities, and it is now visibly the rounding error in the construction rather than its center.

Add them. At the station rate, one usable trajectory-hour costs roughly 740 to 890 yuan before overhead, annotation, premises, and quality inspection. At the facility rate, it costs roughly 3,990 to 4,800 yuan. The market pays 500 to 1,000.

The conclusion follows without needing to choose between the two rates, which is what makes it usable. At its best observed station utilization, a Chinese data factory produces a trajectory-hour for slightly less than the top of the price band and more than the bottom of it, before any cost that is not labor and the robot itself. At the utilization its own cumulative output implies, it produces one for four to nine times the price band. There is no version of these numbers in which the gross margin is comfortable, and the most defensible version, resting on a cumulative figure rather than a single good day at a single station, says the activity loses money at the going rate.

This inverts the reading this series published two issues ago. That piece described the gap between a 26-yuan wage and a 500-yuan price as a window rent being closed from beneath by wages and from above by supply. The rent was measured per input hour, which was the honest measurement available at the time and was labeled as resting on an unobserved yield. Measured per output hour, at the utilization the operator has since disclosed, the rent may never have existed. What looked like a spread was a conversion rate that had not been applied.

Who Is Actually Making Money

If collecting the data does not clear its cost, the economics of everything this series has traced have to be re-read, and they resolve cleanly.

The money in China’s embodied-data economy is in selling the instrument, not in operating it. Leju’s own filing is the demonstration. Its flagship humanoid sold 577 units in 2025 at 308,100 yuan each, revenue on that line grew roughly twelvefold, and the single largest application category, at 44.94 percent, was data collection, sold into buyers named in the filing as municipal industrial development companies and a state research institute. Leju is not exposed to the price of a trajectory-hour. It is exposed to the budget cycles of the entities that buy the machines that produce trajectory-hours, and this series has already shown that those budgets carry the shape of a fiscal year rather than a demand curve.

The buyers are the ones holding the utilization risk, and most of them are public money. That is the capital judgment this construction produces. A state-funded data collection center underwritten on a station-rate projection has an asset whose output, at the facility rates the sector’s most advanced operator has disclosed, is worth a fraction of what the business case assumed. The gap is not a margin compression. It is a factor of five to twenty on the volume side, before the price question is even reached.

Three of the sector’s IPOs route 41 to 51 percent of their combined raise toward model development, and the previous piece established what that money must buy: not card-hours, which are cheap and discounted, but trajectory-hours, which are expensive to produce at realistic utilization and cannot be bought at scale because the supply this piece just measured is a few thousand hours at the frontier. The DeepMind comparison makes the physical scale concrete. Thirteen robots ran for seventeen months to produce roughly 130,000 trajectories, which at this conversion is about 390 hours. The bottleneck is not capital and it is not silicon. It is that a trajectory-hour requires a human hour, in real time, at a ratio of one to one, and no amount of capital compresses a ratio of one to one.

The position this implies is not a short on the sector. It is a preference along the chain. Exposure to the instrument, sold at a filed price into budget-backed buyers, is exposure to a disclosed number. Exposure to the operation of a collection center is exposure to a utilization rate that its own industry does not publish and that the one operator who has published enough to compute it appears to be running at a fraction of its own station-level claims.

The Disclosure That Settles It

Three documents would resolve the denominator, in the order this publication expects them.

First, Leju’s inquiry response at the Shenzhen exchange, still unpublished. The exchange has every reason to force disaggregation of the data-center contracts behind 44.94 percent of the flagship line, and any decomposition that separates hardware from data-service performance obligations reveals whether the buyers are contracting for machines or for hours. If they are contracting for hours, the contract prices those hours, and the first filed trajectory-hour price in China arrives inside a robot company’s inquiry response rather than at a data exchange.

Second, a procurement award from any of the state-funded collection centers that prices data collection per hour or per trajectory. The same document class this series has named twice would settle both the price and, if it specifies volume commitments against a facility size, the utilization.

Third, an operator publishing hours alongside clips. AgiBot has done it once, in a research paper, and that single act of dual-unit disclosure is what made this entire construction possible. Any facility that reports annual output in hours rather than clips is either confident in its utilization or has been made to state it.

Whichever arrives first, the test is the same. This piece says the cost of a Chinese robot-hour is set by a human hour at one-to-one and an instrument depreciating on a calendar, that the resulting cost straddles and probably exceeds the price the data sells for, and that the margin in the sector therefore sits with whoever sells the machine rather than whoever runs it. The three terms are assembled and the denominator is bounded. The first disclosure that narrows it will say which end of a factor-of-eight range this industry actually operates at, and that single number decides whether China’s data factories are an industry or an expenditure.

Inside China’s Machine is research, not investment advice. The trajectory-to-hour conversion is Confirmed from the AgiBot World Colosseo paper on arXiv, which states 1,001,552 trajectories and 2,976.4 hours; note that the project’s repository and successive paper versions give slightly different trajectory counts, and the paper’s figure is used throughout. The staffing ratio, the station rate, the facility’s cumulative output, and the promotional daily rate are all Reported from Chinese press coverage of site visits and interviews with AgiBot’s embodied business president, secondary and not independently verified, and they conflict with one another by design of this argument rather than by resolution of it. The inference that the open dataset approximates the facility’s total collection to that date is this publication’s, not the company’s, and if the internal store is materially larger the facility rate rises and the labor and equipment terms fall proportionally. Shandong training-ground staffing, shift length, and the effective-time definition are Reported from a January 2026 provincial media account. Leju’s unit price, volume, and 44.94 percent application share are Confirmed from its ChiNext prospectus as carried in this series. Wage figures, the 500 to 1,000 yuan selling band, and the compute term are carried from earlier pieces, where their verification status is stated. Depreciation lives are this publication’s assumptions and are run at three and five years. Demand-side hour estimates are Estimated and sourced to an operator with an interest in the scarcity described. Current as of July 24, 2026.

The Third Term

The robot-hour cost construction this series is building needs three observable terms: what the equipment costs, what the labor costs, and what the compute costs. The equipment term was bounded from Leju’s filing. The labor term was bounded in the last piece, a floor of 21 to 26 yuan per shift-hour of collector labor beneath a market that sells the resulting data at 500 to 1,000 yuan per hour. This piece bounds the compute term, and the finding inverts the premise most readers will bring to it.

The premise is that compute is where China’s embodied-AI economics must break, because compute is the input Washington restricted. The observable prices say the opposite. Of the three terms, compute is the only one with a live, quoted, competitive rental market. Per usable trajectory-hour, it is the smallest of the three by the construction below, an order of magnitude under the labor term at domestic card rates. And it is the only term the state directly discounts at the point of purchase, with money raised through ultra-long special treasury bonds, at 10 or 30 percent off the invoice. The tight input in China’s robot-data economy is not the sanctioned one. It is trajectory data, which no export control touches, which has a wage but no public price. The sanctioned input is the slack one, and the distance between those two facts is what this piece measures.

One institutional discovery runs underneath the arithmetic, and it completes a pair with the last piece. The Jiangsu data exchange hosted China’s first embodied-dataset trade and withheld the price, a market demonstrated without the number a market exists to produce. The compute voucher does the mirror image. To claim the rebate, a buyer must file the bill with the actual card model exposed, the price, the scale, and the duration, and the state checks that filing against an internal price cap before paying out. Claim by claim, the rebate machinery is assembling a national census of real transacted card-hour prices, by card model, and publishing none of it. One arm of the institutional machine performs a market without a price. The other compiles the prices without a market display. Same state, same year, same missing page.

Two Curves, Crossing

The data curve and the compute curve crossed in 2026, and they crossed moving in opposite directions.

The data side was the last piece’s subject: a selling band of 500 to 1,000 yuan per hour that is a scarcity rent, being industrialized against from below by wages and from above by open-sourced and self-collected supply. That band, per trade reporting, points down.

The compute side spent two years pointing down and then turned. The decline first, with its source stated: per a Kezhi Consulting series carried by an industry conference organizer, secondary and not verified against the underlying contracts, monthly rental on an eight-card A800 server fell from 60,000 yuan in June 2023 to 28,000 in June 2025, a 53 percent drop, H800 servers fell 34 percent to 66,000, and H20 fell 17 percent to 25,000 over its shorter life. That is the falling series, and it ends in mid-2025. It cannot be spliced onto what follows as one curve, because what follows is a different market.

From late 2025 the direction reversed, and the reversal is stratified by sanction status. Per a 36Kr market report of June 2026, the one-year lease rate on an H100 rose from 1.70 dollars per card-hour in October 2025 to 2.35 dollars in March 2026, a rise of close to 40 percent, H200 rates reached 7.5 to 8.0 yuan per card-hour, and delivery windows for new high-end capacity stretched into 2027, which is the tell that the tightness is real rather than quoted. A rental vendor’s market commentary, a source with a commercial interest in the direction it reports and read accordingly, describes the same structure from inside the trade: Nvidia high-end tiers up 30 to 40 percent, mid tiers up 15 to 20, and domestically adapted packages built on Ascend 910B and 920 and Moore Threads parts up only 10 to 15 percent, with one regional market quoting domestic GPU rental at 1.8 yuan per card-hour against 2.5 for an A100. Every figure in this paragraph is Estimated, trade-sourced, and disclosed as such, because no filing prices a card-hour, which is precisely the institutional condition the next section examines.

Read the stratification, because it is the mechanism. The tiers that are rising fastest are the ones export controls make scarce: restricted silicon, allocated by long contracts, hoarded by the largest buyers, ByteDance reportedly planning to put half of roughly 160 billion yuan of 2026 capital spending into AI chips, a Projected figure from the same market reporting. The tier rising slowest is the one no restriction touches, domestic accelerators in ample supply. For a buyer who can only scale on domestic silicon, which is the legally durable position for every state-funded data collection center this series has traced, the relevant price is the bottom tier, and the bottom tier is both the cheapest in the market and the most insulated from the repricing above it.

One discipline before any of these numbers is used: a card-hour is not a commodity. An H200 hour, an A100 hour, and an Ascend hour do different amounts of work per hour, and comparing their prices without an equivalence basis is the definitional error this publication maintains a rule against. This piece therefore never nets the tiers into one compute price. It carries them as tiers, and the construction below runs on the domestic tier alone, because that is the tier the robot-data economy can actually buy at scale.

The Voucher and the Price Book

Now the fiscal layer, from the one document in this piece read in the primary this session.

On May 1, the Zhejiang Provincial Development and Reform Commission published the claim notice for the 2026 national artificial intelligence voucher, the compute voucher. The design, from the notice itself. The program is national, funded from ultra-long special treasury bonds, the same instrument that funds the country’s flagship infrastructure. It is a post-subsidy: the buyer rents compute first, on a real commercial contract, and claims afterward. Eligible claimants are universities, research institutions, and companies with genuine intelligent-compute rental demand, covering cloud and bare-metal resources, for contracts running May 2025 through April 2026. The support rate is tiered at 10 or 30 percent of the verified contract amount.

Two clauses in the notice carry more than their administrative weight.

The first: the subsidy is computed after a price verification, and where the verified amount sits at or above a price ceiling, the payout is capped at the ceiling. The state, in other words, holds a reference price list for rented compute, detailed enough to verify invoices against, and the notice does not publish it. A ceiling price that caps subsidies is an administered opinion about what a card-hour should cost, and it exists, in a table, somewhere in the program’s files.

The second: the claim materials must prove the card model, the price, the compute scale, the rental period, and the use, and the bill must expose the actual card type. Combine the two clauses and the shape of the machinery is unmistakable. Every claim cycle, the program ingests a province-by-province, model-by-model record of real transacted compute prices, verified against contracts, invoices, and payment records, and checks it against an internal ceiling. This is a price census running inside a subsidy, and it is the exact dataset whose absence defines the gray market described above, where the going rates live in vendor commentary and conference decks because no public document prices a card-hour.

Set this beside the Jiangsu trade and the pairing does the analytical work. China built data exchanges to move pricing into public view, and the first embodied trade on one withheld its price. China built a compute voucher to cut costs, and in operating it the state accumulates precisely the price transparency the exchanges were supposed to create, held privately. The institutional machine is not failing to produce prices. It is producing them selectively, and keeping them. For an investor, the consequence is direct: in both of the robot-hour’s non-labor input markets, the best price data in the country now sits in state files, and every disclosure event that leaks a row of it, a procurement award, an inquiry response, a published ceiling table, moves the informational floor of the whole sector at once.

That completes the free layer: the market stratified, the fiscal wedge documented, the price book located. The paid layer prices the compute term for the robot-hour construction and ranks the three terms.

Pricing the Compute Term

The construction follows the same honesty structure as the labor term: an observable floor, an explicit multiplier for what cannot be observed, and no invented precision.

The observable floor is the domestic card-hour: roughly 1.8 to 2.5 yuan at the regional quotes above, Estimated, trade-sourced. The voucher takes 10 to 30 percent off for an eligible claimant, putting the effective domestic rate at roughly 1.3 to 2.3 yuan per card-hour. Both ends of that band are soft, and the softness does not matter, because the term’s size relative to the other two is decided at any point in the band.

What cannot be observed is intensity: how many card-hours the embodied-data pipeline consumes per usable trajectory-hour, across preprocessing, training runs amortized over the dataset, and the light inference load of teleoperation itself. No filing states this number and this piece will not invent it. Call it k, card-hours per usable trajectory-hour, and hold it as an explicit unknown greater than one.

The construction is then a single line: the compute term equals k times 1.3 to 2.3 yuan. And the line is decisive without k being known, because of where the other terms sit. The labor term landed in the low hundreds of yuan per usable hour under pessimistic yields. For the compute term to reach even the bottom of that range, k would need to exceed roughly fifty domestic card-hours consumed per single usable trajectory-hour, sustained across the whole pipeline. For it to reach the 500-yuan bottom of the data-selling band, k would need to exceed two hundred. Nothing in the model layer’s own filed economics suggests embodied training burns at anything like that ratio against its data intake, and the efficiency direction this series tracks at the model layer runs the other way. Even granting the 2026 repricing full room to run, a doubling of domestic rates, several times the 10 to 15 percent the domestic tier has actually risen, moves the thresholds to roughly twenty-five and one hundred. The ordering does not flip.

So the ranking the keystone needs is now set, and it is the piece’s capital judgment. Per usable robot trajectory-hour: data sells in the hundreds to a thousand yuan, labor costs in the low hundreds, compute costs in the tens at plausible intensities, and equipment amortizes to less than that at any utilization a funded center can sustain. The robot-hour is a labor product with a data margin, and compute, the input the entire China-AI discourse treats as the binding constraint, is its smallest line.

Two model-layer anchors from this series’ filed record show what the opposite structure looks like, and why the embodied layer is not it. Zhipu’s audited first half showed 1,145.1 million yuan of compute service fees against 23.9 million of capital expenditure, a 48-to-1 rent-over-own ratio, and MiniMax’s cost of sales ran at roughly 98 percent compute, invoiced from Alibaba Cloud. That is what a compute-dominated cost structure is: it appears in the filings, unmistakably, as the largest line the auditor signs. No embodied-AI filing this series has read shows that shape. The three robot IPO prospectuses route 41 to 51 percent of their combined raise toward model development, and the last two pieces have shown what that money must actually buy: not primarily card-hours, which are cheap, discounted, and in ample domestic supply, but trajectory-hours, which have a wage, a gray-market price band, and no public market. Capital raised against a compute story will be spent on a data problem.

The sensitivity that matters is not k. It is the two policy variables on the cheap tier: whether the voucher renews past the April 2026 contract window, a program decision on treasury-bond money that can end as administratively as it began, and whether domestic card supply stays ample as the high tiers reprice, which is a substrate question this publication tracks at the fab level. Both risks push the compute term up. Neither, at the magnitudes in view, reorders the terms.

The Document That Reorders the Terms

Three disclosures would test this piece’s construction, listed in the order this publication expects them.

First, Zhipu’s Shanghai listing process. The STAR Market inquiry will force the full-year cost-of-sales decomposition and cloud unit economics that Hong Kong never required. That filing prices the model layer’s compute at audited resolution, and any embodied-adjacent line inside it, data services purchased, trajectory data licensed, becomes the first filed number adjacent to this piece’s k.

Second, a procurement award that prices intelligent-compute services per card-hour or per P. The state-funded data centers rent compute with fiscal money, fiscal money publishes tender results, and the first award notice with a card-hour unit price puts a Confirmed number under the domestic tier this construction rests on. The same document class this series named as the likely first publisher of a data price would, in one award, publish a compute price too.

Third, the ceiling table itself. If any claim cycle of the voucher publishes its price caps by card model, the price book comes out of the file, and the gray quotes this piece was forced to build on become checkable against an administered benchmark in a single afternoon. That publication would be the compute market’s equivalent of the exchange printing a price, and it is the least likely of the three, for the same institutional reasons the Jiangsu trade exhibited.

Whichever arrives first, the test is the same. This piece says the sanctioned input is the cheap one and the free input is the dear one, and that the robot-hour’s cost structure is labor and data wearing a compute story. The first filed card-hour price, the first audited data-purchase line, or the first published ceiling will say whether the construction holds. The equipment, labor, and compute terms are now all bounded. The keystone assembles them, and it is the next piece in this sequence.

Inside China’s Machine is research, not investment advice. The compute voucher’s design, funding source, subsidy tiers, price-cap verification, and card-model disclosure requirements are Confirmed from the Zhejiang Provincial Development and Reform Commission’s claim notice of May 1, 2026, read this session via its official publication. All market rental rates are Estimated from secondary trade sources and none appears in a filing: the 2023 to 2025 declining series is per a Kezhi Consulting compilation, the 2026 repricing and the 160 billion yuan capital-spending plan are per 36Kr reporting of June 2026, and the tier structure and the 1.8 and 2.5 yuan regional card-hour quotes are per a rental vendor’s commentary, a source with an interest in the direction it reports. The intensity multiplier k is an explicit unknown, not an estimate. Zhipu’s 48-to-1 compute ratio, MiniMax’s cost composition, the robot IPO raise allocations, and the labor and data figures are carried from earlier pieces in this series, where their verification status is stated. Current as of July 19, 2026.

The Ladder

China’s robot trajectory data acquired a wage before it acquired a price, and the gap between those two events is where the entire embodied-data economy currently lives.

The wage is at the bottom of a ladder that can now be read rung by rung. A data collector, the person whose recorded movements become robot training data, earns 200 to 250 yuan for a nine-and-a-half-hour day, with night shifts reported up to 370, which the trade shorthands as roughly 30 yuan an hour. One rung up, data collected by workers wearing sensor rigs with no robot present is priced, by trade-press accounts, at a third to a half of real-machine data. Above that, data generated on actual robots through teleoperation sells for 500 to 1,000 yuan per hour, with brokers reselling the same hours to multiple buyers at multiples of their cost. And at the top, in January, the first embodied-intelligence dataset ever traded on a Chinese data exchange sold in Jiangsu on the day it listed, and neither the exchange, the seller, nor the state media that celebrated the milestone published what it sold for.

Every number on that ladder except the last is an estimate from trade reporting, not a filing, and this piece will say so at every rung, because the sourcing is itself the finding. Having a wage means the cost of manufacturing an hour of robot experience is now calculable. Having no published price means the asset that hour becomes still cannot be valued. Between the calculable cost and the asset that cannot be valued sits a spread of seventeen to thirty-three times, taking the trade’s 30-yuan shorthand against the 500-to-1,000-yuan selling band, and everything currently called the embodied-data industry, the collection centers, the brokers, the government-funded data plants this series has traced, is an occupant of that spread. This piece prices the bottom of the ladder, dissects the spread, and ends with the three documents that would finally publish the top.

A Day at 30 Yuan an Hour

Start with the wage, because it is the only rung anyone has stood on and described.

A reporter at the tech outlet Huxiu worked a shift as an embodied-data collector this summer and published the account in late June; the following details are that one person’s one day, reported firsthand and not independently verified, and they should be read as a specimen rather than a statistic. The interview was a group session of more than twenty applicants with near-universal pass rates. The screening that mattered was physical: height and weight checked, palm measured against a fixed glove size, a question about motion sickness in headsets. No degree, no experience. The job is advertised, frequently, under the title robot trainer. On site, phones are surrendered at the door. The work splits into two kinds: teleoperated collection, driving a dual-arm robot through tasks with handheld controllers while every motion is logged, and demonstration collection, wearing sensor equipment and performing tasks directly in staged real environments, a single scene occupying a team for two or three months. The shift ran nine in the morning to six thirty in the evening for a daily wage between 200 and 250 yuan, day-rate settlement available for part-timers, social insurance for full-timers, overtime routine.

