Reporter |
Editor | Wen Shuqi
Before Unitree made a name for itself, Wang Xingxing wasn’t exactly a big deal in the industry.
A few years back, at a domestic industry conference, the organizers had invited a few experts to talk tech. When it was Wang’s turn to speak, he dove deep — a bit too deep, and went on for a while. An official in the room cut him off, saying he was talking too much.
Back then, Unitree wasn’t well-known. Shortly after, the organizers tried to set up another meeting with him, but couldn’t get through. “The man who cut off Wang Xingxing” became a story that circulated internally for a long time.
A source close to Wang shared another tidbit: even when the company’s accounts were flush with cash, he’d still ask who was covering the hotel bill at conferences. Having lived through tight funding days, he’s always kept a tight grip on cash flow.
As Unitree’s fame grew, Wang’s cost-consciousness didn’t budge. It’s also been the company’s fastest track over the past decade: build, cut costs, sell robots.
On August 19, Unitree Technology went public on the Shanghai Stock Exchange’s STAR Market, becoming the first A-share humanoid robot company. Shares opened at 1,100 yuan, surging 629.44% — a jump of 949.20 yuan from the 150.80 yuan IPO price — giving it a market cap of 444.9 billion yuan. The diluted P/E ratio hit 219.23 times, far above the industry average of 38.56 times.
According to the prospectus, Unitree shipped over 5,500 humanoid robots in 2025, with humanoid revenue surpassing its quadruped business for the first time to become the top income source.
Unitree is the first full-machine maker in this humanoid wave to scale up and turn a profit. Its orders are reshaping upstream suppliers, and its pricing is setting the band for the whole industry. On this domestic supply chain, Unitree has become the de facto chain leader.
A mass-producible, profitable humanoid robot sold globally is the most direct product of China’s decades of manufacturing accumulation. But even this chain leader has things it can’t buy or build just yet.
Upstream Transformed by Orders
During Unitree’s subscription period, the A-share robot sector rallied, with component makers like Leader Harmonious Drive, Zhongda Leader, and Estun leading the gains.
The secondary market’s reaction points to a fact: what’s really being lifted isn’t just one company, but a batch of Chinese parts suppliers, all adjusting their product lines to follow robot orders.
One direct change: components are shifting from custom-made to standard parts.
Zhuoyu Technology makes robot drive components. A company spokesperson told NUPIAO that the industry’s first phase was all one-off custom deals. As custom orders piled up, their offerings have converged into three standardized series — planetary modules, harmonic modules, and linear modules — covering over 95% of robots on the market. They claim cooperation with half of China’s robot makers and have built automated lines with annual capacity for 200,000 joint modules and 500,000 motors.
At the same time, product lines are getting more granular, following robot structures. For instance, Zhuoyu’s smallest motor is now 14mm in diameter — about the size of a human fingertip — made specifically for the knuckles of dexterous hands, with a single finger typically using four such micro modules.
Moving from one-off deals to a standard parts catalog marks an industry’s transition from the “handcrafted” stage to the industrial stage.
Huang Jinping, founding partner and chairman of Rongyi Investment, one of Unitree’s backers, believes the robot supply chain has no bottlenecks — everything can be done domestically. Prices still have room to drop because volumes are small now; once volumes climb, costs will inevitably fall.
Another path is the horizontal migration of established parts makers. Audiwei, which has made sensors and actuators for 27 years, has moved from industrial scenarios into what Chairman Zhang Shuguang calls the “low-tier embodied intelligence” market — robot vacuums, pool cleaners, lawn mowers. Last year alone, they sold over 10 million sensors, hitting 110 million yuan in sales, all built from scratch in just three years.
He credits two advantages: first, a vertically integrated chain from materials to algorithms that allows scenario-based customization; second, Chinese firms are generally willing to take custom orders, while overseas makers only offer standard products.
But making something and keeping it running long-term are two different things. Zhang pointed to this year’s humanoid robot half-marathon as an example. In last year’s race, each robot had two handlers — one with a bucket of water, another with a remote. Without water, the robots would collapse. This year, no one was watering; instead, they made bigger joints and stuffed fans inside. “But that’s not a sustainable direction,” he said.
