Unitree Robotics Just Went Public. The Backflips Were Never the Point.

A $900 million IPO, a tenfold collapse in robot costs, and the race to build the hardware platform for physical AI.

Unitree Robotics went public in Shanghai today and briefly became a RMB 400 billion company.

The six-year journey from an unremarkable office in Hangzhou to the STAR Market produced the sort of numbers people remember: roughly RMB 6.1 billion, or $904 million, raised in the IPO; retail demand reportedly more than 8,000 times the shares available; a 629% intraday jump; and a closing valuation of about RMB 342 billion, or $51 billion. The stock finished its first session 460% above the offer price.

Those figures are extraordinary enough on their own. They look even more extraordinary beside the operating business beneath them. Unitree generated approximately RMB 1.7 billion of revenue in 2025, with a gross margin of about 60%. It shipped more than 5,500 humanoid robots during the year, and by July 2026 cumulative deliveries of its bipedal machines had reached roughly 18,000. More than 40% of revenue now comes from outside China.

At the closing bell, the market was valuing Unitree at roughly 200 times its 2025 sales. At the intraday peak, the multiple was higher still.

You can read that as evidence of a market in a speculative mood. You would not be wrong. The offer itself was already priced at more than 200 times earnings, far above the average for comparable manufacturers. Unitree’s machines can dance, box, run and backflip; apparently its shares can do the same.

But focusing only on the first-day multiple misses the more consequential number in the story.

That number is not RMB 400 billion.

It is the difference between $150,000 and $15,000.

This is the number that explains why Unitree exists, why an early investor was willing to back an unusually young founder before the market was obvious, why the comparison with Elon Musk is more than founder mythology, and why people have begun—prematurely, but not absurdly—to ask whether Unitree could become the NVIDIA of robotics.

The backflips made Unitree famous. The cost curve made it important.

A Robot Dog Bows. The Investment Committee Asks: “Then What?”

In August 2020, investor Zheng Juncong visited Wang Xingxing in a modest office near Xixi in Hangzhou. Unitree was still a small company known mainly among robotics researchers. Wang put a quadruped on the floor. It stood up and bowed.

The demonstration was charming. The investment case was not yet obvious.

Zheng asked whether the robot was controlled by AI. Wang’s answer was disarmingly plain: no, it was his own control algorithm. Then came the question that mattered more. What did it cost?

Wang said Unitree could sell it for roughly $10,000 to $20,000—and still make a gross profit.

Zheng had recently examined a Boston Dynamics robot. The comparison he carried into the meeting was roughly $150,000 for a configured, deployment-ready system. To be precise, Spot’s published base price when commercial sales opened in 2020 was $74,500. Once buyers added sensors, payloads, software, support and the requirements of a serious enterprise deployment, the economics could move well into six figures. Wang was not claiming a modest discount. He was proposing a different cost structure.

That distinction is essential.

A product that costs 10% less competes for an existing budget. A product that costs 90% less can move into a different budget, reach a different buyer and create a different market.

During the meeting, Wang did not lead with a total-addressable-market slide. He did not predict that humanoids would soon transform every warehouse, factory and home. He opened the machine.

He explained why Unitree’s motors were lighter, which parts needed to be designed in-house, which could be sourced from the manufacturing ecosystem around the Yangtze River Delta, and how mechanical structure, control software and assembly had to be engineered as one system. He spoke about torque, weight, sourcing and manufacturability. He did not talk much about changing the world.

That presentation eventually reached the investment committee at Vertex Ventures China. The resistance was immediate and sensible. The robot dog was cute. It could move well. But then what?

Where would it sell? China’s quadruped industry was still largely in technical validation. The industrial use cases were immature. Wang was a founder born in 1990 without the pedigree that made traditional venture committees comfortable. Could he recruit and manage a larger organization? Could a small technical team manufacture at scale? Was this a company or an impressive machine looking for a market?

Those were the right questions. Early-stage investing becomes interesting when the right questions do not yet have complete answers.

Zheng’s case was not that Unitree had already found the killer application. It was that exceptional engineering had changed the cost curve enough for applications to become discoverable. The machine in front of the committee was not merely a cheaper imitation of a Western product. Unitree had integrated motors, controllers, algorithms, mechanical design and production in a way that made each generation a foundation for the next. China’s supply chain provided leverage, but the company still had to know what to build, what to buy and how to make the pieces work together.