A trade-press survey of the same labor market, also secondary and also unaudited, adds the distribution around that specimen: hourly rates near 30 yuan, a night-shift worker at 370 yuan a day, and attrition at one collection center reported at 70 percent.

Read the structure, not the anecdote. Group interviews, physical screening, day-rate settlement, night-shift rotations, high churn: this is the recruiting form of Chinese electronics manufacturing, transplanted whole onto the production of the scarcest input in the AI industry. The trade calls the facilities data factories, and the label is not a metaphor. The labor organization of trajectory-data production has already converged on the assembly line, complete with the assembly line’s wage, which runs a few hundred yuan a day, and the assembly line’s turnover. Whatever the embodied-data economy becomes, its manufacturing floor has been designed, and it was designed by copying the one China already had.

Divide the day rate by the shift and the raw number lands: 200 to 250 yuan across nine and a half hours prices collection labor at roughly 21 to 26 yuan per shift-hour before any adjustment for how much of a shift yields usable data. The trade’s 30-yuan shorthand sits just above that band, which is where a shorthand should sit. That is the wage of the robot-hour, and it is the first term of the cost construction this series is building.

The Spread

Between the 26-yuan shift-hour and the 500-to-1,000-yuan data-hour stands a stack of intermediaries, and the stack is where the money currently is.

The output price first, with its sourcing stated plainly. A STAR Market trade daily reported in April that real-machine embodied data trades domestically at 500 to 1,000 yuan per hour, with data collected by sensor-wearing humans and no robot expected to converge to a third to a half of that. A separate industry account puts standard packaged data at 300 to 500 yuan per hour and teleoperated real-machine data above 1,000, and describes a brokerage layer that resells identical hours to multiple model developers at more than ten times cost, with one independent data vendor reported to have booked 550 million yuan of new orders in the first quarter of 2026, its clients including the humanoid firms Zhiyuan and Galbot. None of these figures appears in any filing this publication has read. They are the going rates as the trade press hears them, which is precisely the condition the title of this piece describes: prices that circulate, and nowhere sit still on a public page.

What fills the spread between the wage and those rates is identifiable even where it is not priced. The robot itself, where one is used: this series established from Leju’s ChiNext filing that a humanoid sells to data collection centers at a recognized price of at least RMB 308,000 per unit, so depreciation on the instrument is real, though at any plausible utilization it amortizes to a small figure per hour against a four-digit selling price. The premises and staging. Quality inspection and annotation, a second layer of labor above the collector. And the brokerage margin, which on the reported resale economics is not a margin at all but the dominant term: selling one hour many times converts a labor cost into a licensing business, and licensing is where ten-times markups live.

One reconciliation belongs here, because this series has published a nearby ratio and the two must not be allowed to blur. Leju’s own dataset-construction budget, examined earlier in this series, allocates about seventeen yuan of wages and research expense for every yuan of equipment. That is a ratio of cost components inside one company’s internal build plan. The seventeen-to-thirty-three-times figure in this piece is a ratio of market output price to raw labor input across the whole chain. Same neighborhood of numbers, entirely different definitions, and the coincidence is exactly that. What the two figures jointly establish is one directional fact: at every altitude of this economy, from a filed corporate budget to a gray-market price sheet, the robot-data business is overwhelmingly a labor business wearing hardware as a uniform.

The Trade With No Number

Now the top of the ladder, where the market was supposed to become official.

On January 3, state and financial media reported the first embodied-intelligence dataset ever transacted on a Chinese data exchange: a listing by Jiangsu Zhujing Intelligence on the Jiangsu Province Data Exchange, sold, per the coverage, essentially on listing. The dataset’s contents were described in unusual detail. Roughly 25,000 structured clips across four scenarios, office, retail, dining, and housekeeping. Each clip about ten seconds, tens to hundreds of megabytes, containing the robot-view video stream and the full log of joint currents, angles, and torques, with the task instruction attached. The collection method, as the wire copy described it, was a human demonstrator whose motions were synced in real time to an adjacent robot, the machine’s own joints doing the recording.

Two computations follow from the disclosure, and the first is the arresting one. Twenty-five thousand clips of ten seconds is roughly 250,000 seconds, which is about 69 hours of data. The first exchange-traded embodied dataset in China, the transaction covered by the provincial government, the exchange, and the national financial wires as a zero-to-one breakthrough for the data-element market, Beijing’s policy project of turning data into a tradable factor of production, contained less than seventy robot-hours. At the trade-press rate band for real-machine data, the entire landmark would gross somewhere in the tens of thousands of yuan. The event’s significance was never the volume. It was the venue.

Which makes the second observation the structural one: the price was not published. Not by the exchange, whose institutional purpose is price discovery. Not by the seller. Not by any of the many outlets that carried the story. This publication checked the coverage end to end; the transaction has a date, a venue, a seller, a clip count, a file-size range, a sensor manifest, and no number. China’s data exchanges were built, explicitly, to move data pricing from bilateral gray quotes into public, referenceable markets. The first embodied transaction they hosted did the opposite: it demonstrated that the venue works while keeping the one datum the venue exists to produce. As a market event it discovered nothing. As an institutional event it is legible the same way this series has learned to read the data collection centers: the state side of this economy is currently in the business of demonstrating that a market will exist, which is not yet the business of operating one. The physical machine is being built ahead of its price signal, exactly as the fabs were, and the price signal, for now, lives where it has lived all along, in the trade’s unpublished quote sheets.

That completes the free layer: the wage established, the spread dissected, the official market shown to be a ceremony. The paid layer prices the labor term for the robot-hour construction this series is building toward, and names the documents that will force the number into the open.

Pricing the Labor Term

The robot-hour cost construction this publication announced needs three observable terms: what the equipment costs, what the labor costs, and what the compute costs. The equipment term was bounded earlier in this series from Leju’s filing. This piece bounds the labor term, and honesty about its shape matters more than false precision.

The floor is firm: 21 to 26 yuan per shift-hour of collector labor, from the reported day rates over the reported shift. What is not observable is yield, the fraction of a shift that becomes usable trajectory data after retakes, setup, quality rejection, and annotation. No source this publication has found states a yield figure, and this piece will not invent one. The honest statement is therefore a floor and a multiplier: the labor cost of one usable robot-hour is 21 to 26 yuan times a yield multiplier greater than one, plus the second-layer labor of inspection and annotation. For the construction ahead, the working consequence is that even at punishing yields, several shift-hours consumed per usable hour, the all-in labor term sits in the low hundreds of yuan, which is to say below the bottom of the 500-to-1,000-yuan selling band. The spread survives pessimistic yield assumptions. The current economics of the data factory are real, not an artifact of optimistic accounting.

The question an investor should hold is not whether the spread exists but which direction it closes from, and both ends are visibly in motion.

The wage end points up. Seventy percent attrition, if the reported figure is even directionally right, is the labor market’s verdict on 30 yuan an hour for work that surrenders your phone at the door, and the fix for 70 percent attrition is not culture, it is pay. The physical screening also caps the labor pool in a way electronics assembly never did: the glove has one size. A rising wage floor raises the cost of every hour in the country at once, including the hours the state’s data collection centers are budgeting with fiscal money, and it raises them before those centers reach full operation.

The price end points down. In mid-April, within two days of each other, the tactile-sensor firm Daimon released a dataset it projects at millions of hours within the year, ten thousand hours of it pledged open to the whole industry with the first batch already live on a public model community, and JD announced a full-chain embodied data infrastructure running from collection through its own trading platform, with a stated plan to mobilize up to 600,000 collectors for ten million hours of human-scene video within two years plus one million hours of robot-body data. Zhiyuan’s data subsidiary has launched a one-stop physical-AI data platform of its own. Projections are projections and are labeled as such here, and most of the JD volume is wearable-class video with no robot present, the tier the trade already discounts to a third or a half of real-machine rates. But the direction is unambiguous: open hours compete with sold hours at a price of zero, a retailer collecting inside its own warehouses and storefronts does not need the 500-yuan rate to cover a collector’s wage, and the million robot-body hours land in the exact tier the brokers currently price. The rate band is a scarcity rent, and the scarcity is being industrialized against.

So the labor term for the keystone is set, a floor in the twenties of yuan and a bounded multiplier, and the judgment on the spread is directional and holdable: it is a window rent, being closed from beneath by wages and from above by supply, and the speed of that closing determines what kind of market the state’s fiscally-built data centers open into when they reach capacity. If the spread closes before they do, the centers will have been built for a margin that no longer exists, and the fiscal price this series documented at Leju will stand as the high-water mark of the entire build-out. The compute term is now the only one missing, and it is the next piece in this sequence.

The Document That Publishes the Price

The number this whole piece circles, the public price of one hour of Chinese robot data, does not exist yet. Three document classes would create it, and they are listed in the order this publication expects them.

First, and most likely, a government procurement award. The data collection centers this series has traced are state-funded, and state money that buys data collection services at scale must eventually publish tender results with award amounts, and often unit pricing, because procurement disclosure rules require what commercial contracts do not. The first award notice that prices data collection by the hour or by the trajectory will be the first public price in this market, and it will arrive as a routine administrative document, not a press release.

Second, Leju’s inquiry response at the Shenzhen exchange, still unpublished as this piece goes out. The exchange has every reason to force decomposition of the data-center contracts that produced 44.94 percent of the company’s flagship revenue, and any decomposition that separates hardware from data-service performance obligations puts a filed, audited number adjacent to the trade’s rumored ones for the first time.

Third, the exchanges themselves, if a subsequent embodied listing publishes its price where the first did not. That would mark the moment the demonstration phase ends and the price-discovery phase begins, and it is the least likely of the three to come first, for the institutional reasons the Jiangsu trade just exhibited.

Whichever arrives first, the test it applies is the same. The trade says an hour is worth 500 to 1,000 yuan. The wage says an hour costs tens of yuan to make. The first public number will land somewhere against that spread, and where it lands will reveal whether China’s robot-data market is a real market clearing a scarce input, or a fiscal circuit renting a window. This publication is watching all three documents, and the robot-hour construction resumes the moment any one of them prints.

Inside China’s Machine is research, not investment advice. This piece contains no Confirmed market prices, and that is its subject. Collector wages, shift structure, and attrition are reported from a firsthand Huxiu account of one shift and an ITBear industry survey, secondary and not independently verified. The 500 to 1,000 yuan per hour rate band and the wearable-data discount are per a Kechuangban Daily report of April 17, 2026; brokerage resale economics and the 550 million yuan order figure are per the same ITBear survey. The Jiangsu exchange transaction details are per Nanjing municipal and financial wire coverage of January 3, 2026, which disclosed contents but no price. Leju’s equipment price and dataset budget composition are carried from this series’ earlier reading of Leju’s ChiNext filing, where their verification status is stated. Estimates of the industry’s accumulated data stock conflict by an order of magnitude across sources and are used here only at order-of-magnitude level. Supply-side platform announcements are per Daimon and JD releases of April 15 and 16, 2026, as carried by financial and trade media; their volume targets are Projected. Current as of July 18, 2026.

This publication has already established that roughly half of the RMB 9.31bn the three listed-queue humanoid companies are raising is earmarked for a brain none of them has deployed. Lejuu’s ChiNext prospectus, read in full this week, answers the question that left open. The trajectories that brain needs are not being bought at the going rate, and they are not being built on the company’s own balance sheet either. They are being built inside a network of joint ventures capitalised by local state investment vehicles, in which Lejuu holds a stake in every case and control in none. Manufacturing is moving in the opposite direction, out of the company. What is being assembled here is not a humanoid maker that also sells data. It is a data producer with a robot division attached, and the conversion is being financed by municipal capital rather than by the robots.

The dataset project is the smaller half of the brain budget. Of a RMB 2.6bn raise, the use-of-proceeds table puts RMB 939.67mn into an embodied-intelligence R&D centre and RMB 616.25mn into a high-quality large-scale dataset construction project. That is 36.14% and 23.70%, or 59.84% together, against 8.22% for the manufacturing base and 24.62% for working capital. The company is raising money principally to build a model and to feed it, and only marginally to make more robots.

Six ventures, eight months, and never a controlling stake. Lejuu’s associate company schedule lists six joint ventures entered between June 2025 and January 2026 whose stated business is operating humanoid data collection centres. Beijing Shuju, with the company at 37.76% against Shijingshan district development at 39.31%. Wujiang Zhixun, 49% against Wujiang Big Data and a Suzhou Bay industrial investor. Hefei Shuju, 39% against Hefei Guoxian Holdings at 41%. Qingdao Zhixun, 49% against Qingdao Hi-Tech Industrial Development at 47%. Shandong Zhixun, 49% against Shandong Hongmeng at 51%. Ningbo Juci, 49% against Cixi Municipal Industrial Investment.

The pattern is exact and it is not accidental. In all six the counterparty is local state capital, and in all six Lejuu sits below 50%. These are associates rather than subsidiaries, which places the collection capacity, and the capital expenditure that builds it, outside the consolidated accounts of a company that has lost money for three consecutive years.

One of them carries a floor for the state, and that is the tell. The schedule’s Wujiang entry records a written undertaking from Lejuu’s Suzhou subsidiary to the two state shareholders: if the venture has not turned a positive net profit by the end of the 2027 financial year, the subsidiary will cooperate in acquiring their stakes.

Read what that does. The local government’s money enters as equity and carries a contingent exit at the company’s expense. Equity with a floor under it is not equity in the sense the balance sheet implies. It is closer to a loan whose repayment is triggered by the venture failing to perform. The state is being invited to fund the trajectory pipeline on terms that limit its downside, and the residual risk sits with the listing candidate. Whether the other five carry comparable terms is not disclosed, which is itself the question.

Manufacturing goes the other way. The related-party section records that from March 2026 Lejuu outsources whole-machine assembly of its humanoids to Dongfang Yuanqi, a venture in which it holds 20% and its shareholder Dongfang Jinggong holds 75%. The risk factors cover the input side, where concentration is already extreme: a single supplier, Wuxi Quanzhibo, accounted for 17.20% of total procurement in 2025 and 82.82% of all joint-module purchases, and Lejuu holds 5.73% of it.

So the physical machine is being pushed outward at both ends, to a contract assembler above and a near-single module supplier below, while the data apparatus is being pulled inward and duplicated across six provinces. A company allocates scarce attention to the thing it believes is scarce. This one has told us which.

What the robot business is doing while this happens. Revenue reached RMB 258.19mn in 2025, with the full-size Kuavo line at RMB 177.78mn, 68.86% of the total, up roughly twelvefold. Underneath that, gross margin fell across the three reporting years from 50.45% to 44.30% to 40.78%, the loss attributable to shareholders widened each year to RMB 69.78mn, and operating cash flow was negative in all three, at RMB 27.52mn, RMB 29.41mn and RMB 28.25mn. Revenue is compounding while margin and cash move the wrong way, which is what selling more units into a price-competitive market looks like. The robot business is not going to fund the data build. That is why the data build has two other funders.

The control group. Deep Robotics earned RMB 337.49mn in 2025 and posted its first profit, RMB 28.68mn, with genuine industrial deployments in power inspection, emergency response and policing. Humanoids were 0.24% of revenue. It has no network of data collection ventures, and its own prospectus still earmarks RMB 1,169.26mn, 46.72% of its raise, for embodied algorithms and models. The company that best solved the problem of getting robots into paying industrial work is no closer to the trajectories, because inspection routes do not produce manipulation data. Solving deployment does not solve the input. That is why Lejuu’s answer is structural rather than commercial.

What this is and is not. It is not a scheme. The scarcity is real, the local governments genuinely want the industry, and moving capital-intensive collection capacity off the books of a loss-making company is a rational thing for its board to do. It also means the layer’s data supply is being built with fiscal money on terms the market cannot see, by a company whose own product margin is falling, and priced by an exchange that has not yet asked about any of it.

What to watch. The Shenzhen inquiry response, on three points. Whether the exchange forces disclosure of the Wujiang undertaking’s accounting treatment and quantification, and whether comparable terms exist in the other five ventures. Whether it asks for the unit economics of the RMB 616mn project, meaning cost per collected hour and throughput, which is the number this publication has been unable to source anywhere. And whether the collection ventures will purchase robots from Lejuu in 2026. On that last point the historical record is clean: related-party sales were RMB 737,200 in 2025, 0.29% of revenue, so nothing circular has happened yet, and most of these ventures were formed in the second half of 2025 or later. The question is entirely about the year now running, and the exchange is the only party positioned to ask it.

Every figure in this issue is drawn from the filed prospectuses of Lejuu (ChiNext, 19 May 2026) and Deep Robotics (STAR Market, 18 May 2026), both read in full this week. Anything sourced otherwise is marked in the text.

Inside China’s Machine. China is building the machine that builds physical intelligence. Silicon, models, robots, factories. We read it one layer at a time and turn each into capital judgment.

This is investment research, not investment advice.

The Allocation That Ignores the Income Statement

Of the robot makers now queued at China’s exchanges, Deep Robotics is the one whose business is built on robots holding jobs. Nearly 79 percent of its quadruped and wheeled-leg revenue over the past three years came from industry deployment: machines inspecting substations, entering fire scenes, walking police patrols, and crossing steel-mill floors. That revenue turned the company profitable in 2025, with RMB 28.7 million of net income on RMB 337.5 million of sales, and it did so on motion control, the layer of embodied intelligence the industry calls the cerebellum. Deep Robotics is the purest available proof that the cerebellum, on its own, is already a business.

Now read the company’s fundraising table. Of the RMB 2.50 billion it is asking the STAR Market for, RMB 1.17 billion, which is 46.72 percent, goes to a single project: embodied algorithm and model research. The brain. The same filing reports that the company’s humanoid line, the product family a brain would ultimately animate, generated RMB 823,000 of revenue in 2025. That is 0.24 percent of sales, and it is lower than the year before. The model earmark is one thousand four hundred and twenty times the size of the revenue line it points at.

One company making that bet would be a strategy. What makes this a piece is that they are all making it, in nearly the same ratio. Unitree, whose revenue mix is the opposite of Deep Robotics’ in almost every way, earmarked 48.13 percent of its own STAR raise for model development. Leju, whose largest customer scenario is government data collection centers, directed 59.84 percent of its ChiNext raise to its research center and dataset projects combined. Three companies, three incompatible business models, one capital allocation. The income statements diverge and the fundraising tables converge, and the convergence is the finding: this is not a company’s strategy, it is a layer’s configuration, and the rest of this piece is about what enforces it and what it costs.

What a Cerebellum Business Looks Like When It Works

Before pricing the bet on the brain, establish what the working business actually is, because it is the only one of its kind with audited numbers.

In the industry’s own anatomy, the brain is the slow system: multimodal understanding, task decomposition, long-horizon planning. The cerebellum is the fast one: turning intent into joint torques, keeping the machine upright on stairs and rubble, recovering from a slip in milliseconds. Brains are where the frontier research lives. Cerebellums are where the products live, and Deep Robotics has spent eight years on exactly that layer, through three control paradigms the prospectus lays out in sequence: virtual model control to get robots walking outside a lab, model predictive control fused with whole-body control to cross unstructured terrain, and reinforcement-learning locomotion from 2023 onward to make the crossing robust to mud, gravel, and impact.

The commercial result is a machine that holds a job a human is glad to give up. A substation inspection route covers stairs, cable trenches, and equipment gaps that wheeled robots cannot pass, in weather a human inspector should not be out in. A collapsed building or a smoke-filled corridor is a place fire commanders would rather send a machine first. The filing counts the deployments product line by product line: over 300 units sold into power inspection across more than 100 State Grid and Southern Grid substations, over 150 into fire and emergency response units nationwide, over 300 into police and border systems, roughly 300 into industrial inspection at steel, aluminum, and warehouse sites, and about a hundred more across water, rail, and construction infrastructure. Cumulative production has passed 5,500 units, with 2,908 quadrupeds and wheeled-leg machines sold in 2025 alone.

The economics moved with the deployment mix. Gross margin climbed from 33.48 percent in 2023 to 38.76 percent in 2024 to 52.83 percent in 2025, and the company crossed into profit with RMB 15.1 million of net income even after stripping non-recurring items. Set that against the comparison table in its own filing: Unitree’s cumulative quadruped revenue splits roughly 40 percent research and education, 31 percent consumer, and 29 percent industry application, while Deep Robotics runs 79 percent industry application. Unitree books far more revenue and profit in absolute terms. But its business is substantially selling robots to people who study robots or play with them. Deep Robotics’ business is selling robots to people who put them to work, and the recurring, spec-driven, tender-based nature of that demand is what a hardware investor means when they say a market has arrived.

That is the cerebellum thesis, complete and profitable: motion control plus ruggedization plus scenario integration, sold into jobs that exist today. Hold that picture, because the fundraising table is about to walk away from it.