His take: robot joints face repeated motion and demand high sealing, and traditional fan bearings lose their lubricant quickly at high temperatures. The next-gen solution he’s working on is a built-in water tank in the body, using a micro-pump to deliver water to surface pores for evaporative cooling. If heat dissipation isn’t solved, robots can only do short demos — they can’t enter production lines that require continuous operation.
Joints and heat are just the two most visible issues among many pending problems. Chinese manufacturing has largely passed the “can we make it” test. What’s holding back is whether it can keep working after it’s made, plus a portion of high-end components still relying on imports.
Yu Yiran, managing director at CIC Consulting, told NUPIAO that while leaders like Unitree have achieved self-developed, self-made core components, the industry’s localization still shows significant structural divergence: robot hardware has achieved technological breakthroughs, but high-precision sensors, precision actuators, advanced embodied models, and motion control algorithms still lag.
Breaking it down by component, Yu’s assessment:
For sensors, six-axis force sensors still trail overseas players like ATI in high-precision, high-dynamic, miniaturized products. Tactile sensors are still in the validation phase, far from mass production.
For reducers, harmonic reducers have achieved domestic breakthroughs in routine scenarios, but high-end products still lag significantly behind Japan’s Harmonic Drive. High-end planetary roller screws also heavily depend on overseas suppliers like Switzerland’s GSA.
For motors, basic servos have high adoption rates, but high-torque-density frameless torque motors aren’t fully localized yet, and domestic products underperform overseas ones on multiple performance metrics.
For AI chips, the industry relies heavily on overseas computing platforms like Nvidia. Domestic chips lag significantly in computing power, and high computing costs have led to lower maturity in the “brain and cerebellum” of domestic humanoid robots. Yu noted this is closely tied to China’s semiconductor industry and won’t be solved overnight.
Despite these gaps, this chain can fully support a chain leader. And the one who truly converts this supply capability into price advantage is Unitree itself.
The Rise of a “Chain Leader”
A decade ago, when Wang Xingxing founded Unitree in Hangzhou, the company was so small that no supplier wanted to bother with it.
Huang Jinping decided to invest in Unitree in the second half of 2021 and has since participated in three rounds. His reflection: Unitree’s much-praised high self-development rate in its early days was half forced. The company was too small; others wouldn’t customize for it, so they had to take things apart and build piece by piece themselves.
“But on the other hand, they genuinely had that capability,” Huang told NUPIAO. Once key components became in-house production, cost control became something others couldn’t catch up with. Today, Unitree even sells its own LiDAR to outsiders.
Zheng Juncong, founding managing partner at Vertex Ventures China, was even earlier. In 2018, after Boston Dynamics’ robot dog Spot went viral online, he predicted China would see similar companies and had his team search everywhere — that’s how Unitree came into view. In October 2020, Vertex completed a Pre-A+ round as the sole institutional investor, followed by two more investments.
But at the time, Unitree had only three robot dog models, with sales in the dozens, mostly to research institutions. The unclear application scenarios scared off many investors: what practical use could quadruped robots have beyond research? Even at internal IC meetings, some questioned, “No clear applications — it won’t scale.”
What made Zheng commit was a detail measured in kilograms.
Unitree’s robot dog weighed 12 kilograms — half the weight of the Boston Dynamics product it benchmarked against. Why did it have to be half? Wang explained on the spot: a lighter body means better maneuverability, dramatically lower energy consumption, and cost savings. More importantly, he could articulate exactly how to achieve that — which core components needed customization and where to find partners.
“These details were crucial. They showed he truly understood the technology and knew how to execute, not just paint a rosy picture,” Zheng said.
That cost-obsessive capability ultimately showed up in pricing.
Unitree’s average humanoid robot selling price: 593,400 yuan per unit in 2023, down to 260,400 yuan in 2024, and 166,400 yuan in 2025 — a drop of over 400,000 yuan in two years. Cost reduction continues.
But here’s the twist: as prices fell, Unitree’s gross margins went up, not down.
