"How much can one actually make from winning a single IPO allotment of Unitree? I saw 160,000 RMB this morning, then it changed to 200,000, and now it's 350,000." As soon as the stock market opened this week, someone in the media group chat expressed this confusion. Everyone assumes that winning an IPO allotment is equivalent to getting a windfall for free.
At the same time, others are also celebrating massive payouts—the employees of embodied AI companies.
A friend told me that the headhunting market has already seen algorithm engineer salaries reaching the tens of millions level. Previously, they had recruited for similar positions at embodied AI companies, but the maximum was two to three million. A ten-million-yuan annual salary is unprecedented.
The simultaneous emergence of high P/E ratios and high salaries reveals a contradiction easily overlooked amidst the monetary frenzy: the prospectus shows that Unitree Robotics has a P/E ratio of 219 times. This means that, assuming no profit growth, it would take the company 219 years to earn back its current market capitalization.
The embodied AI companies queuing up for IPO right after, such as AGIBOT, Leju Robotics, and DeepRobotics, will not have low P/E ratios either. Meanwhile, the high-salary talent wars in the industry mean that R&D costs will further increase, which in turn erodes profits. So, when can the high P/E be justified?
Some say that the extremely high valuations of early-stage tech companies are like using a kindergarten child's report card to predict their performance in college.
Our main business is interpreting technology. The technological evolution of humanoid robots from infancy to maturity is perhaps the most certain piece of information in this entire story. Therefore, taking advantage of Unitree's IPO, let's discuss from a technological logic perspective: how many "foldings" does a robot need to go through to turn the future into reality?
The First Folding: Overseas Benchmarking Amid the Global Tech Race
The high valuation of Unitree Robotics is not an isolated case. In fact, against the broader backdrop of the global tech race, as long as there are domestic hard-tech companies with U.S. stock market counterparts, the market is willing to offer extremely high premiums.
For instance, the valuation of CXMT (ChangXin Memory Technologies) is three to five times that of the three overseas memory giants. Cambricon's valuation far exceeds its current earnings growth rate, and Hua Hong Semiconductor boasts a P/E ratio of 800 times. Wrapped within this is a logic of "domestic mega-market + strategic industry security," which further amplifies the value.
Humanoid robots share this same logic.
First, embodied AI has been included as a strategic emerging industry with key cultivation efforts. This year, the outlines of the 15th Five-Year Plan for all 31 provinces nationwide have been fully released, among which 25 provinces mentioned "embodied AI." Especially after the IPO of CXMT, which has spent a decade perfecting its technology in Hefei, the probability of government-guided funds entering the embodied AI industry has significantly increased.
Second, domestic robots are a crucial force for future industrial employment and national defense, possessing a strong logic of massive domestic demand. From another perspective, what the U.S. does not want us to develop is often what we need to focus on. Recently, the Trump administration issued a ban on the import of Chinese humanoid robots, which also indicates that this is a focal point of future competition.
Therefore, granting higher valuations to humanoid robot companies is logically sound in the long term.
However, long-term rationality does not mean that current valuations are reasonable or will definitely be absorbed by the market.
The most typical example is Boston Dynamics. Over a decade ago, Atlas's backflips and Spot navigating through ruins almost defined our imagination of robots. But due to the inability to mass-produce and the mystery surrounding commercialization, Boston Dynamics was ultimately acquired by South Korean capital.
Therefore, when companies like Unitree seek global benchmarks, the most likely reference points are perhaps Tesla and Figure AI. These two U.S. companies represent the two most valuable anchors for current humanoid robots: intelligence level and large-scale deployment.
Intelligence raises the upper limit of capabilities. Tesla's Optimus, relying on its self-developed end-to-end neural network, integrates the Vision-Language-Action (VLA) model into humanoid robots, making it one of the few companies that have genuinely implemented VLA in practice. Figure AI's self-developed Helix 02 model has received widespread industry feedback for its strong collaborative operation capabilities, enabling it to handle generalized tasks. Coupled with the advantages of mass production and large-scale deployment, both Tesla and Figure AI have now sent their robots into factories. Figure 02 is participating in the production of over 30,000 BMW X3s at BMW's U.S. plant; Optimus is also sorting batteries and moving parts within Tesla's own factories, beginning its commercialization run.
So, do domestic embodied AI manufacturers possess these two characteristics? They have both, but they are not yet mature enough.
Domestic enterprises are going all out to follow technological paths like VLA, but technological barriers objectively exist. Currently, giants like Tesla are no longer willing to disclose details of core algorithms such as VLA at launch events to prevent Chinese manufacturers from learning and reverse-engineering them. This means domestic companies must forge their own technological path.
Objectively speaking, benchmarked against Tesla and Figure AI, domestic embodied AI manufacturers have gaps but have not substantially fallen behind. The purpose of embodied AI companies offering annual salaries of tens of millions to algorithm engineers might lie precisely here: they need to spend heavily to build a domestic embodied AI brain comparable to Tesla's. As for how long it will take to emerge, it remains unknown.
Furthermore, while companies like Unitree support the mass production of humanoid robots, how to get these robots onto factory production lines is the truly difficult problem.
Currently, the workplaces for humanoid robots are mostly in markets such as manufacturing, firefighting, and power plant inspections, which are To G (government) or To large B (enterprise) markets. Domestically, perhaps only large government and enterprise entities have the funds and willingness to provide experimental scenarios for early-stage humanoid robots. However, the number of such customers is limited, and they pose high requirements for customer insight and customized services, which is a major challenge for robot companies that sell standardized hardware.
Completing a folding and alignment of capabilities with overseas benchmark tech companies is the first step for domestic embodied AI manufacturers. And the challenges have just begun.