The Brain on the Books

The brain, as it appears in Deep Robotics’ audited statements, is RMB 823,000 of revenue. The DR series humanoid line booked just over RMB 1.1 million in 2024, its first year, and RMB 823,000 in 2025, its second. The line went backward in the year the company’s overall revenue tripled. The second-generation DR02, an IP66-rated outdoor humanoid aimed at power line work and emergency response, launched in October 2025, and the filing describes the whole humanoid effort, accurately, as at the start of commercialization.

The honesty runs deeper than the revenue line, and this is the passage that makes the filing a primary source for the whole series. In its industry analysis, Deep Robotics states that the scarcity of high-quality scenario data is the core bottleneck constraining system optimization, and that unlike language models and autonomous driving, which accumulated massive data, the embodied intelligence industry has not yet formed a data feedback mechanism at application scale. It states separately that embodied foundation models remain unconverged across visual-language models, visual-language-action models, and world models, with no unified technical direction or industry consensus, and that these models are not yet applied at scale in the company’s own products.

Read those sentences as what they are: a legal document, signed by the issuer and its sponsor, asserting that the input the brain needs does not yet exist at scale and the architecture the brain will use has not been chosen. This publication has traced the other side of that hole, the state-funded data collection centers now being built to manufacture exactly the missing feedback, where local government capital is buying robots to produce training trajectories ahead of any market that pays for them. Deep Robotics’ filing is the demand-side admission that matches that supply-side construction. The company proposing to spend RMB 1.17 billion on models is telling you, in the same document, that the fuel for those models is the industry’s binding constraint.

So the humanoid brain, at filing, is a product line with negative momentum, an unconverged architecture, and an input the issuer itself calls scarce. Against that, RMB 1.17 billion. The gap between those two facts is not an oversight. It is the industry’s actual belief structure, written in capital.

The Consensus and Its Enforcer

Line the three fundraising tables up. Unitree: 48.13 percent of the raise to model development, on a business that mints profit from research-lab and consumer sales and books RMB 877 million of humanoid revenue. Deep Robotics: 46.72 percent to models, on a business that profits from industrial deployment and books RMB 823,000 of humanoid revenue. Leju: 59.84 percent to its research center and dataset build combined, on a business whose hardware demand comes largely from state data infrastructure. One seller of research platforms, one seller of working machines, one seller into government data centers. Their model allocations sit inside a fourteen-point band, and two of them within a point and a half.

When companies with opposite income statements file near-identical capital plans, the explanation is not in the companies. It is in the system routing the capital, and the filing shows that system operating on Deep Robotics directly.

In December 2025, the National Artificial Intelligence Industry Investment Fund, the state’s dedicated AI capital vehicle, took 14.4 million shares of Deep Robotics, 3.81 percent of the pre-IPO company. It did not simply buy stock. The filing discloses that the fund, the company, and the controlling shareholder signed a memorandum setting out strategic target commitments on technology, products, and business, with supplementary information rights and liability for breach. The state’s money arrived with a contract about direction. In the same round, China Telecom’s investment arm joined the register; the fifteenth five-year plan proposals had already named embodied intelligence a future industry, with the full plan confirming it in March; and the STAR Market stood ready to refinance the whole layer at scale. Every element of that machine rewards one configuration: keep the profitable body business as the qualifying credential, and lever the raise into the brain, because the brain is the layer the state’s industrial strategy has designated as the contested ground.

Say what this is and is not. It is not companies being forced to misallocate. A rational board, watching foundation-model progress and holding a mandate to survive architectural convergence, might well choose the same split unprompted; the bet is defensible on its own terms. But the uniformity is not organic. It is the signature of a financing system that has made one answer cheap and every other answer expensive, the same institutional machine this publication has watched build fabs ahead of demand and data centers ahead of data markets, now allocating the intelligence layer through the fundraising tables of hardware companies. The physical machine and the institutional machine, one moving because the other pushes.

The free layer ends here with the structure established: a profitable cerebellum business, a pre-revenue brain, a near-uniform allocation across incompatible companies, and a state enforcement mechanism visible in the shareholder register. What remains is the question an investor actually holds: when this company prices, which of the two businesses is the valuation buying, and what would make the answer checkable.

Which Business Is the Valuation Buying

Two price marks exist, and the distance between them is the brain premium made visible.

The first mark is private and recent. In December 2025, the C-round investors, the state AI fund among them, paid RMB 607.17 per unit of registered capital, which against the 8.235 million units then outstanding implies a valuation on the order of RMB 5.0 billion. The state fund’s own entry at RMB 13.89 per share on the post-conversion share count implies roughly RMB 5.3 billion. Call the private mark five billion, dated December 2025, both figures derived from disclosed entry prices rather than stated by the company.

The second mark is arithmetic the filing implies but does not state. The company will issue no fewer than 82.98 million new shares, no less than 18 percent of the post-issue total, and it is asking for RMB 2.50 billion. If the full raise is achieved at the minimum float, the implied post-money value is RMB 2.50 billion divided by 18 percent, which is roughly RMB 13.9 billion. Treat that as a ceiling estimate of what the ask implies, not a prediction of pricing; the actual number waits for the offering. But run the multiples on it, because they are the point. RMB 13.9 billion is about 41 times 2025 revenue and about 485 times 2025 net profit, for a company whose profit engine is a hardware deployment business.

No deployment hardware business carries 485 times earnings. The cerebellum business, valued as the excellent niche industrial it is, supports some fraction of that number; everything above the fraction is the market pricing the brain, the RMB 823,000 product line backed by the RMB 1.17 billion earmark. Which means the multiple, whatever it prices at, will be a measurement of exactly one belief: that the model bet converts before the money runs out, in an industry whose own filings say the data feedback loop does not exist at scale yet.

Follow the sensitivity in both directions. If the brain option expires worthless, the downside floor is the cerebellum business, real, growing, 53 percent gross margin, but small: a few hundred million yuan of revenue growing off a base of substations and fire brigades, with the RMB 1.17 billion of model spending converting to burn. If the option pays, Deep Robotics holds something none of its brain-first competitors have: a fleet already employed in the field, which is to say a proprietary data feedback loop of exactly the kind its filing says the industry lacks. The more than one thousand machines the filing counts at work in grids, fire scenes, and factories are, in that scenario, not the old business. They are the data plant. The cerebellum revenue and the brain bet stop being separate businesses and become one loop, and the company’s 79 percent deployment mix, the thing that today makes it look conservative next to Unitree, becomes the scarcest asset on the cap table.

That is the honest shape of the trade: a profitable floor, a binary premium, and a conversion condition, data at scale from deployed machines, that this publication’s robot-hour work is attempting to price from the outside. The one thing the multiple cannot claim to be is a valuation of the business that currently makes the money.

The Two Inquiries

The thesis is checkable, and the checking mechanism is already scheduled twice.

Deep Robotics’ application was accepted by the Shanghai exchange on May 18. Unitree’s is in the same queue. Two inquiry processes, run by the same exchange on the same layer in the same season, will force the disaggregation this piece is built on, and they will do it in public documents.

Watch the Deep Robotics response for three things. First, the milestone structure of the RMB 1.17 billion model project: what the exchange accepts as success criteria for spending 46.72 percent of a raise on a product line with RMB 823,000 of revenue, and what timeline the company commits to for the humanoid commercialization it currently describes only as begun. Second, the data question: whether the exchange forces the company to reconcile its own statement that no at-scale data feedback mechanism exists with a project plan whose returns depend on one, and whether the deployed fleet is formally claimed as the data source. Third, the consistency test across the queue: whether Unitree’s inquiry produces materially different milestone discipline for a near-identical allocation, which would reveal whether the exchange is pricing the configuration or merely processing it.

The outcomes map. If the responses force hard milestones, dated humanoid revenue commitments, and explicit data-sourcing plans, the configuration consensus acquires a governance structure, and the brain premium becomes a monitorable bet rather than a mood. If the responses accept the allocation with soft language, the STAR Market is confirmed as the enforcement layer of the consensus rather than a check on it, and the premium is policy, not analysis. The middle case, hard questions and soft answers, is the likeliest, and it would leave the burden of verification exactly where it currently sits: on the fleet, in the field, in whether working robots can be turned into the data their own successors require.

Step back to the machine. At the bottom of the stack, the state financed fabs ahead of demand. In the middle, it is financing data plants ahead of a data market. And at the body layer, it has now arranged the capital markets so that even the one company profiting from robots that work will spend nearly half its raise on the part that does not exist yet. Every layer of China’s machine is being configured toward the same missing organ. The organ is the brain, the constraint is the data, and the companies themselves have started saying so in their filings. The judgment of this piece is narrow and holdable: Deep Robotics is the best-collateralized version of the layer’s one bet, and the collateral is the only part of the company the IPO is not really pricing.

Inside China’s Machine is research, not investment advice. Confirmed figures are drawn from Deep Robotics’ STAR Market prospectus, declaration draft accepted May 18, 2026, including revenue, margins, product-line and scenario splits, deployment counts, the fundraising allocation, shareholder entries, and the quoted industry-analysis statements, and from the comparison table that filing presents for peer companies. Unitree’s 48.13% model earmark is carried from this publication’s prior reading of Unitree’s prospectus and was not re-verified against that filing this session. The RMB 5.0 billion private mark and RMB 13.9 billion implied post-money value are derived arithmetic from disclosed entry prices and the minimum-float raise, not company-stated valuations; actual IPO pricing is undetermined. Statements about inquiry outcomes are Projected. Current as of July 18, 2026.

The Price Nobody Paid in a Market

Leju’s prospectus puts robot data collection into a revenue table, and the number it puts there is not the price of data. It is the price the state is paying to build the capacity to produce data. The two are different things, and the difference is the whole document.

The surface reads like a demand story. Leju Intelligence, the Shenzhen humanoid maker that filed for ChiNext in May as the first company under the board’s new fourth listing standard, sold 577 units of its full-size Kuavo robot in 2025, up from 32 the year before. Kuavo revenue rose roughly twelvefold to RMB 177.8 million and carried the company across the RMB 200 million revenue threshold the new standard requires. Third-party trackers put Leju’s 2025 humanoid shipments third or fourth globally depending on the count. The company’s last private round closed at a post-money valuation of RMB 4.33 billion, and it is now asking public markets for RMB 2.6 billion.

Read the revenue table by customer instead of by product and the story changes shape. The single largest destination for Kuavo robots in 2025 was not factories and not universities. It was data collection centers, government-backed facilities that buy humanoid robots in order to record what the robots do, so that the recordings can train the models that will one day run the robots. By the company’s own scenario breakdown, data collection accounted for 44.94% of Kuavo product revenue, ahead of research and education at 32.50%.

Who runs these centers is the question the rest of this piece answers, because the answer is: to a substantial degree, Leju does, in joint ventures with local state capital, some of which also sits on Leju’s own shareholder register. The buyer, the shareholder, and the joint-venture partner are in more than one case the same entity.

That is not a scandal. It is a design, and it is a design China has run before at the bottom of this publication’s stack, when local governments built fab capacity ahead of any market that could pay for it. What it is not, is a market price for robot data. Anyone using Leju’s 2025 revenue line to size the embodied-AI data economy is measuring fiscal willingness to build infrastructure, not commercial willingness to buy data. The distinction decides what a RMB 4.33 billion valuation is actually resting on, and that is where this piece ends up.

What a Data Collection Center Actually Is

Start with why these facilities exist at all, because the mechanism is the part most coverage skips.

A large language model trains on text that already exists. The internet is its warehouse, scraped at near-zero marginal cost. A robot control model has no such warehouse. What it needs is trajectory data: synchronized streams of what the robot saw, what command it received, what its joints did, and what happened in the physical world as a result. This data does not exist until a physical robot performs a physical task somewhere, with sensors recording. Simulation can stretch it but not replace it, because policies trained purely in simulation degrade when they meet real friction, real lighting, real clutter. The industry calls this the sim-to-real gap, and it is the reason embodied AI cannot repeat the LLM playbook. Robots cannot read the internet. Their training data has to be manufactured, one demonstration at a time, on real hardware.

A data collection center is a factory for exactly that. Take a hall, fill it with staged task environments, put in fleets of humanoid robots and human teleoperators, and run demonstrations all day. Leju’s own fundraising documents describe the business model of these centers in two lines: selling data products, meaning standardized training datasets, and renting the training facilities, meaning opening the site and the robot fleet to model developers who want to collect their own custom data.

Notice what the robot is in this arrangement. It is not labor and it is not a product feature. It is capital equipment on a data production line, the way a lithography tool is capital equipment on a wafer line. And that single fact is what connects the middle of China’s machine to the bottom of it. When a government entity buys a fleet of humanoids for a data center, it is making the same category of bet Hefei made when it capitalized a DRAM fab: build the production capacity first, at state expense, and let the market for the output arrive later.

The output here is trajectory data, the scarcest input in the embodied-AI stack. Which means the purchase price of the robots is, economically, the first observable number anywhere in China’s disclosure record that attaches money to the production of that data. It is a capacity price, not an output price. Holding that distinction is the discipline the rest of the analysis needs.

The Buyer, the Shareholder, and the JV Partner Are the Same Entity

Now trace who is writing the checks, because the prospectus lets you do it name by name.

Between June 2025 and January 2026, Leju and its subsidiaries co-founded six data collection center operating companies with local state capital, in six cities. Beijing Shuju was registered on June 17, 2025, with Leju’s Shanghai data subsidiary holding 37.76% and Beijing Shijingshan Industrial Development, a district-government vehicle, holding 39.31%. Wujiang Zhixun followed one day later in Suzhou, 49% Leju, 51% split between two district state-owned investors. Then Hefei Data in November, 39% Leju beside Hefei state capital. Qingdao Zhixun in November, 49% Leju, 47% Qingdao Gaoke Industrial Development. Shandong Zhixun in January 2026, 49% Leju beside a provincial state group. Ningbo Juci in January 2026, 49% Leju beside Cixi municipal industrial capital. Six ventures, seven months, every one of them minority-held by Leju and majority-funded by local government money, every one of them registered to operate humanoid data collection centers.

Hold that list against two other lists in the same document.

The first is the shareholder register. Beijing Shijingshan Industrial Development, the district vehicle that owns 39.31% of the Beijing data center venture, also bought into Leju itself in October 2025, at RMB 72.11 per share, and holds 0.58% of the company at filing. Its sister fund under the same district state-asset umbrella holds another 0.69%.

The second is the customer table. Leju’s five largest customers of 2025 include Beijing Shijingshan Industrial Development and Qingdao Gaoke Industrial Development. The same Shijingshan entity, in the same fiscal year, was Leju’s joint-venture partner in a data center, an equity investor in Leju, and one of Leju’s largest robot buyers. Qingdao Gaoke was a JV partner and a top-five customer. The largest revenue scenario in the company’s breakout, the 44.94%, is substantially a market that Leju co-founded, with counterparties who also sit above it on the cap table.

One more clause in the filing shows how both sides priced this arrangement. In the Wujiang venture, Leju’s subsidiary gave the two state shareholders a written undertaking: if the data center has not turned a net profit by fiscal 2027, Leju will take the state shareholders’ equity off their hands. The government partners, in other words, did not underwrite the data business’s commercial risk. They bought robots, seeded the venture, and secured an exit if the market for the output fails to arrive on schedule. That is not how a customer behaves. It is how a landlord providing policy capital behaves, and it is the single most honest price signal in the entire structure: the state’s own downside case is that trajectory data may not find a paying market by 2027.

Say plainly what this is and what it is not. It is not fabricated revenue. The robots are real, the deliveries are audited, and the exchange will interrogate every related-party linkage in the inquiry process. It is the Hefei playbook applied one layer up the stack: local governments building data production capacity ahead of demand, exactly as they built wafer capacity ahead of demand, accepting years of losses as the cost of owning a strategic input. The physical machine here only moves because the institutional machine is pushing it. But a demand curve made of policy capital tells you what the state will spend, not what the data is worth. Those readings diverge, and the divergence is measurable in Leju’s own accounts, which is where the next section goes.

The Replacement Right: Why Even the Fiscal Price Is Understated

The average selling price of a Kuavo robot in 2025 works out to RMB 308,000 per unit, from RMB 177.8 million of recognized revenue across 577 units sold. Against 2024’s average near RMB 414,000, that is a fall of roughly a quarter, and most coverage filed it as price war. The prospectus itself says the decline has two components, and the second one matters more than the first.

The first component is ordinary: an aggressive market-pricing strategy to win share, the same discount curve every Chinese hardware category rides. The second is an accounting mechanism specific to the data center contracts. In Leju’s own explanation of why its margins sit below Unitree’s, the company discloses that a high share of 2025 revenue came from data collection centers, and that some of those sales contracts grant the buyer a right to one future product replacement. Under revenue rules that right is a separate performance obligation. The contract price gets split, the portion attached to the replacement right is deferred, and 2025 recognizes only the delivered portion.

Follow the consequence through the statements. The RMB 308,000 recognized per unit is a net figure after carving out an undelivered obligation, so the full contract price per robot sits above it by an amount the filing does not, in the sections available, quantify. The same mechanism leans on the margin line: Leju’s blended gross margin fell from 50.45% in 2023 to 44.30% in 2024 to 40.78% in 2025, and the company attributes part of the final leg to recognizing partial revenue against costs already incurred. The reported 2025 margin is structurally depressed, and it will partially reflate in whichever period the replacement obligations settle.

Two readings follow, and both are worth carrying out of the free layer.

For anyone benchmarking robot prices, the first observable fiscal price of data production capacity in China is at least RMB 308,000 per humanoid unit, with the true contracted figure above that line pending the deferred portion. Every cost model of a robot data operation, including the robot-hour construction this publication is building toward, inherits that floor.

For anyone reading Leju’s income statement, the data center contracts cut in both directions at once. They are the growth engine, the margin drag, and a stored release of future revenue, all in the same line item. The buyers embedded an option to swap early hardware for better hardware, which is rational behavior for an infrastructure owner expecting the equipment to be obsoleted by its own output: the data these robots produce trains the models that will make this generation of robots worth replacing. The contracts have the obsolescence loop priced in. The income statement is carrying a liability that is, mechanically, a bet on the machine improving.

That is the free layer complete: what a data collection center is, who actually funds it, and why even the fiscal price on the page understates itself. What remains is the part that moves a valuation. The paid layer prices the fiscal price, triangulates it against Leju’s own RMB 616 million internal cost of building a dataset, and names the disclosure, now pending at the Shenzhen exchange, that will settle whether this structure is a bridge to a data market or a substitute for one.

Pricing the Fiscal Price

Three numbers in the filing let you triangulate what the state is actually paying, and the third one is the surprise.

The first is the equipment price. At least RMB 308,000 per humanoid unit on a recognized basis, with the full contract price above that pending the deferred replacement obligation. This is what a data collection center pays for one unit of data production capacity.

The second is the demand quantum per city. Beijing Shijingshan Industrial Development alone bought RMB 33.4 million of product from Leju in 2025, which is 12.94% of the company’s entire revenue from a single district-government counterparty that also holds its equity and co-owns its Beijing data venture. Scale that across the scenario table and data collection took 44.94% of the relevant revenue base in 2025. Either way, the order of magnitude of one year of fiscal appetite, per participating city, is tens of millions of yuan, and six cities signed up within seven months.

The third number is what it costs to turn that capacity into the actual product, and Leju priced it itself. The RMB 616 million dataset construction project in the fundraising plan breaks down, in the company’s own feasibility arithmetic, as follows: RMB 32.4 million for premises, RMB 31.1 million for equipment purchase and installation, RMB 16.0 million for software, RMB 7.9 million of contingency, RMB 282.2 million for R&D personnel wages, and RMB 246.6 million of other R&D expense, over a 48-month build.

Read that decomposition slowly, because it inverts the picture the hardware revenue paints. In Leju’s own accounting of what a dataset costs to manufacture, equipment is 5.05% of the bill. Wages and R&D expense are 85.8%. The robot, the thing the data collection centers are spending tens of millions per city to acquire, is a rounding error in the cost structure of the output it exists to produce. Data, at 2026 prices, is made of people: teleoperators, annotators, quality reviewers, the human demonstration labor that no amount of capital equipment removes from the loop yet.

Two capital consequences follow.

The first is about measurement. The fiscal price this piece has been tracing, the robot purchase orders flowing from state-backed centers, captures the smallest slice of what trajectory data actually costs. Anyone sizing China’s embodied-data economy off hardware procurement is measuring the lithography tool and ignoring the fab’s payroll. The robot-hour cost this publication is building toward will be dominated by the labor term, and Leju’s own budget just told you the ratio: on the order of seventeen yuan of human cost for every yuan of equipment cost. That is the number the teleoperation piece in this sequence will have to survive contact with.