The investors examined the self-developed content of the machine, motor and joint performance, the repeatability of the demos, whether overseas universities were placing real paid orders, and whether the team could actually build what it claimed. The evidence was encouraging, although far from conclusive.

Zheng later described the decision as 70% diligence and 30% intuition. The intuition was the part that allowed an investor to act before the category had a map.

Vertex invested in Unitree’s Pre-A+ round in October 2020 and followed with two more investments. By the time the company’s humanoids appeared on China’s Spring Festival Gala, capital was no longer difficult to find. The hard decision had been made years earlier, when the prevailing response to a robot dog was still: what is it actually for?

There is a tendency to rewrite successful investments as acts of clairvoyance. That gives both founder and investor too much magic and too little method. In 2020, nobody needed to know precisely how many humanoids would be working in factories in 2030. They needed to recognize that a tenfold change in cost could alter who was allowed to participate in finding the answer.

That is a more modest insight than predicting the future. It is also more useful.

What Changes Between $150,000 and $15,000?

At $150,000, a robot is an institutional capital project.

An operating team needs a defined task. Finance needs a return-on-investment model. Procurement needs competing bids. Safety, legal and IT may need to approve the deployment. A senior executive has to sponsor the experiment, and failure will be visible. The machine is purchased only after the use case has been substantially proven.

At $15,000, the sequence can reverse.

A university laboratory, an AI startup, a corporate R&D team or a well-funded developer can buy the machine before knowing exactly what it will become. The robot shifts from a boardroom question to an engineering question. Instead of purchasing it to execute an approved business case, the buyer purchases it to discover one.

Price, in other words, is not merely a commercial variable. It is an access variable. It determines how many people can run an experiment, how diverse those experiments can be and how cheaply the ecosystem can learn from failure.

This is why a disclosure in Unitree’s prospectus deserves a less conventional reading. In the first nine months of 2025, 73.6% of the company’s humanoid revenue came from research and education. Commercial and consumer applications represented 17.39%. Direct industrial applications accounted for only 9.01%.

A skeptical analyst sees an industry selling expensive laboratory props to researchers who are themselves funded by the same enthusiasm surrounding embodied AI. That interpretation cannot be dismissed. Humanoids still struggle with reliability, dexterity, battery life, safety and the unpredictable disorder of ordinary environments. A successful demonstration is not the same as a useful eight-hour shift. Factories do not buy viral clips. They buy throughput, uptime and payback periods.

But 73.6% research exposure is not necessarily a confession that the market has failed. It may describe the stage before a platform’s uses become obvious.

The personal computer did not arrive with a perfectly defined enterprise role. Spreadsheets helped turn it into one. Graphics processors spent years serving video games before researchers found that parallel computation could accelerate a different kind of workload. Early drones appealed to enthusiasts before filmmakers, surveyors, farmers and emergency services built professional workflows around them.

The first customer for a general-purpose technology is often not the final customer. The first customer is the one willing to experiment before the payoff can be entered into a spreadsheet.

Unitree’s research-heavy mix therefore cuts both ways. It tells us humanoid demand remains immature. It also tells us the company has placed thousands of programmable bodies in the hands of the people most likely to develop new ways of using them.

The company’s quadruped business offers a preview of how that progression can work. Four-legged robots have moved beyond demonstrations into power-line and substation inspection, petrochemical facilities, firefighting, rescue, coal and steel environments. Unitree reports validation with major Chinese industrial groups including State Grid, China Southern Power Grid, PetroChina, Sinopec and Baowu. These deployments are narrower and less glamorous than a general-purpose humanoid, but that is precisely why they matter. A machine does not need to replicate a person to create value. It needs to perform a specific task with an acceptable combination of cost, reliability and safety.

Humanoids will face the same discipline. The form factor is attractive because the world has been built around the human body: stairs, doors, tools, shelves and production lines all assume human dimensions. But a human shape is not itself a business model. The robot must eventually deliver an economic result in an environment that is less controlled than a stage.

Lower cost does not solve that problem. It changes the number of attempts the world can afford.

This is the heart of Unitree’s strategy. The company did not wait in private for a single perfect application. It lowered the cost of the hardware, sold into laboratories and development teams across the world, and allowed a distributed network of customers to probe the frontier. A failed experiment on a $15,000 platform is tuition. A failed experiment on a $150,000 platform can end a program.

Cost also changes the development culture inside the manufacturer. When hardware is precious, teams protect it. When hardware is replaceable, they push it. Robots fall, motors burn out, gearboxes reveal weaknesses and controllers encounter edge cases that no simulation anticipated. The physical world is an unforgiving teacher, but lower-cost machines let more students attend the class.