The Second Folding: Fierce Domestic Competition, Who Can Build a Moat?
Unitree's 219x P/E ratio partly stems from the scarcity of being the "first humanoid robot stock" on the A-share market, as it is currently the only leading humanoid robot integrator in the market. However, with AGIBOT, Leju Robotics, and DeepRobotics successively receiving inquiry letters, is Unitree still scarce?
Or rather, is there any one among them that others cannot catch up with in the long term? That remains highly questionable.
Take Unitree, which has a first-mover advantage, for example; it probably does not possess an insurmountable moat. Its current advantages are reflected in the technological and commercial maturity of its quadruped robots, which have become cash cows; its mass production capability for humanoid robots, as the prerequisite for selling is being able to manufacture them; and its long-established supply chain advantages that dilute costs.
However, looking at the broader market, one will find that the gap does not seem that large.
From a product perspective, hundreds of embodied AI companies on the market are basically clustered in tracks such as humanoid integrators, quadruped robots, and robotic arms. The seemingly diverse product models essentially compete on assembly capabilities. Just like smartphones, integrators mainly focus on minor details such as data collection and model adaptation. It is also inevitable that they will face a "thousand-phone war" similar to the smartphone era, encountering a brutal elimination race.
What if Unitree introduces a disruptive product, just like the iPhone 4 crushed feature phones?
Whoever can first introduce products that lead the market by one or two generations will build an extremely high moat. Who will that be? The answer remains uncertain. Currently, the humanoid robot market is still a fierce competition among many players.
UBTECH, deeply engaged in industrial scenarios, already has full-size robots. Automakers represented by XPeng, Xiaomi, and BYD are directly reusing autonomous driving technology, betting heavily on the robot track, and have even found jobs for the robots—working on their own car factory assembly lines to achieve a direct closed loop. Meanwhile, AI companies represented by AGIBOT, along with numerous university-backed startups, are ambitious and want to rely on algorithm and model advantages to overturn the old guard.
Ultimately, perhaps only 3 to 5 companies will survive and capture 80% of the market, while other manufacturers will just be the "others" in market share. The second folding for embodied AI manufacturers is to build the "monopoly-level moat" that capital markets focus on, enabling them to operate at a higher dimension amidst fierce domestic competition.
The Third Folding: Folding the High-Valuation Future into Current Cash Flow
A friend who has been a retail investor (often self-deprecatingly called a "leek") in the spotlight over the past few months told me that for an emerging stock like Unitree, traditional value investing logic cannot be used. Looking at the P/E ratio is not enough; one must look at the Compound Annual Growth Rate (CAGR) of profits and the Price/Earnings-to-Growth (PEG) ratio... Phew, I almost understood these professional terms.
To summarize simply, a high P/E ratio for humanoid robot companies is not scary, as long as profits can grow every year, it is safe.
It is evident that steadily increasing profits are the only way to allow high valuations to soft-land, serving as the lifeline for embodied AI companies.
The good news is that Unitree's prospectus shows that the expected operating revenue for the first half of 2026 will be between 1.052 billion and 1.128 billion RMB, a year-on-year growth of 35.62% to 45.41%. Further breakdown reveals that the proportion of the quadruped robot business, acting as the cash cow, is declining, while humanoid robot revenue reached 868 million RMB in 2025, accounting for 51.78% of total revenue, officially becoming the main source of revenue.
However, this portion of revenue is highly dependent on the scientific research and education market. From January to September 2025, the proportion of humanoid robot revenue from the scientific research and education sector reached as high as 73.60%, while the proportion truly used in industrial scenarios such as intelligent manufacturing and inspections is relatively low.
Compared to singing, dancing, and bustling exhibition booths, perhaps only the dry financial statements and numbers in Unitree's prospectus represent the real world.
The humanoid robot industry seems bustling, but the jobs they can do are extremely limited, and they cannot yet bring sustained cash flow. The number of companies that can truly make money by selling robots can be counted on one's fingers.
The ChatGPT moment has not yet arrived. Current humanoid robots lack sufficient generalization capabilities, and most struggle to independently complete the closed loop of task decomposition and execution. Manual remote control remains the primary method. And even with remote control, the infrastructure requirements for the operating environment are very high.
A domestic embodied AI manufacturer, while executing inspection tasks, requires highly synchronized manual remote control commands and robot actions. Otherwise, if the commands are not transmitted in real time and the robot itself lacks a sufficiently intelligent brain for autonomous judgment, its movements will instantly become disordered, causing it to fall and convulse. This is unacceptable for both production scenarios and home services.
This is also the current realistic dilemma of humanoid robots. On the industrial end, doing simple transportation and sorting that AGVs and robotic arms can also do is, as the former head of Tesla's Optimus complained, a burden due to the complex structure of the humanoid form in industrial environments. In complex assembly or home scenarios, not only is their success rate substandard, but the costs are also high, making them neither easy to sell to C-end consumers nor easy to force into B-end factories.
Automakers making robots can still produce and sell them internally, but pure embodied AI companies need to weigh this carefully. Unitree Robotics also mentioned in its prospectus that if future new product promotions fall short of expectations, raw material costs rise, and R&D investment intensity increases, it will face the risk of declining operating performance and even losses.
If there is no sustained and rapid growth in profits, the hundred-fold premium for humanoid robots is just a huge bubble that cannot be absorbed at all. The talent war for tens of millions in annual salaries and the tech investment frenzy of guaranteed profits from IPO allotments are all using real money today to purchase a future that has not yet been fully proven.
While waiting for the IPO allotment results, consider this: how many painful "foldings" does the future of the humanoid robot industry need to go through to truly arrive at the present?