The second is about the valuation. Leju’s last private round implies roughly 16.8 times 2025 revenue, and the fourth-standard listing rests on a projected market value above RMB 3 billion. The load-bearing assumption under both is that the 44.94% scenario either persists or converts into something a private customer pays for. The persistence case is policy-cyclical: six ventures in seven months is the signature of a policy window, opened by the August 2025 State Council AI-plus opinion and the Fifteenth Five-Year Plan’s embodied-intelligence training-ground language, and policy windows that open on that schedule can close on one. The conversion case runs through the data centers’ own income statements, which do not consolidate into Leju’s and which the Wujiang undertaking tells you the state partners themselves would not underwrite past 2027. The revenue line is real. The question the multiple is silently answering is whether it is a market or a program, and the filing, on its own, cannot answer it.

Set this beside the series. This publication’s read of China’s three humanoid IPOs found that 41 to 51 percent of their combined RMB 9.31 billion raise was earmarked for model development rather than hardware. Leju pushes the line further: 59.84% of its RMB 2.6 billion goes to the research center and the dataset project combined. The hardware companies are becoming the financing vehicles for the data and model layer, and Leju is the purest case yet, because its hardware revenue itself already comes, in its largest part, from the data layer’s construction budget. The machine is eating its own capital structure from the middle out.

The Document That Will Settle It

The thesis of this piece is falsifiable, and the document that will test it is already in process at the Shenzhen exchange.

Leju’s application moved to inquiry status on May 26. The response, not yet published as this piece goes out, will be the highest-yield document in this story, for a reason internal to the filing itself. The prospectus reports related-party sales of RMB 737,200 in 2025, which is 0.29% of revenue. The same prospectus reports RMB 33.4 million of 2025 sales to Beijing Shijingshan Industrial Development, an entity that holds Leju equity and co-owns Leju’s Beijing data collection venture. Both numbers are audited and both are, under current related-party definitions, simultaneously correct: a 0.58% shareholder that partners with your subsidiary in a joint venture you do not control sits outside the disclosure perimeter of related-party transactions as the rules define them. The buyer, shareholder, and JV partner triangle is legal, audited, and invisible to the one table investors check for exactly this pattern. The inquiry process exists to test perimeters like that one, and a fourth-standard first case will be tested hard.

Watch for three things in the response. First, the customer-by-customer decomposition of data collection revenue, and whether the exchange forces a related-party or quasi-related-party characterization of the shareholder-customers. Second, the replacement-right accounting: the deferred quantum, the expected settlement window, and therefore the size and timing of the margin reflation currently stored in the contract liabilities. Third, sustainability questioning: whether the 2025 data center orders carry into 2026 order books, which the exchange will ask because the fourth standard’s growth requirement makes it existential.

The outcomes map cleanly. If the response shows the co-founded centers already booking meaningful data product revenue from third-party model developers, the fiscal price is functioning as a bridge, a market price is forming downstream, and the bearish reading of the structure weakens. If it shows robot revenue concentrated in co-founded centers with no downstream data sales to speak of, the structure is, for now, a program wearing a market’s clothes, and the 16.8 times multiple is pricing a fiscal appetite with a 2027 sunset written into its own JV agreements. The middle case, partial third-party revenue with heavy concentration, is the likeliest and the hardest, and it would make the 2027 Wujiang test date the single most informative deadline in China’s embodied-data economy.

Step back to the whole machine. At the bottom of the stack, China built wafer capacity ahead of demand with fiscal capital, and a decade later CXMT is pricing an IPO. In the middle, it is now building data capacity ahead of demand with the same playbook, the same district vehicles, the same patient losses. The substrate strategy took ten years to produce a market price. The data strategy has given itself two. The first observable number is on file in Shenzhen, and it is a fiscal one. The market one does not exist yet, and this publication will be watching the exact documents where it would first appear.

Inside China’s Machine is research, not investment advice. Confirmed figures are drawn from Leju Intelligence’s ChiNext prospectus, declaration draft filed May 19, 2026, including shareholdings, joint-venture registrations and stakes, revenue, margins, the related-party sales line, and the fundraising allocation. The scenario revenue split, the Shijingshan customer figure, and the dataset project cost lines are reported from coverage citing the prospectus and reconcile arithmetically to the filed totals, but were not checked against the filing’s own tables this session. The Shenzhen exchange’s inquiry response was unpublished as of this writing. Statements about the inquiry outcome and the 2027 Wujiang profitability test are Projected. Current as of July 17, 2026.

1. Zhipu’s Flagship Runs on DeepSeek’s Attention

Open the GLM-5 technical report and read the first technical contribution the authors list. It says they adopt DSA. Next to DSA is a citation number, and the citation points to DeepSeek.

One of China’s strongest open-weight models runs its core efficiency mechanism on something its largest rival open-sourced three and a half months earlier. DeepSeek released DeepSeek Sparse Attention on 29 September 2025, as the only architectural change in its V3.2 model. Zhipu released GLM-5 on 11 February 2026, and its own paper describes the model as adopting that same mechanism, by name, with the reference attached. The GitHub repository says it plainly: GLM-5 integrates DeepSeek Sparse Attention.

This is not plagiarism. It is an open-source ecosystem working exactly as designed, and DeepSeek published DSA under a permissive licence precisely so that others would use it. But it raises a question that the benchmark tables do not answer. If the efficiency advantage under your flagship model is one your competitor handed you for free, in what sense is it your advantage?

That question is the whole piece, and by the end it will have moved the moat from the model layer down to the silicon.

2. What DSA Is, and Why It Was Worth Copying

Attention is the operation that makes a language model expensive to run. For every new token, the model compares that token against every token already in the context, so a sequence of length L costs on the order of L squared. Double the context and you quadruple the work. On a long document this is where the compute and the memory traffic go.

DSA breaks the quadratic. It adds a small, fast component that DeepSeek calls a lightning indexer, which runs ahead of the expensive attention step and scores which earlier tokens actually matter for the current one. The model then computes full attention over only the top few thousand selected tokens rather than the entire context. The cost falls from quadratic in the sequence length to roughly linear. DeepSeek’s report gives the mechanism as an indexer plus fine-grained token selection, taking attention from order L squared to order L times k, where k is the small number of tokens kept.

Here is why it matters for everything this publication has argued about silicon. The tokens a model attends to are the tokens it must stream out of memory during decode, the answer-writing phase that the ninth issue established is bound by memory bandwidth rather than compute. Cut the number of tokens attended to, and you cut the data moved per output token. GLM-5’s report puts the saving at 1.5 to 2 times on long sequences. That is a memory problem being solved in software.

And it is the same memory bottleneck the ninth issue traced to Huawei’s roadmap. Huawei’s fix is hardware: the 950DT, with wider memory, arriving in Q4. DSA is the software fix, and it is available now. It does not wait for a fab. It moves less data per token, on the silicon you already have.

3. It Took Four and a Half Months to Become Everyone’s

DeepSeek published DSA on 29 September 2025, with a paper, open weights, and an MIT licence. Zhipu shipped it inside GLM-5 on 11 February 2026. Four and a half months from one company’s research result to the efficiency core of a rival’s flagship.

And the adoption is explicit rather than reinvented. GLM-5’s technical report does not describe an independently discovered mechanism that happens to resemble DSA. It names DeepSeek Sparse Attention and cites the DeepSeek paper, the way you cite a dependency you have imported. Zhipu’s subsequent releases keep it: GLM-5.1 carries DSA forward, and GLM-5.2 adds a refinement on top of it called IndexShare that reuses the indexer across layers, which is an optimisation of DSA, not a replacement for it. MiniMax and Qwen have moved along the same sparse-attention path.

So within one release cycle, a single company’s architectural advance became the shared default of the Chinese open-weight field. This is what open source does, and it does it fast.

A thing that becomes everyone’s in four and a half months is not a moat. It is public infrastructure. The company that invented it gets a few months of lead and a citation. It does not get a durable cost advantage, because the mechanism that produced the advantage is now running inside every competitor that wanted it.

4. The Precision Is Not Free

Sparse attention has a cost, and the balance of this piece depends on naming it.

When a model keeps only the top few thousand tokens and discards the rest of the context from each attention step, it is making a bet that the discarded tokens did not matter. Usually that bet is fine. Sometimes it is not, and the model loses information it would have used. DeepSeek’s own ablations on V3.2 show a mixed picture against the dense baseline: performance holds on most tasks, improves on some, and regresses on a few. Zhipu reports a related difficulty from the memory-compression side, that its latent-attention variant initially underperformed standard grouped-query attention until an optimisation it calls Muon Split closed the gap.

The point is not that DSA is flawed. It is widely adopted because it works. The point is that architectural efficiency runs into a floor. You cannot compress the KV cache to nothing, and you cannot drop attention to no tokens at all without the model getting worse. There is a limit to how much data movement software can remove, and once a design is near that limit, the only way to serve tokens more cheaply is better silicon underneath.

5. If Everyone Has the Architecture, Where Does the Margin Come From

Follow the logic to its end. If DeepSeek, Zhipu, MiniMax and Qwen are all running the same sparse-attention mechanism, then at the architecture level their per-token cost structures are converging. The efficiency that DSA delivers is available to all of them, on the same terms.

A cost advantage cannot come from a thing everyone has. So if these companies earn different margins on inference, and the eighth issue showed they earn very different margins, the difference has to come from somewhere the architecture is not.

There are two places left, and both are silicon.

The first is how deeply a company’s model is fitted to its chips. The same DSA mechanism does not yield the same throughput on every accelerator. It has to be compiled, quantised and scheduled onto the specific hardware, and the quality of that model-silicon integration decides how many tokens per second actually come out. Running a model at low precision without losing accuracy, fusing the attention kernels to the memory system, keeping the accelerator fed: this work is worth real points of gross margin, and it does not transfer with the open-source weights.

The second is which silicon a company can get at all. Better memory bandwidth, a newer process, more of it. That is the ceiling the sixth and ninth issues described, and it is set in fabs and export-control offices, not in model architecture.

Architecture pulled everyone up to the same line. Then the gap reopened on silicon. The moat did not disappear when DSA became universal. It moved down a layer, from the model to the chip, where it is harder to copy because you cannot fork a fab.

6. Re-diagnosing Zhipu’s 18.9 Percent

The eighth issue of this publication found Zhipu earning an 18.9 percent gross margin on its cloud and API line, the figure it reported as roughly 19 percent, against far higher margins elsewhere, and could not fully explain the number from the income statement alone. The ninth issue traced part of it to silicon: Zhipu runs on domestic accelerators whose memory bandwidth caps its decode throughput, and the chip that would lift that cap does not ship until Q4. That issue left one thread deliberately loose. It said the ceiling is set by the silicon, but the distance to the ceiling is set by engineering, and that distance is the part Zhipu still owns.

This piece closes that thread, and the answer is not the one the loose end implied.

The tempting story was that Zhipu earns less because it did less of the architectural work, that DeepSeek compressed its memory traffic and Zhipu did not. That story is false. Zhipu runs DSA. It adopted DeepSeek’s mechanism in GLM-5 and carried it through every release since. On sparse attention, the thing that determines how much data decode moves per token, Zhipu and DeepSeek are running the same design. The architectural distance between them, on what matters most for inference cost, is close to zero.

So the 18.9 percent is not a verdict on Zhipu’s architecture, because its architecture is DeepSeek’s architecture, and DeepSeek’s is available to everyone. Once the model layer converges, the margin stops being a statement about the model. It becomes an almost pure reading of something underneath it: how well Zhipu’s identical-on-paper model is fitted to the domestic silicon it is required to run on, and how good that silicon is.

That is the sense in which the distance Zhipu owns is real but small, and shrinking. It does not own an architectural edge, because there is no architectural edge left to own. What it owns is the integration work between a model everyone has and a chip it did not choose. That is a narrower thing to compete on than a better model, and it is silicon work, not model work.

7. The Strongest Objection: Does Open Diffusion Erase the Silicon Gap

Here is the case against this entire thesis, stated at full strength, because it is a good one.

If DSA lets every model move half as much data per token, then it helps the memory-constrained domestic chips most. A company running GLM-5 on a 1.6-terabyte-per-second Ascend 950PR, cutting its memory traffic in half with sparse attention, might reach the throughput that a competitor gets from twice the raw bandwidth without sparse attention. On this reading, open-source architecture is not a sideshow to the silicon story. It is precisely how China routes around the export controls: if you cannot buy the memory bandwidth, you open-source the software that halves your need for it. Diffusion is not eroding the moat. It is China’s answer to not having one.

That objection is partly right, and it matters. DSA genuinely does relieve the domestic bandwidth constraint, and the fact that it spread across the Chinese field in a single release cycle is a real strategic asset that a closed ecosystem would not have.

But the relief is not exclusive, and that is where the objection fails. Attention is symmetric. DeepSeek open-sourced DSA to the entire world, not to China alone. The same mechanism that halves memory traffic on an Ascend 950PR halves it on a B300, and the B300’s memory bandwidth was about five times higher to begin with, so the tokens-per-second a frontier operator gains from adding DSA is larger, not smaller. Software that helps everyone helps the operator with the better silicon at least as much. DSA lifted the whole field, including the part of the field that was already ahead. It shortened China’s distance to the ceiling. It did not raise China’s ceiling, because the ceiling is bandwidth, the bandwidth is silicon, and the fast silicon still ships in Q4 to one side and is already shipping to the other.

8. The Moat Went Down a Layer

Step back to the machine, because this is where the three issues close.

The sixth issue argued that China’s compute ceiling is a memory ceiling, not a logic one. The eighth found that ceiling resting on a model company’s gross margin, on Zhipu’s 18.9 percent. The ninth showed Huawei splitting a chip in two along the memory seam, with the half that lifts the ceiling scheduled for Q4. This issue adds the last turn: software can help you reach the ceiling faster, but it cannot raise it, and the software that helps is shared, so it does not separate you from anyone.

The forward marker is the diffusion clock itself. Watch the next architectural advance after DSA, whatever compresses memory traffic further, and time how long it takes to appear across the open-weight field. If the answer is again a few months, then architecture as a non-moat is not an observation about one mechanism, it is a standing property of an open-source ecosystem, and every efficiency story in Chinese AI should be read as temporary by default. If some lab instead holds a real architectural advance closed for a year and no one can reproduce it, this thesis needs revising, and I will say so.

A model company’s margin is not set inside its model, and now we can say why with the last piece in place. It is not set there because the model is not where the difference is. Everyone downloaded the same attention. What is left to compete on is the fit between a shared model and an unshared chip, and the quality of the chip, and both of those are silicon. DeepSeek open-sourced the ladder to the ceiling, so everyone has the ladder. The ladder is not the moat. The height of the ceiling is, and it is written on a fab’s roadmap, not in a model’s weights.

Inside China’s Machine is research, not investment advice. Nothing here is a recommendation to buy or sell any security.

1. Huawei Built Two Chips Where There Used to Be One

In 2026 Huawei’s flagship AI accelerator is not one chip. It is two, sharing the same core compute architecture, separated by nothing but the memory packaged alongside it.

The Ascend 950PR carries 128 gigabytes of Huawei’s own high-bandwidth memory at 1.6 terabytes per second. The Ascend 950DT carries 144 gigabytes at 4 terabytes per second. Same DaVinci cores, same interconnect, same instruction set. The difference the company chose to build a second product line around is memory bandwidth, and a factor of two and a half separates the two.

A chipmaker does not split its flagship in two without a reason that costs it money to act on. The reason is printed in Huawei’s own architecture whitepaper, and it is worth reading in the company’s own words.

This is a story about silicon. It does not stay there. The seam Huawei cut through its own product line is the same seam that runs through the income statement of every Chinese company selling inference, and by the end of this piece the 1.6 and the 4 will be numbers about gross margin.

2. Why Inference Splits Into Two Halves That Want Opposite Things

Serving a large model happens in two phases, and they place opposite demands on the hardware.

The first phase is prefill. The model reads your prompt, the whole thing at once, and builds its internal representation of it. This is a wide, parallel, compute-heavy operation. Many numbers multiplied at the same time. It leans on arithmetic throughput and it is comparatively relaxed about memory.

The second phase is decode. The model writes its answer one token at a time, and each new token has to see every token before it. At each step the accelerator streams the model’s weights and the entire accumulated context out of memory and into the arithmetic units, does a comparatively small amount of maths, and writes one token back. Then it does the whole thing again for the next token. The arithmetic units are mostly idle. What they are waiting for is memory.

Prefill is compute-bound. Decode is memory-bound. This is not an opinion, and it is not this publication’s framing imported onto Huawei’s hardware. It is Huawei’s framing.

The whitepaper states, in its opening, that agent workloads with long context and multi-turn interaction drive KV Cache storage up exponentially, and that on-chip memory alone can no longer support the business. The KV Cache is the model’s working memory of the conversation so far, the stored representation of every token already seen, and it is what decode has to stream through memory at each step. It grows with context length, which is why longer prompts and longer sessions push the memory demand up rather than the compute demand. Huawei is telling you, in its own document, that the binding constraint on modern inference is memory.

3. The Names Are the Argument

Read what Huawei called the two chips.

The 950PR stands for Prefill and Recommendation. The 950DT stands for Decode and Training. The company did not split its chip by customer, or by price tier, or by market segment. It split it by which half of inference the chip is for, and it put the high-bandwidth memory on the decode side.

A product line is the most honest thing a company publishes, because it is the one claim it has spent capital to stand behind. Marketing can say anything. A second tapeout and a second memory supply chain cannot. When Huawei put 4 terabytes per second behind decode and 1.6 behind prefill, it was pricing the exact proposition this publication argued from the silicon side in its sixth issue and from the income-statement side in its eighth: that the cost of serving a token is set by memory bandwidth, because decode is where the tokens are made and decode is memory-bound.

You do not have to take my word that decode is the expensive half. You can take Huawei’s capital allocation, which says the same thing and cost more to say.

4. Self-Sufficient and Two Generations Back, at the Same Time

Here is where the balance has to hold, because the story breaks if you drop either half.

The high-bandwidth memory in both chips is Huawei’s own. HiBL 1.0 on the 950PR, HiZQ 2.0 on the 950DT. The whitepaper calls it high-speed on-chip memory and describes it as the DRAM that holds the model’s global data, which is the same role HBM plays on an Nvidia part, so the bandwidth figures compare on a like axis rather than against a cache. For a company cut off from SK Hynix, Samsung and Micron by export controls, designing and producing its own such memory is a real achievement, and it removes the single most effective chokepoint the controls were built around. This is not a company that failed to get memory. It is a company that built its own.

And on that one axis, bandwidth, the memory it built is behind. The 950DT’s 4 terabytes per second sits at the level of Nvidia’s H200 in its HBM3e configuration, roughly 4.8, a chip that reached volume in 2024. Nvidia’s current B300 runs at 8. I am comparing memory bandwidth and nothing else here: interconnect, software ecosystem and total chip capability are separate questions on which the two companies sit very differently, and Huawei’s interconnect in particular is a genuine strength. But on the decode axis that sets inference cost per token, Huawei’s best available memory is about half of Nvidia’s current part and level with Nvidia’s part from two years earlier.

Both of these are true and neither cancels the other. Self-sufficiency solves whether you can buy the chip at all. It does not solve what the chip costs to run per token. The first is a survival question and Huawei has answered it. The second is a margin question and the answer is a ceiling.

5. The Chip That Relieves the Bottleneck Ships in Q4

Now the timing, which is the whole point and the part the roadmap coverage skipped.

The 950PR, the prefill chip with 1.6 terabytes per second, reached mass production in the first quarter of 2026. The 950DT, the decode chip with 4, is scheduled for the fourth quarter. For most of 2026, the only Ascend 950 a Chinese company can actually buy in volume is the half built for prefill.

That is the chip whose memory is slowest, deployed into the phase that needs memory most. A company standing up domestic inference capacity in mid-2026 is either running decode on a 1.6-terabyte-per-second part built for prefill, or still buying restricted Nvidia, or renting from a cloud that does one of those two things on its behalf.

This is the sentence the roadmap turns into once you read the memory column and the ship dates together. The ceiling on Chinese inference margin in 2026 is not that the country cannot build a high-bandwidth decode chip. It has designed one. The ceiling is that the decode chip is not here yet, and until it is, the economics of serving a token run through a part that was built for the other half of the job.

Which is where the silicon rejoins the income statement.