That is why the occasional public fall is not the most interesting risk. A machine falling in a demonstration can be repaired. A cost structure that prevents anyone from testing the machine is harder to recover from.

Of course, cheap hardware can also become commodity hardware. If every competitor gains access to similar components and manufacturing capacity, price leadership may compress margins rather than build a durable moat. Unitree’s 60% gross margin in 2025 suggests it has not merely subsidized its way into demand. The harder question is whether that margin can persist as rivals scale and whether the installed base can be converted into something more defensible than units shipped.

That is where the argument moves from manufacturing to platform economics.

The Musk Parallel: An Invoice Is Not a Law of Physics

Elon Musk is invoked too casually in founder stories. The comparison often means little more than ambition, difficult engineering and a willingness to ignore conventional advice. Unitree’s case allows for a narrower and more useful parallel.

Musk has described first-principles reasoning as breaking a problem down to the things known to be fundamentally true, then building upward from there. In explaining SpaceX, he used the cost of the raw materials in a rocket as a thought experiment. If the material inputs represented only a small fraction of the market price, then the prevailing price was not dictated by physics. It reflected a production system, a supply chain, a set of volumes and a collection of inherited choices.

That did not mean rockets were secretly easy. It meant the industry’s invoice should not be confused with a natural law.

SpaceX brought critical capabilities in-house, redesigned around manufacturability, accepted iteration as part of development and pursued reusability to attack the cost of access to orbit. Its achievement was not simply negotiating harder with suppliers. It changed the system that made the suppliers’ prices possible.

Wang’s question about legged robots belongs to the same family.

Why was a capable machine so expensive? Which costs came from the actual demands of balance, torque, sensing and control? Which came from low production volume? Which reflected specialized suppliers, organizational handoffs or designs optimized for technical spectacle rather than repeatable manufacturing? What happened if the company designed the motor, controller, structure, algorithm and production process together?

The answer was not a single breakthrough. It was thousands of engineering decisions that compounded into a different bill of materials.

This is what “first principles” looks like when stripped of motivational-poster language. It is often less cinematic than invention. A founder questions a specification. An engineer removes weight from a motor. A team integrates two systems that incumbents purchase separately. A production line exposes a tolerance problem. The next design corrects it. Unit volume rises, the learning curve improves and fixed development costs spread over more machines.

The elegance lies in the accumulation.

There are also important differences between SpaceX and Unitree. A rocket is produced in low volumes, operates in an extreme environment and carries consequences that make failure extraordinarily expensive. A robot has the potential to become a mass-produced endpoint, owned by thousands of organizations and modified by outside developers. SpaceX’s vertical integration was partly about controlling a mission-critical system. Unitree’s integration is also about making a complicated machine cheap and repeatable enough to distribute.

The shared method is not “build everything yourself.” Total vertical integration can become an expensive religion. The method is to identify which components determine performance, cost and iteration speed—and refuse to outsource those simply because the industry traditionally does.

Unitree appears to have made that choice around motors, drives, controllers, core algorithms and system design, while using China’s dense electronics and manufacturing base where external specialization is genuinely advantageous. That balance matters. Supply-chain proximity is not a magic wand; many companies have access to the same geography. The advantage comes from knowing exactly what the product requires and turning that knowledge into specifications, supplier relationships and production feedback faster than competitors can.

China is unusually fertile ground for this approach. The country combines large engineering talent pools, deep component ecosystems, aggressive suppliers, rapid prototyping and a domestic market willing to test hardware in messy real environments. A robot may begin in a university laboratory, appear at an exhibition, entertain shoppers, inspect a substation and then be redesigned for a task nobody originally placed in the market-size deck.

That progression can look chaotic. It is also how an industry learns.

The global robotics debate often starts at the top of the technology stack: foundation models, reasoning, vision-language-action systems and the possibility of general intelligence embodied in a machine. Those questions are important. But physical AI has a second bottleneck that software-native analysis tends to underestimate: industrialization.

An intelligent model cannot create economic value in the physical world if the body is too expensive, fragile or difficult to manufacture. Intelligence needs an affordable endpoint. Unitree’s wager is that the company able to industrialize the body will have strategic leverage over the intelligence that eventually inhabits it.

That is the bridge to NVIDIA.

Could Unitree Become the NVIDIA of Robotics?