From 2023 to 2025, its main business gross margin rose from 44.22% to 60.13%, exceeding the industry average since 2024.
Cutting prices while raising margins — that’s uncommon in hardware.
Typically, a full-machine maker lowers prices by sacrificing its own profit or passing pressure upstream. Prices go down, and gross margins follow. Unitree cut its own manufacturing costs instead — what it gave up in pricing was made back on the cost side.
Unitree’s prospectus attributes this to its full-stack self-development model for robot bodies and core components: motor drives, joint modules, body structure, and motion control systems are all in-house, keeping hardware costs low. Plus, as production and sales scale up, process improvements, procurement bargaining, and cost structure optimization all play a role.
The result: Unitree posted 1.699 billion yuan in revenue for 2025, with non-GAAP net profit of 591 million yuan and a non-GAAP net margin of 34.77%. Among the four comparable companies listed in the prospectus — UBTech, Dobot, and Leju Robotics — all posted losses in the same period, while CloudMinds posted non-GAAP net profit of 15.12 million yuan.
In other words, in this track, Unitree is currently the only one running both scale and profit at the same time.
The Missing “Brain”
From motors and reducers to LiDAR, Unitree’s parts are either self-made or available on this mature supply chain. Except for one thing.
In Unitree’s prospectus, there’s an unusual self-assessment.
From 2023 to 2025, Unitree’s R&D expenses were 49.95 million yuan, 70.02 million yuan, and 145 million yuan respectively — growing fast, but relatively steady compared to some industry players in terms of scale and revenue share.
The prospectus states that early R&D focused on body structure and motion control — the body and cerebellum. It was only from 2024 that it gradually increased investment in embodied large models, the brain. Given unclear global technology paths and data scenarios, it hasn’t conducted large-scale real data collection or factory deployment training during the reporting period.
A company that’s shipped the most humanoid bodies globally, admitting in an IPO filing that it underinvested in the brain — that’s an uncommon statement.
The numbers back it up. In 2025, Unitree’s R&D spending was 8.53% of revenue, down from 31.39% in 2023. For comparison, UBTech’s R&D ratio was 35.1% in mid-2025.
In the IPO fundraising plan, of the 4.202 billion yuan Unitree aims to raise, 2.022 billion yuan — nearly half — goes to the embodied intelligence model R&D project. That’s higher than the 1.11 billion yuan for robot body R&D and 624 million yuan for the manufacturing base construction project.
A company that went public on manufacturing strength is putting its largest fundraising chunk into the part it’s least good at and can’t simply buy.
Yu Yiran told NUPIAO that the assessment of China’s embodied intelligence industry as “strong body, weak brain” still holds at this point.
Relying on China’s mature manufacturing supply chain, domestic robot makers like Unitree are leading globally in joint motors, motion control algorithms, and whole-machine cost control. But on the model front, Physical Intelligence and Generalist, which focus on general-purpose foundation models, still lead.
“Domestic companies’ models perform well in specific scenarios, but once they enter high-frequency, unstructured complex environments, their capabilities drop off dramatically,” Yu said.
The real bottleneck is data. Yu told NUPIAO that collecting data from the physical world is difficult, costly, and inefficient — it’s hard to form an effective data flywheel before commercial deployment.
This cost is quantifiable on the ground. A company executive working with Unitree told NUPIAO that a major tech firm once explored this internally, pouring hundreds of millions of yuan in, with data collection alone costing several million. Online videos look usable, but they lack physical properties, fluid dynamics, and friction information — useless for actual operations. You have to collect it yourself.
In their view, robots need three capabilities: understanding, prediction, and action intervention with feedback. 2D understanding is currently achievable, but 3D still has depth-space issues to resolve. Applied to scenarios, robots for short-cycle, shrinkable processes in factories can already do decent work — but open-ended tasks are a no-go.
Wang Feili, UBS Securities’ China industrial sector analyst, made a similar comparison in an interview with NUPIAO and other media this May. She believes general-purpose humanoid robots are harder on the brain than autonomous driving. If autonomous driving is graded L1 to L5 and humanoids by generality, humanoid L5 would be harder than autonomous driving L5. Currently, humanoid intelligence is roughly equivalent to autonomous driving L3 — capable of one or two tasks in controlled environments.
