Just under a month after its debut, Unitree Technology has seen its market value erased by over ¥240 billion, with shares now trading more than 57% below their peak. On September 15, the stock broke below ¥500, closing at ¥469.8, which brought the company's market capitalization down to ¥190.02 billion. This staggering loss is equivalent to wiping out the entire market value of 0.98 JD.coms, 3.1 Baidus, or 5 Haidilaos. A week prior, on September 9, a report sent a chill through the humanoid robot industry: the China Securities Regulatory Commission had reportedly issued informal "window guidance" to some investment banks and institutions, raising the IPO bar for humanoid robot companies. Aspiring public listings must now prove they have sustainable revenue, are on a path to reducing losses, or possess genuinely critical technological innovation. Capital, which spent the past two years paying for future promise, is now demanding that robot companies demonstrate their ability to make money today.

Why Can't the High Valuation Hold Up?

On the day Unitree Technology went public, the capital market celebrated, but the manufacturing sector was puzzled. As a company in the general equipment manufacturing industry, Unitree generates annual revenue of over a billion yuan and profits of two to three hundred million. An opinion piece cited by Economic Observer Online, from Shi Huashan, Chairman of Guangdong Hanyu Group, suggested that by standard manufacturing valuations, such scale and profit could not justify a market cap of over ¥200 billion. He estimated a reasonable valuation would be around ¥70 billion—which, within manufacturing circles, was already considered generous. However, Unitree is not merely a company with a compelling narrative. In the first half of 2026, revenue reached ¥1.152 billion, a 48.54% year-on-year increase, with net profit attributable to parent of ¥274 million, a turnaround from a loss in the same period last year. According to Counterpoint data, Unitree's humanoid robot shipments in H1 were approximately 5,900 to 7,000 units, capturing about a 31% global share, second only to Zhiyuan, with its flagship G1 model dominating the scientific research and education segment.

But the ¥444.9 billion peak valuation on its first day was not paying for these results. The IPO price was ¥150.8, corresponding to an offering price-to-earnings (P/E) ratio of 219 times, nearly six times the general equipment industry average of 38 times. On its first trading day, the stock opened at ¥1,100, with the dynamic P/E ratio briefly exceeding 700 times. This pricing effectively discounted all future expectations of humanoid robots entering factories and homes years in advance. Another contributing factor is the extremely small public float: only about 30.08 million shares, or 7.44% of total share capital, were initially available for trading. With roughly 70% of that held by institutional offline subscribers, a small amount of capital could drive the price to extremes, and conversely, a small number of sell orders could break it when sentiment shifted. Yet, high valuations require high growth to sustain them, and that growth is decelerating. Revenue growth slowed from 332.64% in full-year 2025 to 68.49% in Q1 of this year, and further to 48.54% in H1. Non-GAAP net profit fell 19.34% year-on-year to ¥244 million. The revenue structure also raises questions: over 70% of Unitree's humanoid robot revenue comes from scientific research and educational institutions, while actual adoption in industrial manufacturing scenarios remains in the single digits.

The reference point here is DJI. In its early days, DJI sold drones to research institutions and for city light shows, hitting a visible ceiling. It later made two pivotal transitions: turning drones into a mass consumer product, and then using software and algorithms to become an imaging technology company. Today's Unitree, selling primarily to labs and stages, is receiving a valuation befitting a futuristic tech giant. Fu Mengzhen, head of MIR Ruigong's embodied intelligence team, told Pen & Pencil that short-term stock price fluctuations won't change the industry's direction, noting that "in the next few years, OEMs must focus on application scenario deployment." But the timeline of the capital market and that of the industry have become misaligned.

Who Are the Robots Actually Sold To?

To understand this disconnect, one must look at who actually buys the robots. Counterpoint data shows that global humanoid robot shipments exceeded 22,000 units in H1 2026, a nearly 300% year-on-year surge, with full-year projections exceeding 50,000 units. However, over 60% of these went to entertainment performance, commercial shows, education, and data collection; intelligent manufacturing accounted for only 13% and warehousing/logistics 5%. For every ten robots sold, more than six end up on stages, in labs, or in data collection workshops; fewer than two actually work in factories. This is precisely the issue highlighted in the earlier report and the reason regulators are demanding "sustainable revenue"—sales volumes can be inflated by a specific type of buyer, and the quality of that revenue is now being publicly questioned.