6. Put the Bandwidth on Zhipu’s Gross Margin

The eighth issue of this publication took apart the valuation gap between Zhipu and MiniMax and found it sitting inside one number: Zhipu earns an 18.9 percent gross margin on its cloud and API line, the business of selling inference by the token. Call it 19. That number was a mystery on the income statement. It is not a mystery on the silicon.

An 18.9 percent gross margin means the cost of serving is 81 percent of the revenue from serving. The largest movable part of that cost is the number of tokens a given accelerator can produce per second, because that sets how much hardware and power it takes to meet demand. Decode throughput per chip is, to first order, memory bandwidth divided by the size of the model’s active parameters and context per token. Bandwidth is the numerator. Raise it and cost per token falls. Cap it and cost per token has a floor.

Zhipu runs its inference substantially on domestic silicon, which its filings describe as co-design and which its addition to the US Entity List, effective 16 January 2025, makes a requirement rather than a preference. In 2026 domestic silicon at volume means the 950PR and its 1.6 terabytes per second, or Huawei’s prior 910 generation, or Cambricon, an independent Chinese AI chip designer. None of those is the 950DT. The memory bandwidth that caps Zhipu’s decode throughput, and therefore floors its cost per token, and therefore ceilings its cloud gross margin, is a number set in Shenzhen and Huawei’s HBM line, not in Zhipu’s model.

The 18.9 percent is not a verdict on GLM. It is the reading, on an income statement, of a memory-bandwidth number the company does not control.

7. Whether the Split Even Holds

The honest risk to this thesis is that the prefill-decode split does not survive contact with deployment.

Nvidia explored a version of disaggregated inference with its Rubin CPX, a prefill-oriented part using cheaper GDDR7 memory, and that approach has reportedly been revised since, which is a caution against assuming a clean prefill-decode hardware split is where the industry settles. Running prefill and decode on physically different chips adds scheduling complexity, network traffic between the two pools, and utilisation risk if the mix of the two phases does not match the mix of hardware you bought. It is entirely possible that Huawei’s customers run both phases on whichever 950 they can get and never deploy the clean split the product names imply.

But notice that this does not rescue the margin. The demand that decode places on memory bandwidth does not disappear because the workload runs on a prefill chip. It reappears as idle arithmetic units and low throughput, which is to say as higher cost per token. The split is Huawei’s attempt to price the two phases efficiently. If the split fails to deploy, the memory bottleneck does not go away. It just stops being labelled, and shows up unpriced in the cost of sales instead.

There is a second caution, and it cuts the other way, so it belongs here. If memory bandwidth alone set gross margin, every company on the same domestic silicon would earn the same margin. They do not, and DeepSeek is the reason to believe it. Its V4 models, released in April 2026 under an open licence, are built around a hybrid sparse-attention architecture that compresses the KV cache, the exact thing decode has to stream through memory, to roughly a tenth of the previous generation’s, and cuts single-token inference to about 27 percent of the FLOPs. Those figures are from DeepSeek’s own technical report. The point is not the specific number. It is that decode’s memory load is not a fixed tax. It is an engineering variable, and a model architected to move less data per token gets closer to the bandwidth ceiling than one that is not. Bandwidth sets the ceiling. Architecture decides how close you get. So the claim is not that Zhipu’s 18.9 percent is fixed by Huawei alone. It is that the ceiling above it is, set by the memory Zhipu can buy, and that no amount of engineering lifts that ceiling. It only closes the distance to it, and that distance is the part Zhipu still owns.

8. A Layer’s Margin, Written on a Roadmap

Step back to the machine.

The sixth issue argued that China’s compute ceiling is a memory ceiling. The eighth found that ceiling resting on a model company’s gross margin. This issue is the seam between them, and Huawei drew it in public: two chips, one bottleneck, and the relief scheduled for Q4.

The forward marker is specific and it is dated. The 950DT enters mass production in the fourth quarter of 2026. When it does, the bandwidth available to Chinese decode roughly doubles at the leading edge, from the 950PR’s 1.6 to the 950DT’s 4, and the cost-per-token floor drops with it. That is the moment the ceiling on Chinese inference margin lifts, if it lifts. Watch the ship date, watch whether the 4 terabytes per second is real in volume rather than on a slide, and watch whether Zhipu’s next disclosure shows a cloud gross margin climbing off 18.9 as the better memory arrives.

If the 950DT ships on time and Zhipu’s cloud margin is still stuck a year from now, the problem was never the memory and this thesis is wrong. If the 950DT slips, or ships thin, the ceiling stays where the 950PR put it, and the 18.9 percent is not a transitional number but a structural one.

A model company’s gross margin is not set inside its model. It is set on a wafer it does not own, in a fab it does not run, under a memory bandwidth it did not choose and cannot lift, on a schedule published by the company that does. Huawei cut the seam through its own product line and told you what it was called. It runs through everyone else’s income statement too.

Sources and Claim Types

Primary, from Huawei’s own document. The Ascend 950PR and 950DT memory capacities and bandwidths, the prefill and decode designations, the third-generation DaVinci architecture, the KV Cache statement, and the Q1 and Q4 2026 production timing come from Huawei’s Ascend 950 NPU Architecture Whitepaper, copyright 2026, and from Xu Zhijun’s roadmap presentation at Huawei Connect 2025 on 18 September 2025. The 950PR memory is 128/112 GB at 1.6/1.4 TB/s and the 950DT is 144/96 GB at 4 TB/s per the whitepaper’s specification table.

Primary, Nvidia. The B300 and GB300 figures of 288 GB HBM3e and 8 TB/s, and the H200 comparison at roughly 4.8 TB/s in its HBM3e configuration, are Nvidia’s published specifications.

Primary, US government. Zhipu’s addition to the Entity List, effective 16 January 2025, is from the Federal Register notice by the Bureau of Industry and Security. The Rubin CPX reference is Nvidia roadmap coverage and is treated as reported, not filed.

Confirmed, from filings. Zhipu’s 18.9 percent cloud gross margin and 48.8 percent on-premise gross margin are from its 2025 annual results announcement, reconciled to the audited gross profit and cross-checked against the prospectus cost-of-sales disclosure, as set out in full in the eighth issue. The blended figure is 40.96 percent.

Estimated, secondary. 950PR shipment and order figures, and named early customers, come from Chinese technology press citing supply-chain sources and are not company-confirmed. DeepSeek’s reported inference gross margin above 50 percent comes from secondary coverage of a third-party report, not a filing, and is used only as a directional counterweight, not as a load-bearing number. Both are flagged where they appear.

Inside China’s Machine is research, not investment advice. Nothing here is a recommendation to buy or sell any security.

Three Prospectuses, One Allocation

Three Chinese robot makers are in the listing queue on two exchanges. They build different machines, sell to different customers, and run opposite income statements. One leads global humanoid unit shipments and earned RMB 591mn on an adjusted basis last year. One is a niche industrial player that crossed into profit for the first time in 2025, by RMB 15mn. One has lost money in every year of its reporting period and tells the exchange it may not break even before 2028.

They have written the same use-of-proceeds table.

Unitree will raise RMB 4.202bn and has assigned RMB 2.022bn of it, 48.13 percent, to a single line called the intelligent robot model R&D project. Deep Robotics will raise RMB 2.503bn and has assigned RMB 1.169bn, 46.7 percent, to an embodied algorithm and model R&D project. Leju will raise RMB 2.600bn and has assigned RMB 616mn, 23.7 percent, to building a high-quality large-scale dataset, plus RMB 940mn, 36.15 percent, to an embodied intelligence R&D centre.

Those four lines total RMB 4.75bn against RMB 9.31bn of combined proceeds. Fifty-one percent.

That is the ceiling, and one boundary has to be drawn on it before it can be used. Unitree’s and Deep Robotics’ projects are pure model programmes. Leju’s R&D centre is not: its disclosed scope covers toolchain development and core algorithm work, but also robot body structure design, which belongs to the hardware layer. Strip it out, keep only the unambiguous brain-and-data lines, and the figure is RMB 3.81bn, or 40.9 percent.

So between 41 and 51 percent of every yuan these three companies are asking the public market for is earmarked for a model and the data to train it. At the bottom of that range it still exceeds what they are spending on factories, robot bodies and everything else combined.

This is not a robot financing. It is a model financing, and the robot business is the collateral.

Claim type: Confirmed. Prospectuses filed with the Shanghai and Shenzhen exchanges.

The Money Is Not Buying Robots

Start with what the raise is not for, because the market’s read of these companies is a hardware read, and the hardware read is correct as far as it goes.

Unitree’s hardware business is exceptional and there is no need to hedge that. Revenue went from RMB 159mn in 2023 to RMB 1,699mn in 2025. Main-business gross margin went from 44.22 percent to 60.13 percent. Operating cash flow in 2025 was RMB 670mn.

It did this while cutting prices. Average humanoid selling price fell from roughly RMB 593,000 in 2023 to roughly RMB 166,000 in 2025, a decline of about 72 percent. Quadruped pricing fell far less, from roughly RMB 38,600 in 2022 to RMB 30,300 in 2025, so this is not a company-wide fire sale. It is one product line being driven down a cost curve.

A company that raised gross margin by sixteen points while cutting the price of its fastest-growing product by seventy-two percent is not running a hardware problem. It has solved one, by building the motors, the reducers, the encoders, the dexterous hands and the lidar itself, and by keeping bought-in components to between 14 and 18 percent of cost.

Which raises the question the use-of-proceeds table answers. If the hardware is solved, why is only 14.9 percent of Unitree’s raise going to a factory?

Because the thing all three are short of is not a robot. It is a brain.

What the Filing Admits

The most valuable sentence in Unitree’s prospectus is a disclaimer. The company asking for RMB 2.022bn to build an embodied foundation model is telling you it does not currently have one in a product.

The filing states that because embodied large-model technology is worldwide still at the research and testing stage, the company has not, during the reporting period, deployed its self-developed general embodied model at scale in any robot product. What ships inside the machines it sells is the self-developed motion-control model, which the industry calls the cerebellum, plus a third-party large language model for voice interaction.

That is the leader in humanoid unit shipments on its own disclosed count, more than 5,500 units in 2025, saying the intelligence inside its product is a control loop and someone else’s chatbot.

Deep Robotics says the harder thing, in the industry section of its filing. It tells the exchange that no embodied large model capable of supporting the full chain of multimodal perception, real-time interaction and fully autonomous decision execution is yet mature, and that the field is still exploring several routes in parallel with no converged direction and no industry consensus.

The routes it names are genuinely different bets. A vision-language model maps pixels to words, and something else must turn the words into motion. A vision-language-action model skips the words and maps pixels straight to joint commands, which is faster and far hungrier for recorded robot trajectories. A world model does neither: it learns to predict what happens next in the physical scene, and lets the robot plan against its own simulation of the future. Different data, different compute, different timelines.

So Deep Robotics is requesting RMB 1.169bn from public investors to spend on a technology, and telling the regulator in the same document that the field has not agreed on which technology it is.

None of this is a scandal. It is unusually candid disclosure and the system worked. The point is narrower, and it is a pricing point.

A Programme Neither Has Ever Run

Deep Robotics itemises its research spending by year in the filing. RMB 32.18mn in 2023, RMB 38.21mn in 2024, RMB 84.30mn in 2025. RMB 154.69mn across the entire reporting period. The single model project it is asking the public market to fund is RMB 1,169mn, which is 7.6 times everything it has ever spent on research and development.

Unitree is the same shape at larger scale. Its R&D expense for the first nine months of 2025, the largest research period in its history, was RMB 90.21mn, a research intensity of 7.73 percent of revenue against a comparable-company average of 27.92 percent. That is the cost structure of a premium consumer brand, not of a frontier technology company, and until now it was a compliment about discipline rather than an accusation. The model project is RMB 2,022.46mn, roughly twenty-two times that nine-month figure.

Name the definitional boundary rather than hide it. Expensed R&D and a multi-year use-of-proceeds budget are not the same measure, and dividing one by the other produces no meaningful multiple in any accounting sense.

So the claim is not arithmetic. It is this: two companies whose entire institutional experience is of running research programmes in the tens of millions of yuan a year have told the exchange they will now run one in the billions. That is not a scale-up. It is a change of species, attempted by hardware organisations, at the same moment, on the same thesis, with no shipped product to validate it.

The Brain It Must Beat Is Free

On 8 July 2026, six days after the CSRC waved Unitree’s registration through, Ant Group’s embodied-AI subsidiary Ant Lingbo open-sourced LingBot-VLA 2.0. Weights on Hugging Face and ModelScope, code on GitHub. Free.

The specification is the part that matters. The model was pre-trained on 60,000 hours of real physical data: 50,000 hours cleaned from a 90,000-hour corpus of real-robot recordings, plus 10,000 hours distilled from 20,000 hours of first-person human manipulation footage. Its pre-training coverage spans 17 robot brands and 20 morphologies. Two of those brands are Unitree and Leju. The efficient post-training variant runs inference in under 130 milliseconds on an RTX 4090, which is a consumer graphics card.

Claim type: Confirmed as a company announcement and technical report, published 8 July 2026 and carried by Chinese technology press. Coverage of a brand in pre-training data is a capability claim, not an adoption claim. Nobody has said Unitree uses this model, and this piece does not say so.

Now hold the two documents next to each other.

Unitree is asking public investors for RMB 2.022bn to build an embodied foundation model it has not shipped. A week ago, a company with Ant Group’s compute and data behind it published a competent one for nothing, pre-trained on Unitree’s own morphology, running on a gaming GPU.

The honest counters are real and they must be said. A generic cross-embodiment policy is not the same thing as a model trained on your own fleet’s failure data; owning the model owns the improvement loop, and it owns the margin. Open weights today do not imply open weights at the frontier tomorrow, and Ant is a strategic actor with strategic reasons. And LingBot has not displaced anything: it is a floor, not a ceiling.

But the floor is exactly what moved. The thing RMB 2.022bn has to beat is no longer zero.

The Input Is Not for Sale

Here is why the free model is not merely competitive pressure but a statement about what the money can and cannot purchase.

RMB 2.022bn buys compute. It buys researchers. It buys simulation infrastructure and a data-collection organisation. It does not buy the scarce input, because the scarce input is recorded trajectories of a physical body doing physical work, and those are not sold on a market. They are manufactured, one hour at a time, by a robot doing a job.

So how much work are these robots doing?

In the first nine months of 2025, Unitree’s humanoid revenue was 73.60 percent research and education, 17.39 percent commercial and consumer, and 9.01 percent industry applications. That industry-application slice was RMB 53.60mn of main-business revenue.

The second-round inquiry response then does what only an inquiry response does. It forces the company to take that slice apart, and the company writes this:

Enterprise tour-guiding is roughly 50 to 70 percent of it. Sales into clearly identified working scenarios, meaning smart manufacturing, intelligent inspection and logistics delivery, come to RMB 15.70mn, or 29.29 percent of industry applications. And, in a sentence the company had no reason to volunteer, it tells the exchange that because downstream customers develop their own applications on top of the machines, it does not fully know what its industry customers use the robots for. The RMB 15.70mn is a preliminary count. The remainder is tour-guiding “as understood.”

Claim type: Confirmed. Primary source, second-round inquiry response filed with the Shanghai Stock Exchange, read this session.

Read the scope precisely, because it is doing load-bearing work. Nine months, not a year. Humanoids, not the whole product line. And it is the working slice of the industry slice.

The company raising RMB 2.022bn to train an embodied model does not know what its robots are doing. That is not an accusation, it is a disclosure, and it is the most honest sentence in the filing. But it is also a statement about the fuel, because the fuel is precisely a record of what the robot is doing.

The humanoid shipment leader put RMB 15.70mn of machines into identifiable work in nine months. That is the size of the factory floor from which the training data is supposed to come. Ant published 60,000 hours of it for free.

Readers of the weekly will recognise this constraint, and this is the piece where it acquires a price. The model layer’s bottleneck was never GPUs or headcount. It is that robots cannot read the internet, so the fuel has to be burned into existence by deployment, and deployment at the market leader is currently a rounding error that the company cannot fully see.

What You Are Actually Buying at RMB 42 Billion

Unitree’s registration became effective on 2 July 2026, and every outlet has reported the listing valuation as approximately RMB 42bn. Readers of this publication should recognise the shape of that number, because it is the construction we took apart in the CXMT piece.

RMB 4.202bn of planned proceeds, divided by a minimum issue of 40,446,400 shares, gives RMB 103.89 a share. Applied to the post-issue share count that a minimum ten percent float implies, that gives RMB 42.02bn. The valuation is not an appraisal. It is a project budget divided by a share count, and it is a floor derived from how much money the company needs rather than from what the company is worth.

The floor is not a bad guide to the issue price, and it is worth saying why. Moore Threads listed on the STAR Market in December 2025 with a planned raise of RMB 8.0bn and an actual raise of RMB 7.997bn. Under the current regime the book is built to land on the project requirement, so the price converges on the floor. What the floor tells you nothing about is the second number. Moore Threads issued at a market capitalisation of RMB 53.7bn, opened at RMB 650 against an issue price of RMB 114.28, and traded above RMB 300bn on day one. MetaX listed twelve days later and closed its first session up 692.95 percent.

So the range that matters is not RMB 42bn. It is what a market that has repriced two GPU issuers by five to eight times in a single session does to the only listed pure-play humanoid maker in China.

And the earnings you would price it against are about to become unreadable. Unitree’s 2025 reported net profit was RMB 278.21mn; adjusted net profit was RMB 590.75mn. The gap is a RMB 349.07mn non-cash share-based payment on a pre-listing equity grant, booked to administrative expense and classified as non-recurring. It landed almost entirely in the first half of 2025, which is why the company reported a net loss that half. Guidance now has first-half 2026 reported net profit swinging to RMB 258mn to 306mn while adjusted net profit falls 6.43 to 21.97 percent. Add the charge back to the 2025 base and the underlying reported number is flat to down. The turnaround the headline will show is the absence of an expense that was never cash. The number that tracks the actual bet, adjusted profit, is going the other way, because the model spending has started.

The first instalment is already billed. In the first quarter of 2026 revenue rose 68.49 percent while adjusted net profit fell 52.55 percent, driven by an incremental RMB 38.33mn of quarterly R&D. That is under two percent of the model project, and the market has already flinched at it.

So price what the prospectus separates for you.

The first thing is a business. A cash-generative actuator and motion-control franchise with a real cost moat, sixty percent gross margins, global unit leadership, and a humanoid revenue base still 73.6 percent research and education. It is worth a great deal more than the RMB 12.7bn post-money at which it last raised privately in June 2025.

The second thing is an option, and the prospectus prices it for you at RMB 2.022bn of your money. It is an option on an embodied foundation model whose technical route the industry has not chosen, whose scarce input arrives at RMB 15.70mn of identifiable working deployment per nine months, at a company that told the exchange it cannot fully see where that deployment goes, and whose competent free substitute was published last week by a subsidiary of Ant Group.

The judgment is not that the strategy is wrong. It is close to the only strategy available: a humanoid maker that does not own the brain becomes a contract manufacturer for whoever does, and Unitree has looked at that outcome and declined it. The allocation is correct.

The judgment is that the market is applying a hardware multiple to a document that describes a model bet, and the arithmetic that makes the bet legible is sitting in the use-of-proceeds table, unread.

What Would Settle It

The weekly said to watch industrial revenue share and not R&D expense. That was right, and it was answering a different question, so the two belong on the same dashboard rather than in competition.

Industrial revenue share tells you whether the bet has fuel. It is the leading indicator of whether trajectory data is being generated at all, and it is the one thing Ant cannot give Unitree for free, because a general policy trained on other people’s robots is not a closed improvement loop on your own. Watch also whether the company stops saying it does not know what its robots are used for. A firm that can see its own deployment has begun building the loop. A firm that cannot has bought GPUs.

R&D expense tells you whether the bet is being placed. The model project is the thesis of this IPO and, on the available evidence that Unitree does not capitalise development spending, the expense line is the only place it will ever be visible. If R&D does not step up by an order of magnitude from the RMB 90mn nine-month run rate, the RMB 2.022bn is a story and not a programme, and what investors own is the hardware business, which should be priced as one.

Neither line is revenue and neither is reported net profit. Both of those will look fine, and both will be lying.

Before either, the pricing announcement lands, probably within two weeks. It discloses the issue price, the issue P/E, and the earnings base used to compute it. Chinese issuers must use the lower of the pre- and post-adjustment figures, which for 2025 is RMB 278.21mn, not RMB 591mn. [VERIFY against the announcement when it appears.] If so, the headline multiple will be roughly 151 times rather than the 71 times every piece of coverage has implied, and it will be double for a reason that has nothing to do with the business.

Inside China’s Machine is research, not investment advice. Figures are sourced to filed prospectuses, inquiry responses, and company announcements, with claim types marked where the source is secondary or forward-looking.