Her benchmark: when autonomous driving achieves L5 and is widely applied, humanoid generality might not be far behind.
Of course, this isn’t just Unitree’s shortcoming.
Domestic companies lead on the manufacturing side but haven’t built equivalent capability in the model layer. The AI chips Yu mentioned are part of it — the gap between domestic computing platforms and overseas ones, plus high computing costs, are slowing the maturity of the “brain and cerebellum,” directly tied to China’s semiconductor progress.
Going a layer deeper, embodied intelligence models globally still haven’t reached “ready to work” levels. Physical Intelligence and Generalist are also grappling with generalization.
Components can be self-developed or bought upstream. But the brain has neither path.
At the August 7 roadshow, Wang positioned the industry: embodied intelligence is still in its early stages overall, similar to the early days of home computers. The generalization capability for devices to adapt to unfamiliar environments still needs industry-wide improvement.
Who’s Paying for Robots?
The body’s built, the brain’s not quite there — so who’s actually buying these robots now?
Market research firm Smart Analytics Global (SAG) latest data shows global humanoid robot shipments hit about 19,100 units in the first half of this year — more than triple the 5,100 units in the same period last year. Chinese makers contributed over 97%, with the top four all being Chinese companies. Agibot topped the list for the first time with 8,400 units and 44% share, Unitree was second with 5,900 units and 31%, followed by UBTech and Leju.
More telling than rankings is where the money comes from. In 2025, Unitree’s top five customers combined accounted for just 12.08% of revenue. The largest single customer was 3.11%, and the second-largest was JD.com with 48.87 million yuan in sales. From 2023 to 2025, this ratio hovered around 12% — no industry customer has yet scaled up.
According to Unitree’s response to the STAR Market’s inquiry letter, in the first three quarters of 2025, humanoid robot revenue breakdown: research and education accounted for 73.6%, commercial consumption (including malls, exhibitions, performances) 17.39%, and industry applications 9.01%. As previously compiled by Caijing, the main buyers of humanoid robots remain government and state-owned entities.
Wang Feili noted that among last year’s top five suppliers by shipments, most downstream deployments are still research, data collection centers, and entertainment — true industrial applications remain rare.
Yu Yiran gave a more direct estimate: in 2026, among global humanoid and quadruped robots, only about 20% will actually match production needs, operate stably, and achieve commercial deployment.
Upstream players are even more cautious. Zhang Shuguang told NUPIAO that early humanoid robot companies are basically research-oriented. “Right now, volumes haven’t picked up, and they haven’t figured out how to use them,” so Audiwei plans to wait until the commercial stage before entering.
However, the value of the body as a development platform is already paying off. Zheng Juncong told NUPIAO that many US startups working on brain training are using Unitree’s robots because they can’t buy Tesla or Figure products.
“It’s like when PCs appeared — Apple had PCs, then IBM PCs came along. Once everyone had PCs, they could develop software,” he said. Unitree’s position now resembles that stage.
An industry insider from one of Unitree’s core suppliers told NUPIAO that whether Unitree can build a moat comes down to one question: for people developing brains, is Unitree’s body the easiest to use and the most universal option?
They noted that companies like Tesla build both body and brain, which is obviously better from a software-hardware synergy perspective. But universities and small startups are using Unitree’s body to develop brains — that’s Unitree’s opportunity.
In other words, most of the robots Unitree sells today are bought for R&D, not for actual work. It’s hard to call that true deployment.
Unitree has spent a decade proving one thing: Chinese manufacturing can build a humanoid robot, build it cheap, sell the most globally, and bring an entire supply chain along with it. That chain can now support a chain leader.
But what it hasn’t yet proven is that this robot can actually work. That’s no longer within manufacturing capability — it depends on models, data, and the pace of global AI progress. That’s also why Unitree is putting its largest fundraising chunk into the brain.
(NUPIAO reporters Cheng Lu and Li Shuyao also contributed to this article)