“The NVIDIA of robotics” is the kind of phrase that spreads quickly because it compresses an entire investment thesis into four words. It is also dangerous because it can substitute analogy for analysis.

NVIDIA did not become one of the world’s most valuable companies merely by selling a large number of fast chips. Its durable advantage grew from the interaction of architecture, software and developers. CUDA gave programmers a stable way to use NVIDIA hardware for general-purpose computing. Libraries, tools, documentation and years of backward compatibility turned the chip into a platform. Researchers built on it. Universities taught it. Companies designed systems around it. Each new user made the ecosystem more useful, and every application increased the cost of switching away.

Today, NVIDIA describes a CUDA-X community of more than six million developers and nearly 6,000 applications. The moat is not one GPU benchmark. It is the accumulated work of millions of people who have chosen NVIDIA as the environment in which to build.

For Unitree to earn the comparison, it would need to do something similar in the physical world.

The company already has several necessary ingredients.

First, it has an expanding installed base. More than 5,500 humanoids shipped in 2025 and roughly 18,000 cumulative bipedal deliveries by July 2026 are significant in an industry that has historically measured many programs in dozens or hundreds. The quadruped fleet expands the developer audience further. Unitree’s low entry prices, including consumer-oriented quadrupeds priced far below industrial systems, make the hardware accessible to a much broader population.

Second, it is exposing tools around that hardware. Unitree provides SDKs, ROS and ROS2 interfaces, simulation support, reinforcement-learning environments, model and dataset releases, teleoperation resources, and compatibility with widely used frameworks such as MuJoCo and LeRobot. Its open-source work includes vision-language-action research and manipulation data. These are not yet a robotic equivalent of CUDA, but they indicate that Unitree understands a body without a development environment is only a product.

Third, the company controls enough of the hardware stack to coordinate improvements across components. A model trained in simulation must survive contact with motors, joints, sensors, latency, heat, battery limits and imperfect floors. Integration between the digital and mechanical layers determines whether an impressive policy becomes a dependable machine. A manufacturer that owns the relevant interfaces can shorten the loop between data, control and redesign.

Fourth, Unitree is international earlier than many Chinese hardware companies. With more than 40% of revenue generated overseas, its developer base is not confined to one country or one industrial system. That matters because platforms become more powerful when third parties invent applications the platform owner did not anticipate.

The optimistic version of the thesis is straightforward.

Unitree sells affordable bodies into labs, startups and industrial teams. Those users generate code, workflows, data and trained behaviors. More applications make Unitree hardware more useful. Greater demand increases production volume, lowering unit costs and funding better machines. Better machines attract more developers. The cycle repeats until “build it on Unitree” becomes a default choice for physical-AI experimentation.

If that happens, hardware revenue is only the opening act. The company could capture value through development software, model deployment, simulation, fleet management, maintenance, application marketplaces, data services and certified third-party modules. Hardware would create distribution; software and services would deepen the economics.

That is the bull case. It is plausible. It is not yet proven.

CUDA works because it offers more than an installed base. It offers continuity. Code written years ago often remains valuable on newer NVIDIA hardware. Developers trust the roadmap. Libraries solve difficult problems. The ecosystem contains complementary businesses, not merely customers. Most importantly, NVIDIA captures substantial economic value when demand for accelerated computing rises.

Unitree still has to establish each of those properties.

Can developers move code and learned behaviors reliably from one generation of robot to the next? Will third-party companies build profitable products on top of Unitree machines? Does Unitree own the interface that developers most value, or will that interface belong to a model company, an operating-system provider or an industrial integrator? Will customer-generated data flow back to Unitree, given privacy, security and intellectual-property constraints? Can the company create recurring software and service revenue, or will buyers treat the robot as interchangeable hardware?

These questions are not footnotes. They determine whether Unitree becomes a platform or an excellent manufacturer in a market that eventually commoditizes.

There is also a strategic tension in openness. Easy access to SDKs, standard frameworks and open-source tools can accelerate adoption. It can also make switching easier if competitors support the same abstractions. NVIDIA’s moat emerged from an ecosystem that was open enough to attract developers but sufficiently tied to its architecture to preserve economic advantage. Unitree will need to find its own version of that balance.

Nor is the relevant competition limited to other robot manufacturers. The platform owner could emerge at several layers. A foundation-model company may provide the intelligence that works across many bodies. A simulation provider may own the development workflow. An industrial automation group may control customer integration. A component maker may standardize actuators and controls. Or a robot manufacturer may integrate enough of the stack to become the default.