1. The Market Compared Two Numbers That Do Not Measure the Same Thing

Zhipu is priced on a gross margin, and the gross margin does not survive being taken apart.

The price is not a small thing to be wrong about. The stock has risen roughly 1,650 percent since it listed in January, closing on 9 July at a market value of about 906 billion Hong Kong dollars. That same day the company placed new shares at a 13 percent discount and raised 31.4 billion, six times what its IPO brought in six months earlier.

The consensus chain underneath that price is short. Zhipu told investors in March that it raised API prices 83 percent in the first quarter of 2026 and that call volume grew 400 percent anyway. Price and volume rising together is the textbook signature of pricing power. Pricing power means a model customers cannot substitute. And the proof appeared on the income statement: Zhipu’s blended gross margin for 2025 was 41.0 percent against 25.4 percent at MiniMax, the other Chinese model company that listed the same week. One company sets prices. The other takes them.

Both gross margins are audited and both are correct. Neither is a measurement of the same business.

Zhipu’s 41 percent is mostly a services margin. MiniMax’s 25 percent is mostly a consumer margin. On the one line where the two companies sell the same product, at almost exactly the same scale, the ranking reverses, and it reverses by fifty points.

2. What Is Inside Zhipu’s 41 Percent

Zhipu reports two deployment lines and discloses the gross margin of each. None of what follows is hidden. It is simply not what the market has been reading.

On-premise deployment produced 534 million yuan of revenue in 2025, 73.7 percent of the company. Cloud deployment produced 190 million yuan, 26.3 percent. Total revenue was 724 million.

On-premise ran at roughly 49 percent. Cloud ran at roughly 19 percent.

Now set the 49 against its own history, which is where the story is. In the accountants’ report filed with the prospectus, the on-premise line earned 68.2 percent in 2023, 66.0 percent in 2024, and 59.1 percent in the first half of 2025. The full-year 2025 figure is 49. The company’s own explanation is that serving customers required more delivery resource. More delivery resource means more people.

A margin that falls when you add people is not a software margin. It is a delivery margin, and it is priced in engineer-months. What Zhipu sells a state enterprise is a model. What it delivers is an installation. The gross margin of that line tells you how many engineers the last contract needed and almost nothing about the model.

Confirmed. Zhipu prospectus, accountants’ report, audited by KPMG. Full-year 2025 segment figures from the annual results announcement, read through broker research quoting the announcement.

3. What Is Inside MiniMax’s 25 Percent

MiniMax splits differently, and the split is just as decisive.

AI-native consumer products produced 53.1 million US dollars in 2025, 67.2 percent of revenue. The open platform and enterprise services line produced 26.0 million, 32.8 percent, growing 198 percent. Blended gross margin was 25.4 percent, up from 12.2 percent.

The consumer line ran at a gross margin of 5 percent through the first nine months of the year, and that number inverts the economics that made consumer software work. Software distribution has a marginal cost near zero, which is why user growth and margin expansion were historically the same event. Inference does not behave that way. Every conversation buys another block of compute. Every generated video is a bill. Users do not dilute the cost curve, they walk along it. MiniMax had 1.77 million paying users at the end of September against 212 million people using its products. Fewer than one in a hundred pays, and the other ninety-nine still cost money to serve.

So MiniMax’s blended 25 percent is a consumer inference margin with a high-margin enterprise business sitting on top of it. Zhipu’s blended 41 percent is a project delivery margin with a low-margin cloud business sitting underneath it.

The blends are what the market compared.

Confirmed. MiniMax 2025 annual results announcement, audited by EY. Segment gross margins from the prospectus, read through broker research quoting the prospectus.

4. The Only Line They Both Sell

Strip out the parts that are not the same business and one line remains. Both companies sell access to a model through an API, metered by the token.

Zhipu’s cloud line, 2025: 190 million yuan.

MiniMax’s open platform line, 2025: 26.0 million US dollars, which at the 7.06 yuan to the dollar used in Zhipu’s own prospectus is 183 million yuan.

Two companies. Same product. Same year. The revenue on that line differs by under four percent.

Zhipu earned 19 percent on it. MiniMax earned 69.

The objection that arrives first is that the 69 percent must be a small-base artifact. It is not. The bases are the same size. The second objection is that MiniMax’s open platform line carries dedicated compute arrangements and model licensing alongside metered API calls, which will flatter it against pure inference. That one is fair, and it is worth something, and it is not worth fifty points. I have tried to construct a reclassification that closes fifty points and I cannot.

Which leaves the sentence the market has not said out loud. Zhipu’s 41 percent is not a report on its models. It is the accounting shadow of a services business, and the services business is the part that is shrinking as a share of the company.

5. What Is Actually Inside Cost of Sales

Here is where the comparison stops being an inference and becomes a fact, because both companies disclose the composition of the line the market was comparing.

MiniMax’s 2025 cost of sales was 59.0 million US dollars. The note to loss before tax gives cost of services provided, excluding employee benefit expenses, as 58.1 million. Read those two together, which is the reading the note invites, and the labour sitting inside MiniMax’s cost of sales is roughly 0.9 million dollars.

Under two percent of MiniMax’s cost of sales is people. The rest is compute.

Zhipu’s prospectus gives the same decomposition by nature. In the first half of 2025, salary cost was 39.4 percent of cost of sales and compute service fees were 37.6 percent.

Roughly four in every ten yuan of Zhipu’s cost of sales is payroll.

Two things have to be said about that comparison before it can be used, and the first is that I am about to be held to my own standard.

The periods do not match. MiniMax’s figure is full-year 2025. Zhipu’s is the first half, because that is where the prospectus stops and Hong Kong never required the rest. I cannot align them from outside, and the piece that opened by objecting to a period-blind comparison does not get to make one quietly.

And Zhipu’s salary share is falling. It was 54.4 percent of cost of sales in 2024 and 39.4 percent by mid-2025. The full-year number is very likely lower again. Extend the trend and Zhipu’s cost of sales converges on MiniMax’s shape: mostly compute, barely any payroll.

Follow that where it goes, because it does not rescue the 41 percent. It buries it. The more Zhipu’s cost of sales becomes a pure compute bill, the more honestly its cloud margin measures what inference actually costs the company. And that margin is 19 percent. The payroll inside cost of sales is not what is dragging the cloud line down. It is what is holding the blended figure up.

So the two numbers the market set side by side are not two measurements of the same thing under different conditions. They are two different line items wearing the same name. MiniMax’s cost of sales is an infrastructure invoice. Zhipu’s is substantially a wage bill with an infrastructure invoice inside it. A gross margin computed on the first tells you what inference costs. A gross margin computed on the second tells you how many engineers the last delivery took.

That is a definitional asymmetry. It is also, in this case, the valuation gap.

Confirmed. MiniMax 2025 annual results announcement, note to loss before tax, audited by EY. Zhipu prospectus, cost of sales by nature, audited by KPMG. The subtraction, and the reading of the note that permits it, are mine.

6. Pricing Power Is a Demand Curve, and Cost Per Token Is Set Somewhere Else

Now the other side of the argument, at full strength, because it deserves it.

Zhipu raised prices 83 percent and volume rose 400 percent. That is real, it is rare, and a company with a substitutable product cannot do it. And the cloud line has moved fast. The prospectus shows it earning 3.4 percent in 2024 and negative 0.4 percent in the first half of 2025, which is to say that eighteen months ago Zhipu was losing money on every token it sold. The full year came in at 19. Something changed in the second half, and it changed steeply.

Zhipu has also put a serious share of its research budget into what the filings call co-design, meaning adapting the model and the domestic accelerator to each other rather than porting one onto the other. As of June 2025 its models ran on more than forty chip platforms. That work is unglamorous, difficult, and the reason the models run at all under an export ceiling. It is not soft.

But two claims have been welded together, and the weld is where the mispricing lives.

The price went up. The unit cost did not come down enough to notice. After an 83 percent increase, the cloud line still earns nineteen cents on the dollar. Pricing power tells you what a customer will pay for a token. It tells you nothing about what the token cost to make.

And what the token cost to make is not, in the end, a question about the model.

Serving a large model is not compute-bound. It is memory-bound. At every generated token the accelerator streams the active weights and the accumulated context out of memory and into the arithmetic units, and the arithmetic finishes long before the data arrives. The chip idles, waiting. What sets throughput is memory bandwidth, not the headline compute figure on the brochure. Throughput per chip is how many tokens one accelerator produces per second. Divide the cost of owning and powering that accelerator by the tokens it produces and you have cost per token, which sits in the denominator under every gross margin in this industry.

Memory bandwidth sets throughput. Throughput sets cost per token. Cost per token sets the gross margin of anyone who sells tokens.

This publication has priced that chain before, from the other end. The sixth issue argued that China’s compute ceiling is a memory ceiling rather than a logic ceiling: the dies are fine, high-bandwidth memory is the scarce input, and a stockpile is not a capability. That was an argument about silicon.

It has arrived somewhere new. It is now sitting on an income statement.

7. Fifty Points Have Only Two Places to Come From

Two companies sell inference through an API at the same scale. One earns 69 percent and one earns 19. Given the mechanism above, a gap that size has exactly two possible sources: what the compute costs, and who sends the bill.

What the compute costs. Zhipu’s inference runs substantially on domestic accelerators, which is what the co-design spending buys and what the January 2025 Entity List designation makes necessary. Domestic accelerators are memory-constrained by construction. A memory ceiling is therefore a gross margin ceiling, and it has landed on the income statement of a company that never bought a wafer.

Who sends the bill. MiniMax’s cost of sales is, on its own disclosure, roughly 98 percent infrastructure. It buys that infrastructure from Alibaba Cloud. Alibaba was also its largest institutional shareholder before listing, holding roughly 15.66 percent. And Alibaba contributed approximately 22 percent of MiniMax’s revenue in 2024.

A supplier, an owner, and a customer, in one counterparty, sitting on top of a cost line that is almost entirely that supplier’s invoice.

I want to be careful here, because the careless version of this paragraph is worth nothing and the careful version is worth the subscription. I am not asserting that Alibaba subsidises MiniMax’s compute. Neither filing decomposes the fifty points, and I cannot decompose them from outside.

What I am asserting is narrower and much harder to argue with. A 69 percent gross margin, earned on a cost line that is 98 percent an invoice from your largest shareholder, is a number whose composition nobody has asked about. A 19 percent gross margin, earned on memory-constrained domestic silicon and diluted with the payroll of a delivery organisation, is a number the market has read as a verdict on the model. Both readings are unexamined. They point in opposite directions. And the valuation gap between the two companies has never rested on anything else.

8. The Two Lines Are Crossing, and Where They Meet Is a Silicon Number

The prospectus gives audited half-year figures. The results announcement gives the full year. Subtract the first from the second and the second half of 2025 falls out.

Cloud: 29.1 million yuan of revenue in the first half at negative 0.4 percent. 190 million for the year at 19 percent. The second half therefore ran roughly 161 million yuan at about 22 percent.

On-premise: 162 million yuan in the first half at 59.1 percent. 534 million for the year at 49 percent. The second half therefore ran roughly 372 million yuan at about 45 percent.

The two reconstructions add to 298 million yuan of gross profit against 297 million reported. The derivation holds.

It rests on two rounded full-year figures, so it is worth stating how much the rounding can move it. Flex the full-year cloud margin between 18.5 and 19.5 percent and the second half lands between 22 and 23. Flex on-premise between 48.5 and 49.5 and the second half lands between 44 and 45. The direction and the magnitude survive the rounding. The decimal places do not, and I am not claiming them.

Read the trajectory rather than the snapshot.

In one half-year the gap between Zhipu’s two business lines closed by thirty-six points. The high-margin line is falling and the low-margin line is rising, and they are converging on each other.

That matters because of an identity. Blended margin is on-premise margin times its revenue share plus cloud margin times its revenue share. When the two line margins are equal, the blend stops caring about the mix entirely. Zhipu’s blended gross margin is therefore converging on the point where its two lines cross, and it will get there regardless of what happens to the revenue mix.

The 49 percent full-year figure the market is anchored to is already stale. It contains a 59.1 percent first half. The exit rate is 45, and falling. The exit rate on the cloud line is 22, and rising. Naively extended, they meet somewhere in the mid-thirties.

Hold on-premise at its exit rate of 45 and vary the cloud line, and the blend looks like this.

Scenario grid, not a forecast. On-premise held at its derived second-half exit rate of 45 percent, which was 66 percent eighteen months earlier and is still falling. Segment margins for full-year 2025 per broker research quoting the results announcement; half-year figures per the audited prospectus; the second-half derivation is mine.

Zhang Peng has said he wants API at half of revenue. Read the grid. At the cloud line’s exit rate, half of revenue puts the blend at 34 percent, seven points below what the market is capitalising.

The strategy succeeding and the margin falling are the same event.

Now the number that decides everything, and it is buried in the prospectus rather than the headline.

In the first half of 2025, Zhipu spent 1,145 million yuan on compute service fees inside research and development, 71.8 percent of its R&D budget. In the same six months it spent 35.8 million yuan on compute inside cost of sales.

Zhipu bought thirty-two times more compute to train with than to serve with.

The 19 percent cloud margin, and the 22 percent exit rate the bull case leans on, are earned on a serving business that consumes three percent of the company’s compute. It has never been tested at scale. Scale it ten times and the cost structure it rests on is an assumption, not a result. And the price at which that compute is bought is set by Ascend memory supply and Cambricon’s production ramp, not by GLM.

Layer on the open-source floor. DeepSeek released V4 in April, reported to run natively on Ascend and Cambricon with its Flash variant priced in single-digit yuan per million tokens. That specification comes from secondary technology press and I have not verified it against DeepSeek’s own release. If it is even directionally right, Zhipu’s 83 percent price increase has an open-source floor sitting directly above it, running on the same domestic hardware.

And then, in the week this piece was written, Zhipu answered the question itself.

On 9 July it placed new H shares at 1,588 Hong Kong dollars, a 13 percent discount to the previous close, raising 31.4 billion. That is six times what its January IPO raised. It comes after the company had already spent more than 93 percent of the IPO proceeds by the end of June. Six months of runway from a listing, then a raise six times larger.

Now read the stated use of proceeds. The two items the announcement leads with, in every account of it I have seen, are base model research and development and compute infrastructure. Commercialisation appears further down the list, and the accounts diverge on what sits beside it. I have not seen the announcement’s own allocation, and if it gives percentages I do not have them.

I do not need them. Zhipu has just told the market that twenty-seven billion yuan is going, first, into the base model and the compute to train it. That is the same place the thirty-two times ratio said the money was already going. Not into the cost of serving a token. Into the cost of training the next model.

9. What the STAR Market Will Force Zhipu to Disclose

There is a document coming that settles this, and it is worth waiting for rather than arguing about.

Zhipu’s board proposed a 15 billion yuan A-share issue on the STAR Market, with 12 billion earmarked for the base model, 2 billion for the model-as-a-service platform, and 1 billion for working capital. Shareholders approved it at the annual general meeting on 22 June 2026. Before a Chinese company can file to list domestically it must complete a supervised preparation period with a sponsoring broker, registered with the securities regulator. On 7 July Zhipu publicly denied press reports that it had withdrawn from that process, stating that the preparation is complete. MiniMax is pursuing a STAR listing of its own.

A STAR listing brings an exchange inquiry process, and the inquiry is where the Shanghai Stock Exchange forces companies to disaggregate what they would rather present blended. Cost of sales by nature, for the full year rather than the half. The unit economics of the cloud line under its commercial name. What twelve billion yuan of base-model spending actually buys.

Those are the three disclosures this piece turns on. Hong Kong required the first only up to June 2025, and the other two not at all.

10. The Best Argument Against This Piece Was Made by Zhipu, on 11 July

Tang Jie, one of Zhipu’s co-founders, circulated an internal letter on 11 July. Its reported content: while the industry rushes to monetise, Zhipu has decided to push upward instead, and will not chase near-term commercialisation of applications. The letter reached the press rather than the exchange, so treat it as reported and not as filed.

Take it at face value and the bull case becomes coherent in a way the gross margin cannot make it. A company that is not trying to earn a software margin cannot be faulted for failing to earn one. What shareholders are buying is not a cash flow. It is an option on capability. And a 1,650 percent return since January is the market saying, without ambiguity, that it was never reading the income statement in the first place.

That case is real. It is the one a reader should weigh against mine, and it may well be the one that pays.

But notice what it concedes. It concedes that the gross margin will not fix itself, and that nobody at the company is currently trying to make it. This piece has argued that the fifty points are set by silicon rather than by GLM. The letter answers that the fifty points do not matter. Those are not the same claim, and only one of them survives the day the market decides it wants to see a cash flow.

So the thesis is falsifiable on a date, by a filing, in public.

If the inquiry response shows cloud gross margin already through 40 percent, this piece is wrong and the next issue will say so, because more than half the required cost decline would already have happened and the bull case would be intact on its own terms rather than on the letter’s. If it shows compute rising as a share of cost of sales while salaries fall and the on-premise margin keeps compressing, then the 41 percent was always a services margin in retreat, and the fifty points were never about intelligence at all.

A model company’s gross margin is not set inside its model. It is set on a wafer it does not own, in a fab it does not run, under a memory ceiling it did not choose and cannot lift, and it is reported in a line item whose contents almost nobody opened. That is the machine. Zhipu is a cross-section of it, and so is the price.

Sources and Claim Types

Primary, audited. Zhipu prospectus of 30 December 2025, accountants’ report by KPMG: all segment revenues and gross margins through the first half of 2025, cost of sales by nature, research and development compute service fees. MiniMax annual results announcement of 2 March 2026, audited by EY: full-year revenue by segment, cost of sales, gross profit, the note to loss before tax, headcount.

Primary, reported through broker research. Zhipu’s full-year 2025 segment gross margins come from its annual results announcement, which I have read through Soochow Securities research quoting it rather than through the announcement itself. Those figures reconcile to the audited gross profit, which is the best check available without the filing in hand, and I have said where they are load-bearing. The Alibaba shareholding and revenue concentration reach me the same way, through research quoting the MiniMax prospectus.

Market data and July events. The placement of 9 July, its price, discount and size, the 93 percent utilisation of IPO proceeds, the 22 June shareholder approval of the A-share issue, the 7 July denial of the withdrawal reports, and Tang Jie’s letter of 11 July all reach me through Chinese financial press reporting company announcements, not through the announcements themselves. I have not read the placement announcement, and I do not have its allocation of proceeds by percentage. The accounts I have seen agree on the two items it leads with and diverge on what follows, and the piece claims no more than that. The market capitalisation and the return since listing are prices, not disclosures. They move daily, they are quoted as of 9 July 2026, and every conclusion here is about the numerator rather than the denominator. Tang Jie’s letter is internal correspondence reported by the press and is treated as reported throughout.

Company-reported, unaudited. The 83 percent price increase, the 400 percent volume increase, and the MaaS annual recurring revenue figure, all from the March 2026 results call.

Derived, mine. The second-half 2025 segment margins, the yuan-dollar comparison of the two API lines, the labour share of MiniMax’s cost of sales, the sensitivity grid, and the ratio of training compute to serving compute.

Secondary and unverified. The DeepSeek V4 specifications and pricing.

Inside China’s Machine is research, not investment advice. Nothing here is a recommendation to buy or sell any security.

CXMT opens investor subscription on July 16, under the code 688825, seeking 29.5 billion yuan on the STAR Market. It is the largest A-share offering of 2026 and, by planned proceeds, the second-largest STAR Market listing ever, behind SMIC.

Run the obvious calculation and the stock looks like a gift. The company plans to raise 29.5 billion yuan for a stake of not less than 10 percent of its post-issue equity. Divide the one by the other and the planned raise implies a market capitalization near 295 billion yuan. Its own guidance puts first-half 2026 net profit attributable to the parent at 50 to 57 billion yuan. Annualize the midpoint and you get a company trading at under three times forward earnings.

Three times. For the national champion of Chinese memory, at the top of the sharpest memory upcycle the industry has produced.

That number is now circulating everywhere, and it is the reason Chinese market commentary is discussing a post-listing valuation of two to three trillion yuan.

It is also not a price.

The multiple everyone is quoting is a placeholder

CXMT’s book-build opens on July 13. The final offering price is published on July 15. Subscription is on the 16th. As of this writing, no investor has told anyone what CXMT is worth.

The 295 billion yuan figure is what you get when you divide a planned fundraise by a minimum float. Chinese coverage is precise about this and treats it as a floor. The share count is capped; the price floats with the book, and the raise floats with the price. Whatever the institutions bid, the capitalization moves with them. SMIC is the precedent that matters here: in 2020 it planned to raise 20 billion yuan on this same exchange and walked away with 53.2 billion.