At the end of August, Shao Tianlan, whose company Mech-Mind Robotics was just listed on the Hong Kong Stock Exchange, posted a series of critical comments on his WeChat Moments in early September, attacking the industry's "project-creation" startups. He alleged that some high-profile, high-valuation humanoid robot companies in Beijing and Shanghai, famous for appearances on Spring Festival Gala, were constructing trading relationships around "data collection centers" and "leasing companies." By generating orders and revenue through related parties like local governments, investors, and suppliers, they continuously release fundraising and contract announcements to inflate valuations, aiming for IPO within two to three years of founding. In his comments, he directly named Galaxy General, calling on regulatory authorities to verify his claims. This operational chain works as follows: embodied intelligence companies and local governments jointly build data collection factories, which become a major shipment destination for the robots, creating substantial purchase orders. The data is then bought back by the manufacturer. The contracts and payments are real, but the question remains—without this intricate relationship, would external customers still buy at the same price? Can these orders be sustained next year?

Galaxy General responded with a lengthy statement on the evening of September 10, saying it would "not engage in unproductive verbal battles," and introduced its self-developed large models and application scenarios in industry, retail, and pharmacies, without directly addressing the related-party transaction concerns. Founded in 2023, the company has accumulated approximately ¥7 billion in funding, boasts a valuation exceeding ¥20 billion, and is actively pursuing an IPO. On August 28, the National Development and Reform Commission publicly called on the robotics industry to prevent "blind imitation and rushing into action," repeatedly emphasizing the need for "realistic scenarios" and "genuine demand." Just over ten days later, word of the window guidance emerged.

The Next Round is About Repeat Purchases

The tightening regulation stems from the long queue of companies waiting. Industry statistics indicate that over 40 robotics companies are in various stages of capital raising, with Zhiyuan, Galaxy General, LimX Dynamics, and Galaxea AI having already submitted listing materials. In H1, total industry financing reached ¥93.5 billion, with humanoid robot financing at $5.4 billion (Crunchbase metric), more than double the entire 2025 figure. With abundant capital, numerous companies, and scarce real revenue, the door naturally narrows. Behind that door lies a single test: replicable, real-world scenarios. Fu Mengzhen pegs the timeline for this at 2028—before that, more manufacturers will develop replicable applications in industrial and commercial settings. She stresses that the key word isn't "scenario," but "replicability": a robot moving boxes in one auto plant is a project, but being able to replicate it from one factory to ten, or from one workstation to a hundred, without months of recalibration in each new environment, is a business.

Changes are already emerging. Jiao Jichao, Vice President of UBTech, revealed in August that some industrial clients are beginning to make repeat purchases of humanoid robots. UBTech sold 1,079 full-size humanoid robots in 2025, with over 80% entering industrial scenarios. However, a robot's initial time to move a material bin was 4.5 minutes, optimized to 2 minutes, while a worker takes under 1 minute. Its overall efficiency is about 30% of a skilled worker. UBTech has set an internal target of reaching 60% efficiency by 2026, which is viewed as the break-even point for large-scale owner adoption. The cost of deploying a humanoid robot in a factory is currently ¥600,000-700,000, but auto manufacturers say it only becomes economically viable at around ¥250,000. Fu Mengzhen notes that capital and public patience are limited, making "a market shakeout inevitable." The shakeout will separate winners not by who can do backflips more reliably, but by two questions: Will customers buy a second batch after the first? And can success in one factory be replicated across a hundred? Before OEMs achieve scale profits, money may flow to more certain parts of the supply chain: as long as shipment volumes keep rising, reducers, ball screws, motors, and sensors will see continuous orders, and service providers focused on data, integration, and retrofitting around factories will have revenue first.

On the day of Unitree's big drop, its TTM P/E ratio still stood at 345 times. This suggests the market hasn't completely abandoned the industry, but it no longer accepts that all companies should be priced on "future" potential alone. Can it be replicated? Can it recoup costs? Can it make a profit? These three questions have replaced backflips and funding records as the new litmus test for humanoid robots. This article does not constitute investment advice.