The phrase “NVIDIA of robotics” therefore should be treated as a question, not a conclusion.

Unitree has proved that it can build capable machines at prices that expand access. It has established meaningful volume, attractive gross margins and a global research footprint. It has begun assembling developer tools around the fleet. These achievements buy the company a credible option on platform status.

An option is valuable. It is not the same as ownership.

What the Next Five Years Need to Prove

The public market has already assigned Unitree an enormous portion of the future. The company now has to turn that expectation into operating evidence.

The first test is industrial utility. Research laboratories can train machines and discover behaviors, but industrial buyers pay for reliable work. Watch whether humanoid revenue moves from education toward production environments; whether deployments expand beyond pilots; and whether customers reorder after measuring uptime, safety and economics. The most revealing metric may not be units shipped, but productive hours per unit after installation.

The second test is the quality of the ecosystem. Downloads and GitHub interest are useful signals, but applications matter more. Are outside developers releasing products that customers will pay for? Are universities teaching on Unitree hardware because it is the default, or merely because it is inexpensive? Is there a growing market of sensors, grippers, tools, software and services that becomes more valuable as the fleet grows?

The third test is recurring revenue. A robot sold once may generate an attractive gross profit. A platform compounds when software, support, orchestration, upgrades and services continue throughout the machine’s life. Unitree’s financial disclosures today still describe a hardware-led company. The NVIDIA comparison becomes materially stronger only when the economics begin to reflect the software layer.

The fourth test is whether cost leadership survives competition. China’s robotics field is crowded, capitalized and moving quickly. A tenfold advantage can attract imitators as effectively as it attracts customers. Unitree will have to keep reducing cost while improving durability and performance—without sacrificing the margins that made its 2025 results so unusual.

The fifth test is geopolitical resilience. Unitree is already international, but international exposure carries regulatory risk. Restrictions affecting Chinese connected devices, components, data flows or access to the US market could fragment the developer ecosystem. A global platform must navigate security concerns and local rules without losing the network effects that make global scale valuable.

None of this diminishes what happened today. It clarifies it.

Unitree’s IPO is not proof that general-purpose humanoids have arrived. A first-day valuation is not the same thing as a defensible platform. A dancing robot on national television does not establish industrial ROI, and a research-heavy customer base does not guarantee the next application layer will appear.

But the company has done something more substantial than produce a good demonstration.

Wang Xingxing did not begin by proving that the world needed millions of robots. He began by proving that a capable robot did not need to cost what the world assumed it cost. Then he placed enough machines in enough hands for the world to start looking for reasons to use them.

That is first-principles thinking in its most commercially powerful form. Do not begin with the market forecast. Begin with the constraint. Ask whether it is imposed by physics or inherited from an industry’s habits. If it is only a habit, redesign the system until the economics change.

There is something appropriately youthful in that approach—not because young founders are automatically wiser or braver, but because they have had less time to internalize which assumptions respectable people have agreed not to question. Wang was not burdened by the belief that advanced legged robots naturally belonged in a six-figure procurement category. So he opened the machine, examined the parts and asked a simpler question: why must it be this expensive?

The answer created a company worth billions. Whether it creates the dominant platform for physical AI remains undecided.

That uncertainty is not a weakness in the story. It is the reason the next chapter matters.

Will the most important robotics company be the one that builds the body, the one that trains the brain, or the one that owns the interface between them?

Will low-cost hardware create developer lock-in—or accelerate commoditization?

Are research laboratories a weak end market, or the seedbed from which the next industrial software ecosystem will grow?

And if a capable robot really does become a $15,000 development platform, what problem would you ask it to solve—not after the business case is proven, but in order to discover one?

Unitree has not answered those questions for us. It has made answering them much cheaper.

That may turn out to be the more important achievement.

---

Sources and notes

IPO and first-day trading figures are drawn from reporting by the Financial Times, Reuters, Associated Press, and The Asset. Financial, customer-mix and operating data are based on Unitree’s STAR Market prospectus and contemporaneous public reporting. Unitree’s developer resources can be reviewed on its open-source page. Boston Dynamics’ 2020 Spot pricing was reported by TechCrunch and The Boston Globe. The SpaceX comparison draws on Musk’s explanation of first-principles reasoning in Wired. NVIDIA’s published CUDA ecosystem figures are available in its company materials. The early investment account is based on a public recollection by Vertex Ventures China founding managing partner Zheng Juncong, supplemented by Unitree’s public filings.