So the honest version of the opening arithmetic is this. Under three times forward is not the market’s judgment of CXMT. It is the lower bound of a judgment that has not been made yet. The event is not the listing. The event is Wednesday.

And the exchange has built a tell into the process, which is worth reading carefully because it measures two different things. Under the terms of CXMT’s July 9 book-build announcement, the issuer and its underwriters must publish a special investment risk announcement before subscription opens if either of two conditions is met. The first is that the final price implies a P/E above the trailing one-month average for the industry published by China Securities Index Co. That tests whether the price is high against the sector. The second is that the final price exceeds the lower of two reference points drawn from the book itself: the median and weighted average of surviving offline bids after the highest are struck out, and the median and weighted average of the bids from public funds, the social security fund, pensions, annuities, insurance capital and qualified foreign investors.

Read the second condition again. It tests whether the underwriters priced the deal above where China’s longest-horizon institutional money was willing to go. That is not a valuation test. It is a test of who set the price, and it is the more informative of the two.

That document is either published on Wednesday or it is not. It is the cheapest possible read on what the A-share market decided memory is.

The control group priced itself six days ago

Before guessing what CXMT will fetch, look at what a competitive book-build just paid for the best asset in the category.

On July 10, SK Hynix listed American depositary receipts on the Nasdaq, pricing at 149 dollars and raising 26.5 billion dollars. The stock closed its first session at 168.01, valuing the company near 1.27 trillion dollars. Per FactSet, that puts the world’s HBM leader, holder of roughly 58 percent of the HBM market and the sole reason a large share of Nvidia’s accelerators function, at about 5.4 times forward earnings. Micron sits near 6.7 times forward, against roughly 22 times trailing.

These are the numbers being paid for companies whose results are, on every measure they themselves report, the strongest in their histories.

Micron’s fiscal third quarter delivered revenue of 41.5 billion dollars, up 346 percent year on year, with non-GAAP gross margin at 84.9 percent and non-GAAP earnings per share of 25.11 dollars against 1.91 a year earlier. That is a thirteenfold increase in twelve months, and management then guided the following quarter to 50 billion dollars of revenue and 31 dollars of EPS. SK Hynix posted first-quarter revenue of 52.6 trillion won, operating profit of 37.6 trillion won at a 72 percent operating margin, up 405 percent year on year, and net profit of 40.3 trillion won at a 77 percent net margin. All three incumbents crossed a trillion dollars of market value in May.

And then they fell. As of the second week of July, Micron, Samsung, SK Hynix and the Roundhill memory ETF are each more than 20 percent below recent closing highs. Semiconductor market value has contracted by roughly 1.5 trillion dollars since June 25 on Yahoo Finance’s basket, with Micron alone down close to 350 billion. The Kospi, where Samsung and SK Hynix now account for about half the index weight, triggered circuit breakers twice, on June 23 and again on July 2, the latter after Meta signalled it would resell surplus AI compute rather than keep absorbing it.

Note carefully what did not happen in those three weeks. DRAM contract prices did not fall. TrendForce’s June 30 update has them still rising on tight supply and low inventories. The drawdown happened anyway.

That is the lesson the memory industry has beaten into every investor who has ever touched it, arriving in real time. Peak-cycle earnings are not a base. They are a peak. A single-digit multiple on record earnings is not a market calling a company bad. It is a market saying the denominator is borrowed. The multiple compresses precisely because the earnings are inflated, and it starts compressing before the price series turns, because the market is pricing the turn rather than reporting it.

The incumbents’ cheap multiples are cheap because the fall has already been discounted. CXMT’s cheaper multiple is cheaper because nobody has discounted anything yet, because nobody has priced it at all.

What the incumbents are doing that CXMT is not

There is a serious counterargument to all of this, and it deserves its hearing before the argument continues.

Nomura’s position is that memory is short, not long: the big three are cannibalizing commodity DRAM capacity to feed HBM, which tightens DDR5 rather than loosening it, and the market is pricing a supply risk a decade early. UBS and Bank of America have both framed the July drawdown as a reset inside a supercycle rather than a break in it. The HBM bottleneck is widely expected to persist into 2027. Micron guided up, not down.

More important than any of that, Micron is doing something structural. It has signed sixteen multi-year strategic customer agreements with take-or-pay terms, covering roughly 20 percent of DRAM and 30 percent of NAND volume, and it puts the floor-price revenue under those agreements at about 100 billion dollars, or roughly a quarter of revenue over their term. That is a company converting cyclical earnings into contracted earnings. It is the most credible answer anyone in memory has ever given to the “it is just a cycle” objection.

CXMT has disclosed no equivalent. Its earnings are spot earnings. Hold that thought.

CXMT’s margin is a price, not a cost position

Now go back to the denominator and ask what is inside it.

CXMT earned its first annual profit in 2025, at 1.875 billion yuan attributable to the parent. Its guidance for the first half of 2026 alone is 50 to 57 billion. In six months, it is guiding to roughly thirty times what it earned in the whole of the prior year.

Nothing about the company improved by a factor of thirty in twelve months. The price of DRAM did.

And here is the number that settles what kind of profit this is. SemiAnalysis, reconciling the prospectus against its own memory model, finds CXMT’s cost per bit on DDR5 remains more than 30 percent above that of the three global incumbents, while its average selling price sits only 5 to 10 percent below theirs. It is not the low-cost producer and it is not undercutting anyone. It is a price-taker inside a shortage, and its gross margin of over 70 percent is the arithmetic output of that shortage rather than of any cost or technology advantage.

This distinction is the whole thing, and most coverage misses it. When Micron earns an 85 percent gross margin, that margin sits on top of the industry’s best cost structure, which means it survives a price decline longer than anyone else’s. When CXMT earns 70 percent on a cost base 30 percent worse than the leaders, the same price decline reaches it first. In a boom, a high-cost producer is indistinguishable from a low-cost one. In a bust, it is the first thing that stops working.

So the sub-three-times multiple is real arithmetic on a temporary denominator, computed off a price that does not yet exist.

The option the company declined to fund

There is a version of the bull case that survives everything above, and it deserves a fair hearing, because it is the case that would actually justify a franchise multiple.

If CXMT is not really a DRAM company but a future HBM company, then commodity DRAM earnings are merely the funding mechanism for something far more valuable, and pricing it on the DRAM cycle is a category error. Under that reading, a rich multiple is not exuberance. It is a call option on China’s compute independence, and cheap at almost any price.

That is a real argument. It has one problem, and the company put it in the prospectus.

Of the 29.5 billion yuan in net proceeds, 7.5 billion funds a wafer-manufacturing line upgrade and 13 billion funds DRAM technology upgrades, together 69.5 percent. The remaining 9 billion, or 30.5 percent, funds forward-looking DRAM research. There is no dedicated HBM project and no separate HBM funding line. SemiAnalysis, going through the use of proceeds line by line, reports that the prospectus does not mention HBM at all. Some generalist coverage has characterized the raise as funding an HBM3E push. The filing is the arbiter, and the filing is silent.

The supporting evidence points the same way. Roughly 99 percent of CXMT’s 2025 revenue came from DDR and LPDDR. SemiAnalysis puts HBM at about 5,000 of roughly 265,000 monthly wafer starts at the end of 2025, rising to perhaps 30,000 by the end of 2026, and describes the company as still unable to stabilize 8-high HBM3, with 12-high harder again.

None of this is irrational. It is the opposite. TrendForce found that in the first quarter of 2026, the per-wafer revenue and profitability of 64GB DDR5 server modules surpassed HBM for the first time in the industry’s history. HBM dies are larger, because of the through-silicon vias, so the same wafer yields fewer of them, and then stacking takes another cut. For a company optimizing its own economics right now, every wafer moved to HBM is a wafer that earns less.

And this is where the seam opens. Beijing needs HBM, because the model layer cannot run without it and export controls have made the shortfall a national problem. CXMT’s unit economics say make commodity DRAM. With 29.5 billion yuan in hand and a prospectus that does not name HBM, the company has told you which instruction it is following. The IPO does not merely fail to fund the option the market may be paying for. It funds the wafer capacity that competes with it.

What is actually being priced

Strip it down. At any price the book produces on Wednesday, an investor is buying two things bolted together.

The first is a high-cost commodity memory manufacturer at some multiple of peak-cycle earnings, with no take-or-pay contract structure disclosed, whose costs sit 30 percent above the leaders and which therefore breaks before they do when DRAM rolls over.

The second is a lottery ticket on a national HBM program that the company’s own use of proceeds does not finance, and which fails slowly, quietly and unnewsworthily, as nothing happening for several years.

Both are real. The DRAM franchise in particular is enormous and China genuinely needs it. But they have to be priced separately, because they fail for entirely different reasons and on entirely different clocks.

The market is not pricing them separately. It is pricing a story. The arithmetic behind the two-to-three-trillion talk is visible if you look for it: assume full-year 2026 attributable profit of 150 to 200 billion yuan, apply the 20 times that Chinese brokers apply to strategic semiconductor assets, and you land at three to four trillion. Every step is defensible in isolation. The first step annualizes a peak. The second step applies a franchise multiple to it. Do both at once and you have paid a franchise price for a cycle.

For scale, at three trillion yuan CXMT would be the largest company on the A-share market, ahead of ICBC at roughly 2.7 trillion. The last private round, in which Alibaba put in 6.1 billion yuan for 3.85 percent, valued the company at 158.4 billion. The market is discussing a step-up of roughly twenty times, in about a year, on a company whose cost position has not moved.

What to watch, and when

Three things, in order of how quickly they resolve.

Wednesday, July 15: the offering price, and whether the special risk announcement appears. It appears if the price is rich against the sector, and it also appears if the price was set above where the long-horizon institutions bid. Either way, the issuer must publish a document explaining why. That document is the A-share market stating its theory of memory out loud, in a filing, on a date.

Monthly: the DDR contract price. TrendForce publishes it. Every yuan of the denominator under every multiple discussed above rests on it. It has not turned. It was still rising as of June 30, and the incumbents fell 20 percent anyway, which tells you the equity market does not wait for the print.

Quarterly: CXMT’s HBM wafer allocation, if it is ever disclosed. Not HBM press releases, which will be enthusiastic and unfalsifiable. The wafer count. That number, and only that number, tells you whether the option is being funded.

Read the seam between the cycle and the story

Silicon is the substrate of the machine this publication reads, and memory is the substrate of the substrate. Nothing above it runs without it. The model layer’s ceiling is set by HBM, and HBM is set by wafers, and wafers are allocated by whoever is optimizing the margin on them.

Which is why the most important line in this prospectus is one that isn’t there. A company can be strategically indispensable and financially cyclical at the same time, and the market’s characteristic error is to collapse those two facts into one. The world’s best memory companies are indispensable and they trade at five to seven times earnings, because indispensable and cheap are not a contradiction in a commodity business. They are the normal condition of one.

This week, the global market marked that down by a fifth while contract prices were still climbing. Five days later, in a market that has just pushed SMIC past Moutai, the same commodity comes up for sale again.

The A-share market gets to state its theory on Wednesday. It will be published, it will be dated, and it will be wrong or right for reasons that are already visible in a document the company has already filed.

Inside China’s Machine is research, not investment advice.

Verification appendix.

Confirmed, from CXMT’s STAR Market prospectus and its July 9 book-build announcement filed with the SSE: offering size of 29.5 billion yuan for not less than 10 percent of post-issue equity, with a greenshoe; book-build July 13, pricing July 15, subscription July 16; use of proceeds of 7.5 billion, 13.0 billion and 9.0 billion yuan across three named projects; 2025 attributable net profit of 1.875 billion yuan; first-half 2026 guidance of 110 to 120 billion yuan of revenue and 50 to 57 billion yuan of attributable net profit; the two conditions triggering the special investment risk announcement. The 295 billion yuan capitalization is arithmetic on the planned raise and the minimum float, not a price.

Confirmed, from company reports: Micron fiscal Q3 2026 revenue, gross margin, EPS, Q4 guidance and the sixteen strategic customer agreements; SK Hynix Q1 2026 revenue, operating profit, operating margin and net margin; SK Hynix ADR pricing, proceeds and first-day close.

Market data, current as of July 10 to 12, 2026, and moving daily: forward multiples for Micron and SK Hynix, per FactSet; the drawdowns in Micron, Samsung, SK Hynix and the Roundhill memory ETF; aggregate semiconductor market-value decline, per Yahoo Finance’s basket; ICBC and SMIC A-share market capitalizations. Forward multiples for CXMT are computed by annualizing company guidance against a market capitalization implied by the planned raise, and are illustrative of a floor, not a projection.

Estimated, and attributed to SemiAnalysis, which reconstructs these from supply-chain sources and the filing rather than reading them off it: CXMT’s DDR5 cost per bit relative to the incumbents; its DRAM ASP relative to the incumbents; its HBM wafer allocation and trajectory; its HBM3 8-high yield difficulties; the absence of HBM from the use of proceeds. The DDR5-versus-HBM per-wafer profitability crossover is attributed to TrendForce.

Market commentary, cited as evidence of what is being discussed rather than as forecast: the two-to-three-trillion-yuan valuation range; the 150 to 200 billion yuan full-year profit and 20 times multiple used to reach it; Nomura’s shortage thesis; the UBS and Bank of America characterizations of the July drawdown.

Current as of July 12, 2026.

This was the week the brain got a valuation before it got a body of evidence. Unitree’s registration took effect, the pricing window opened, and every other humanoid company in the listing queue now has an anchor to be priced against. Read the three filings side by side and they tell one story from three angles: the money is flowing to the model, the model’s input is not there, and the input has become so scarce that it is now being sold as a product line. The exchange has already noticed. Its inquiry letters are the best sell-side research being published on this sector, and they are free.

The anchor is set. The CSRC approved Unitree’s registration on 2 July, and the exchange updated the file to effective on 6 July. The offering is at least 40,446,434 new shares for at least 10% of the enlarged capital, with RMB 4,201.71mn of proceeds, which sets an implied floor market capitalisation of roughly RMB 42bn. [Confirmed: prospectus. The price line is still blank, so RMB 42bn is a floor derived from the raise and the minimum float, not a market capitalisation.] Preliminary inquiry and pricing land in the next fortnight. From that moment, Deep Robotics, Lejuu, and everything behind them get priced off a single multiple.

The anchor’s own filing prices a brain it has not deployed. Unitree has earmarked RMB 2,022.46mn, 48.13% of the raise, for embodied model R&D. [Confirmed: prospectus, use-of-proceeds schedule] Against that, industrial deployment of its humanoids produced RMB 15.70mn of revenue in the first three quarters of 2025, which is 2.64% of a humanoid line worth RMB 595.19mn. [The humanoid revenue line is confirmed in the prospectus. The RMB 15.70mn figure is disclosed in the second-round inquiry response and is reproduced here as reported, reconciled against the prospectus denominator.]

That is 129 yuan committed to building the brain for every one yuan the body has earned doing the work the brain is supposed to enable.

Research and education account for 73.6% of that humanoid line. Industry applications account for 9.01%, and between half and seventy percent of those are corporate tour guiding: a humanoid standing in a company showroom, greeting visitors. What remains resolves into three named pilots. Live-line work at a 500kV substation in Zhejiang. Materials handling at CRRC Zhuzhou. Handling and assembly at NIO.

Three projects. Fifteen million yuan. That is the entire industrial trajectory pipeline of the company that shipped more humanoids than anyone on earth last year, over 5,500 units, bipedal, excluding wheeled dual-arm. [Confirmed: prospectus, definition as stated]

The company does not hide any of this. Its risk factors state plainly that it has not yet applied its self-developed general embodied model at scale in any product, and has run it only as a pilot inside its own facilities. [Confirmed: prospectus, risk factors] The gap is not an inference. It is disclosed.

The data is now a product, which is the tell. Lejuu’s ChiNext application was accepted on 19 May, raising RMB 2.6bn. It is the first company to list under that exchange’s fourth standard, the one that does not require profit, and it lost RMB 69.78mn on RMB 258mn of 2025 revenue. Two details matter more than the losses. Its largest revenue application scenario is data collection. And its use of proceeds includes a project to build a large-scale, high-quality dataset. [Reported: prospectus, via Xinhua Finance and Sina. Not checked against the filing.]

Read those two facts against the one above. The second humanoid company in the queue earns its biggest revenue line by manufacturing training data, while the first is raising RMB 2bn to acquire it. Trajectories have become a traded good before the robots that were supposed to generate them as exhaust have been deployed. That is not a business model. That is what a supply shortage looks like at the moment it gets priced.

The company in the queue that makes money does not make humanoids. Deep Robotics was accepted on the STAR Market on 18 May. It earned RMB 337mn in 2025 and posted its first profit, RMB 28.68mn, against a loss the year before. Quadrupeds and wheel-leg machines are over 95% of revenue. Humanoids are 0.24%. [Reported: prospectus, via Xinhua Finance. Not checked against the filing.] The revenue comes from power inspection, mining survey, and emergency response, which is to say from machines doing work somebody was already paying to have done.

This is the layer connection, and it is the whole argument in one comparison. The industrial revenue that humanoids cannot find, quadrupeds already have. The binding constraint is therefore not the body, and it is not manufacturing, and it is not cost. It is the distance between what a cerebellum can do, which is locomotion and inspection along a known route, and what a cortex must do, which is manipulation in a space nobody mapped in advance. Unitree’s own numbers say the same thing from the inside. Its quadruped line completed this migration, with research and education down to 31.6% of quadruped revenue. Its humanoid line has not started.

The control group. UBTech took the opposite route and paid for it in cash. It shipped 1,079 Walker units in 2025, grew that revenue line by over 2,000%, and lost roughly RMB 790mn doing it, with machines on production lines at BYD, Geely, and SF Express. [Reported: company disclosures, via trade press. The weakest sourced item in this issue.] Two strategies, one variable. Unitree monetised the laboratory and is now buying the brain with IPO proceeds. UBTech bought the factory access first and paid in losses. Both are purchasing the same scarce input. Within four quarters we find out which one bought the cheaper trajectory, and the answer sets the multiple for the entire layer.

What to watch. The pricing itself, and the issue P/E it implies. But the disclosure that actually resolves this thesis is narrower. Watch whether Unitree’s H1 report breaks out revenue by application scenario. If the industrial share rises from 9.01%, the earmark has an input supply and the model bet is live. If it does not move, RMB 2,022.46mn is funding a research programme rather than a model, and the multiple should be a hardware multiple. Watch also for the inquiry responses to Lejuu and Deep Robotics, because the exchange will force from them exactly the disaggregation it forced from Unitree, and that is where the numbers nobody wants to publish get published.

The layer is being priced in the next month. The evidence that would justify the price is being generated at the speed of one factory pilot at a time.

Inside China’s Machine. China is building the machine that builds physical intelligence. Silicon, models, robots, factories. We read it one layer at a time and turn each into capital judgment.

This is investment research, not investment advice.

Take apart an Ascend 910C, Huawei’s flagship AI accelerator, and you find something the press releases do not mention.

TechInsights did exactly that, and their teardown reports logic dies fabricated by TSMC in 2020, sitting alongside HBM2E memory stacks from Samsung and SK Hynix. Chinese trade coverage of the same chip has repeatedly described its localization rate as exceeding 90 percent.

The press release and the X-ray disagree. That gap is where this issue lives, because it points at the thing almost every analysis of China’s AI compute position gets wrong: the chokepoint is not where the export controls aimed it.

Everyone is watching the wrong layer

The public conversation about Chinese AI silicon runs on lithography. EUV machines, ASML, SMIC’s 7-nanometer process, the question of whether China can reach 5. It is a coherent story and it produces a clear scoreboard, which is probably why it dominates.

It is also, at this moment, not the binding constraint.

SemiAnalysis, working from supply-chain data, puts the position bluntly: China could easily build more than 805,000 Ascend accelerators in a year from available TSMC and SMIC capacity, but it will not, because it does not have enough HBM. The wafers are not the problem. The logic dies are not the problem. China has fab capacity it cannot convert into accelerators, because it lacks the memory to package next to the compute.

This deserves to sit for a moment, because it inverts the entire framing of the export-control debate. The controls were designed around the assumption that if you deny a country the ability to print advanced logic, you deny it advanced compute. China’s answer was to route around lithography with chiplets, older nodes, and brute force. That answer worked, in the narrow sense that the dies exist. And it ran straight into a wall nobody was pointing at, which is that an AI accelerator is not a logic die. It is a logic die surrounded by high-bandwidth memory, and the memory is the part China cannot yet make at scale.

The workaround for one bottleneck created another

Look at how the 910C is actually built, because the engineering tells you the cost of the detour.

Unable to print a large monolithic die at an advanced node, Huawei took two 910B dies and packaged them together, each on its own interposer, connected through an organic substrate. This is a legitimate strategy. It is cheaper, it yields better, and it ramps faster than the advanced packaging Nvidia uses.

It also has a price, and the price is bandwidth. Independent analysis suggests the die-to-die bandwidth of this arrangement runs an order of magnitude or more below Nvidia’s packaging, and the chip-to-chip interconnect sits around 400 GB/s, well under half of what NVLink provides. For workloads dominated by all-reduce communication, which is to say large-scale distributed training, that is precisely the wrong place to be slow.

So the substitution chain runs like this. Denied advanced lithography, China substitutes chiplets. Chiplets require advanced packaging to preserve bandwidth. Advanced packaging is also restricted. The result is a chip whose compute is respectable and whose ability to talk to itself, and to its neighbors, is not.

Reported memory bandwidth figures for the 910C vary across sources by nearly a factor of two, which is itself worth noting: this is a domain where confident numbers should be treated with suspicion. But every source, including the favorable ones, places it below the H100. The compute gap is manageable. The memory and interconnect gap is the one that decides what can actually be trained.

Huawei is running on inventory, and inventory has a maturity date

Here is the fact that turns a technical observation into a capital judgment.

Before the export controls on high-bandwidth memory took effect, Chinese buyers acquired a large stockpile of HBM stacks. SemiAnalysis, again from supply-chain reconstruction rather than any disclosure, puts it at roughly 13 million. An Ascend 910C carries eight. That stockpile is therefore sufficient to package on the order of 1.6 million accelerators, which is a large number and which is also, precisely, a finite one.

Treat the figure as an estimate, because it is one, and no Chinese entity has published it. But the shape of the claim does not depend on the precision of the number. Whatever the exact count, it was bought once, it is being drawn down, and nobody is selling more.

The same is true of the logic. TechInsights found TSMC dies from 2020 inside a 2025-era chip, and SemiAnalysis estimates that this die bank depletes within months rather than years.

Huawei’s AI accelerator business is currently running on a warehouse. That is not a criticism of Huawei, which stockpiled intelligently and under duress, and it is not a prediction of collapse. It is a statement about what kind of asset the current production run represents. Capability compounds. Inventory depletes. The two look identical on a shipment chart and they are not remotely the same thing, and anyone reading Ascend volumes as evidence of a structural position is reading an inventory drawdown as if it were a capacity curve.

Which means the entire question becomes: what replaces the warehouse?

The company that decides this is not the company everyone is watching

The answer is CXMT, ChangXin Memory Technologies, and it is one of the most remarkable industrial stories in China right now, for reasons that have almost nothing to do with AI.

CXMT is winning, spectacularly, in conventional DRAM. One caveat has to come first, and it matters more than it looks: CXMT is private and pre-IPO, so none of its financials are audited public filings. Every figure below is a reported number, and the reporting traces back largely to a single analysis house. Hold them accordingly.

With that said, the reported figures are extraordinary. First quarter 2026 revenue of about 50.8 billion yuan, up more than 700 percent year on year. Net profit around 33 billion. Margins near 70 percent, which would place it in the same band as SK Hynix. A first annual profit in 2025. SemiAnalysis expects its wafer capacity to approach Micron’s by the end of 2026, at roughly 350,000 wafers per month against Micron’s 385,000. Omdia has its global DRAM revenue share rising from under 4 percent in mid-2025 to around 8 percent. From a standing start a decade ago, this is one of the great catch-up performances in semiconductor history, and it should be described as one.

But read the margin honestly, because the company’s defenders and its critics both misread it. SemiAnalysis finds CXMT’s DDR5 cost per bit still runs more than 30 percent above the big three. The 70 percent margin is not a cost victory or a technology victory. It is a price. DRAM prices exploded, and CXMT is holding a lot of DRAM.

None of that gates AI compute.

Conventional DRAM is not HBM. HBM is DRAM stacked vertically and connected with through-silicon vias, and the stacking is the hard part. And in HBM, the picture is entirely different.

At the end of 2025, SemiAnalysis puts CXMT’s HBM allocation at roughly 5,000 wafers per month out of about 265,000 wafers of total capacity. Under two percent. The same analysis has CXMT still struggling to stabilize production of 8-high HBM3, with 12-high harder still, and industry estimates put the 8-high yield near 25 percent. Roughly 99 percent of CXMT’s 2025 bit shipments were conventional DDR and LPDDR.

None of that is a scandal. It is a rational allocation, and the reason is the most important sentence in this issue.

Bulk DRAM currently carries better margins than CXMT’s HBM, and it produces more than three times the bits from the same wafer area. For a company optimizing its own economics, every wafer moved to HBM is a wafer earning less. The commercially correct decision and the nationally required decision point in opposite directions, and the company is currently making the commercially correct one.

Which brings us to the document that settles the argument.

The use of proceeds is the company telling you what it will actually do

CXMT opens IPO subscription on July 16, raising 29.5 billion yuan on the STAR Market. It is the largest A-share listing of 2026 and the second-largest in STAR Market history, behind SMIC’s 53.2 billion yuan offering in 2020.

Read where the money goes. Roughly 69.5 percent of net proceeds funds wafer production lines and DRAM technology upgrades. The remaining 30.5 percent funds forward-looking DRAM research. There is no dedicated HBM project and no separate HBM line in the use of proceeds, an absence noted independently by SemiAnalysis, by the Seoul Economic Daily, and in Korean industry coverage of the filing.

This is worth stating carefully, because it is the load-bearing claim of the issue. The prospectus is not silent about HBM as a subject; a filing of that size discusses its technology roadmap and its risks. What it does not contain is money earmarked for it. And a use of proceeds is not a press release. It is a legal commitment about where four billion dollars will go, and it goes to commodity memory.

So the schedule the market is pricing, in which CXMT ramps HBM and relieves the constraint on Chinese AI compute, is not the schedule the company has just funded.

What this does to a valuation, and to our own prior work

At the indicated offering price, CXMT lists at a market capitalization near 295 billion yuan. That is not the number worth looking at. Chinese market commentary is discussing a theoretical post-listing valuation of two to three trillion yuan.

Sit with the upper end of that. The largest company on the entire A-share market today is China Construction Bank, at roughly 2.63 trillion yuan. The number being floated for a memory manufacturer would make it the most valuable listed company in China.

Read the two halves of the business separately, which is the discipline this publication keeps arriving at. The conventional DRAM franchise is genuine, profitable, and world-class, and it is enjoying a historic price cycle. On its own it justifies a large number. But it is a cyclical commodity business whose current margin comes from scarcity rather than from cost advantage, and it is being valued as though the cycle were a moat. The HBM business, meanwhile, is the strategic asset of enormous national importance, and it is the one the company has just declined to fund.

Those two things are being sold as one company, and at IPO they will be discounted at one rate.

The number to watch is not SMIC’s next node announcement, which will be covered everywhere. It is CXMT’s HBM wafer allocation and its 8-high stacking yield, which will be covered almost nowhere. That is what decides how many accelerators China can build once the warehouse is empty, and therefore how much compute the entire Chinese AI industry has to work with.

There is a consequence for our own argument too, and it needs stating.

Two issues ago we argued that the binding constraint on embodied AI is data rather than compute, on the grounds that GPU clusters capable of training vision-language-action models are commercially available. That was accurate then and it is accurate now. What this issue adds is a maturity date on it. Compute has been available in China because of a stockpile bought before a door closed. If CXMT’s HBM ramp slips and the warehouse empties, compute stops being the abundant input and becomes a contested one, and the data argument, while still correct on its own terms, stops being the only thing that matters.

A constraint that is currently non-binding is not the same as a constraint that has been solved. The data thesis holds. It now holds conditionally, and the condition has a name, and the name is a yield rate in Hefei.

Read the seam between the memory and the model

Silicon is the substrate. Models are the intelligence that runs on it. The seam between those two layers is not measured in nanometers, and it is not measured in floating-point operations either. It is measured in how fast you can move a parameter from memory into a multiplier, and how fast two chips can agree on a gradient.

That is why the chokepoint moved without anyone noticing. Everyone was watching the layer that gets photographed, the fab, the machine, the node. The constraint quietly relocated to the layer that gets soldered next to it.

China’s AI compute ceiling in 2027 will not be set by a lithography breakthrough. It will be set by two things happening inside one company in Hefei: whether it can push its DRAM core die to a competitive node without EUV, and whether it can stack eight of those dies reliably, at volume, in the back end. The first is a process problem. The second is a packaging problem. Neither is what the export controls were designed around, and neither is what the market is watching.

And there is a third thing, which is not an engineering problem at all. It is whether the company chooses to. This week, with four billion dollars in hand, it chose commodity DRAM.

Inside China’s Machine is research, not investment advice. The 910C teardown findings are Confirmed from TechInsights. The IPO size, subscription date, and the allocation of net proceeds are Confirmed from CXMT’s prospectus and the exchange filings as reported by Caixin, the South China Morning Post, and Reuters. The absence of a dedicated HBM line in the use of proceeds is reported independently by SemiAnalysis and by Korean industry coverage including the Seoul Economic Daily; readers should note that some generalist outlets have characterized the proceeds as funding an HBM push, and the two readings are in tension. Ascend production capacity, HBM stockpile volumes, die-bank depletion timing, CXMT’s HBM wafer allocation, and its DDR5 cost position are Estimated and attributed to SemiAnalysis, whose supply-chain figures are reconstructions rather than primary filings. The 8-high HBM3 yield figure is an industry estimate. The two to three trillion yuan figure is market commentary, not a forecast by this publication, and is cited as evidence of what is being priced rather than as a valuation. Reported memory bandwidth figures for the 910C conflict across sources and are treated accordingly in the text. Current as of July 12, 2026.

This publication has spent four issues pricing China’s humanoid stack. We valued a robot maker on its margin, tore down what that margin was funding, followed the profit into the supply chain, and argued that the data bottleneck resolves into a deployment question.

Every one of those arguments held the same variable fixed, and said so at the time. Unitree’s revenue growth was never in doubt, only its margin. The sensitivity grid put revenue on the x-axis precisely because we were not forecasting it. Leader Harmonious Drive’s roughly 570 times trailing earnings was priced on humanoid units that have not shipped. The data curve’s slope was a fleet size.

Holding a variable exogenous is a legitimate move, and it is honest as long as you eventually come back for it. This is coming back for it.

Four issues, one load-bearing assumption, taken as given each time. This is the issue that stops taking it as given.

The number the whole sector is priced on cannot survive being looked up

Start with what should be the easiest question in the industry. How many humanoid robots did China ship in 2025?

The answers, all from 2026 industry reports and all reported as fact:

• 14,400 units

• 18,000 units

• about 20,000 units

• 28,000 units

That is a spread of nearly two to one on the most basic figure in the sector. It gets worse. One set of reports puts global 2025 shipments at about 17,000 units, with China holding 84.7 percent of them. Multiply those together and you get 14,399, which is where the lowest number comes from. But other reports put China alone at 20,000 to 28,000, which is more than the entire world is supposed to have shipped.

The Chinese figure exceeds the global figure. Both cannot be true, and both are in circulation.

This is not a rounding disagreement. It is evidence that the industry has no shared definition of what it is counting. Does a unit ship when it leaves the factory, when it is invoiced, when it is delivered, when it is installed? Does a research platform sold to a university count the same as an industrial unit installed on a line? Does a robot leased to a rental platform count once, or every time it is redeployed? Different answers to those questions produce every number on that list, and each report is internally honest while the set of them is incoherent.

Anyone building a valuation on this number is building on sand and calling it granite. The honest move is not to pretend the sand is granite, and it is not to refuse to build. It is to go and look at what the sand is made of. So let us look at the one dataset that is actually granular enough to interrogate.

The demand curve has the shape of a fiscal year

China discloses public tenders. According to the 2025-2026 industry report published by the Humanoid Robot Scenario Application Alliance in March 2026, disclosed humanoid procurement awards in 2025 numbered 292 projects worth more than 1.81 billion yuan. An independent tally by a Chinese financial outlet, working from the same public tender disclosures through September 2025, arrived at a compatible picture: roughly 600 million yuan of awards covering more than 1,000 robot units across 20-plus manufacturers. Two separate counts, same source material, same shape.

That is not the whole market, and neither source claims it is. Procurement does not have to go through tender, so these awards understate total demand. But this is the part with its receipts published, and its structure is the most honest picture available of who is actually buying.

Look first at how the awards are sized.

• Projects under 1 million yuan: 54 percent of all awards

• Projects between 1 and 10 million yuan: 33 percent

• Projects over 10 million yuan: 10 in total

• Projects over 100 million yuan: 4 in total

Eighty-seven percent of every disclosed award in the country came in under 10 million yuan. The alliance report puts education and research institutions at 66 percent of the 292 projects by count, and notes without apparent discomfort that this price band matches the budgets of university teaching platforms and small-scale training systems. The independent tally reaches the same conclusion from the other direction: most tender projects originate with universities and research institutes, the robots are bought as research tools, and a typical order is one or two units.

These are not production lines buying labor. These are laboratories buying a robot.

Now look at when the awards landed. In January and February the count declined. Through the second quarter it oscillated between roughly 15 and 23 projects a month with no sustained rise. Then, from July through December, it held above 27 projects a month for six consecutive months, peaking in October at 42.

The industry’s own analysis of this pattern is blunter than anything I would have written. It concludes that local fiscal budgets and university budgets run on concentrated execution cycles, and that those cycles have a decisive influence on the tender rhythm. In other words: the curve rises in the second half because that is when the money has to be spent, not because that is when the demand appeared.

A demand curve does not have that shape. A budget-execution calendar does.

The largest orders come from the balance sheet of a local government

The four awards above 100 million yuan are the ones that lift the total, and it is worth asking who wrote them.

They came from local state-owned investment platforms, industrial investment companies, and municipal construction entities, in cities including Fangchenggang, Zigong, Jiujiang, Liuzhou, and Zhumadian. These are not the manufacturing centers a Western reader would expect to lead a robotics rollout. And the awards, by the industry’s own description, are not primarily for teaching or research. They fund robot industrial parks, intelligent demonstration zones, and regional application-scenario construction.

That is a specific and recognizable thing. It is the local government industrial-policy playbook, the same structure that built solar, batteries, and electric vehicles, applied to humanoids. A local platform company borrows against future land and tax revenue, buys the assets that qualify the city as a designated industrial base, and books the deployment.

I want to be careful here, because the lazy version of this observation is that the demand is fake. It is not fake. Solar was built this way and it now dominates the world. Batteries were built this way and CATL is real. Chinese industrial policy has a track record of converting subsidized demand into genuine global competitiveness, and betting against it has been a losing trade for two decades.

But subsidized volume and market volume have different risk properties, and they must be discounted differently. Market demand persists as long as the buyer earns a return. Policy demand persists as long as the fiscal cycle funds it. One is underwritten by a customer’s profit and loss, the other by a municipality’s balance sheet, and the local government financing vehicles doing this buying are operating under the most sustained deleveraging pressure they have faced in years.

The market is currently discounting both at the same rate. That is the mispricing.

What survives the audit, and what does not

Honesty requires saying clearly what this does not prove, because there is real industrial deployment underneath the noise and it is growing.

UBTech delivered more than 500 industrial humanoids in 2025 against production capacity above 1,000 units, brought its thousandth Walker S2 off the line, and reported an order book approaching 1.4 billion yuan, with Walker unit costs down about 25 percent from 2024. Note the distinction that most coverage blurs: an order book is a promise, and deliveries are the fact. UBTech’s own numbers show the promise running roughly three times ahead of the delivery, and the company posted a loss of 439 million yuan in the first half of 2025 while carrying that book. Galbot has agreed to deploy over 1,000 embodied robots with a manufacturing partner and its ecosystem. One research report puts 65 percent of Chinese humanoid shipments into factory settings. Whatever the tender data says about the tail, the head of this market contains real industrial customers buying real robots for real work.

So the finding is not that the demand is hollow. It is narrower and more useful than that:

The market has two demand curves stacked on top of each other and reports them as one number. Underneath is a genuine, growing, slow industrial curve, where a factory buys a robot because it pays for itself. On top of it is a fast, lumpy, fiscally-timed curve, where a laboratory or a municipal platform buys a robot because a budget exists. They have different growth rates, different durability, and different sensitivity to a downturn. Blending them into a single shipment figure, and then pricing a supply chain off that figure, is how a sector talks itself into a multiple.

The industry’s own segmentation says exactly this, for anyone willing to read it. The alliance report breaks 2025 demand into five scenarios. Education and research buys the most units but in ones and twos at low prices. Data collection is bought by government and state-asset bases, in batches of a hundred or more. Interactive service goes to corporate showrooms, museums, and tourism venues. Entertainment and performance goes to rental and event companies, at high volume and low unit price. And industrial logistics, the one scenario where a robot is bought to do work that pays for it, is described by the alliance itself as still small in scale because humanoids remain at a preliminary stage.

Read that list again. The scenario that justifies the valuations is the one the industry says is smallest.

There is a detail here that closes a loop from our last issue. UBTech’s final award of 2025, worth 59.6 million yuan, was to build a humanoid robot data-collection center in Huizhou’s Huiyang district. We argued last issue that embodied AI is bottlenecked by data and that data comes from deployment. Here is what that looks like in practice in China: a municipal government writing a check for a facility whose output is training data. The physical machine and the institutional machine are not adjacent. They are the same transaction.

And the industry knows. A venture investor quoted in the Chinese press observes that most domestic humanoids remain in demonstration and performance roles, running preset motions, and that this reveals the brain is not merely immature but arguably not yet formed. A widely read Chinese tech outlet notes that a machine costing hundreds of thousands of yuan is still valued by many buyers for what it calls emotional value rather than productivity. Chinese analysts have begun asking openly whether the sector’s order book contains related-party volume of the sort that once inflated the electric vehicle industry.

None of that appears in the shipment number. All of it is in the shipment number.

Two of our own conclusions have to be revised

Letting a variable move changes what depends on it. Two of the conclusions in earlier issues were conditional on the volume assumption, and now that the assumption has structure, those conclusions need recalibrating.

Issue three argued that Leader Harmonious Drive has a floor under it, because its harmonic reducers still sell into industrial robots and machine tools even if humanoids disappoint. That still holds. But the humanoid growth on top of that floor is now partly identified as fiscally-timed procurement, which means the incremental volume driving a roughly 570 times multiple has a different persistence profile than the industrial base underneath it. The floor is real. The thing being priced above the floor is softer than it looked.

Issue four argued that the data bottleneck resolves into a deployment function, because trajectories come from robots doing real work. That argument survives, but it acquires a filter. A robot in a university lab generates a trajectory. A robot in a demonstration zone generates a photograph. If 87 percent of disclosed awards are laboratory-scale, then a large share of China’s deployed fleet is not generating the industrial manipulation data that the fleet-size arithmetic assumed. The 100,000-robot fleet that collapses 212 years into two only does so if the robots are working. Robots that are exhibiting produce nothing an embodied model can learn from.

The deployment layer, in other words, is not just a destination or even a data source. It is the audit function for every layer above it. What actually happens on the floor is the only thing that can validate the numbers being priced at the top.

Read the receipts, not the forecast

The stack has a bottom and a top, and this publication has argued that the value lives in the seams between them. This issue argues something narrower and more uncomfortable. The seams are only as trustworthy as the measurement at the bottom, and the measurement at the bottom is currently a number that four different research houses cannot agree on, whose seasonality tracks a budget calendar, and whose largest line items are written by the investment arms of prefecture-level cities.

That is not a reason to be bearish on Chinese humanoids. It is a reason to price the two demand curves separately, and to treat any valuation built on a blended shipment figure as a valuation built on a number that has not yet been defined.

The forecasts for 2026 run from 62,500 units to 200,000. Before believing any of them, it is worth asking a question the forecasts never answer. When those robots ship, who signs the check, and what do they do with the robot on Monday morning.

Inside China’s Machine is research, not investment advice. Tender counts, award-size distribution, monthly seasonality, scenario segmentation, and the identity of the largest procuring entities are drawn from the 2025-2026 industry report published by the Humanoid Robot Scenario Application Alliance in March 2026, and are cross-checked against an independent tally of the same public tender disclosures. Both are compilations of public procurement records rather than primary filings, and are labeled accordingly. Shipment totals are Estimated and are reported here precisely because they conflict. Company delivery figures for UBTech and Galbot are Confirmed from company announcements. Views on demonstration-versus-productive use are Estimated and attributed to named industry participants. Forward volumes are Projected. Current as of July 11, 2026.