Unitree Hits All-Time Low Five Days After $66B IPO Debut on STAR Market

Five trading sessions after its record-breaking Shanghai debut, Unitree Robotics hit an all-time low of ¥602.05 on Monday, August 24, having shed approximately 45% of its peak market value in the span of one week. The company that briefly commanded a market capitalization of ¥444.9 billion (approximately $66 billion) at its intraday opening price on August 19 now trades at roughly ¥603 per share, giving it a market cap of about ¥272 billion (approximately $40 billion) — still nearly 300% above the ¥150.80 IPO price at which it raised ¥6.1 billion ($904 million) just days ago.

The gap between those two numbers tells the actual story. The crash from ¥1,100 to ¥603 is not a verdict on whether Unitree is a real business — it is. It is a verdict on how China's IPO machinery, combined with extreme investor appetite for the world's first mainland-listed humanoid robot company, produced an opening price that reflected a structural scarcity premium rather than a fundamental assessment of what profitable-but-research-concentrated robotic hardware is worth at scale. The market is now doing the work that retail frenzy could not: pricing in the commercialization gap between Unitree's viral backflipping demonstrations and the industrial deployment rates that would justify its valuation. The numbers that are driving that correction were all available before the first bell rang on August 19. The market is just reading them now.

How China's IPO Structure Built the Pop That Became the Crash

Understanding the August 19 surge requires understanding how China's securities regulator, the CSRC, governs initial public offerings. Unlike US and European exchanges, where IPO prices reflect extended investor roadshow negotiations and can be set closer to anticipated market value, the CSRC deliberately underprices listings to protect retail investors from immediate post-listing losses. This structural feature means that high-demand STAR Market listings routinely produce first-day gains that have no direct equivalent in Western markets. CXMT, the Chinese memory chipmaker that listed on the same exchange just weeks before Unitree, surged 466% on its debut day — a jump that was treated as remarkable in isolation but was partly the same institutional mechanism at work.

Unitree's 629% intraday peak and 460% closing gain exceeded even the STAR Market's inflated standard — the exchange's average first-day gain in 2026 has been running above 279% — but additional structural factors pushed the magnitude further. Unitree offered only 10% of its enlarged share capital to the public, the regulatory minimum. Of that already-thin float, strategic investors including DeepSeek (which committed approximately ¥140.8 million, or roughly $21 million), Tencent, Alibaba, Ant Group, Geely Capital, Meituan, China's National Social Security Fund, and several state-linked entities received allocations carrying lock-up periods of 12 to 36 months. That further reduced the number of shares that could change hands on day one. The online subscription tranche was oversubscribed more than 8,000 times, with 9.78 million investor accounts competing for approximately 9.7 million available retail shares — a lot-winning rate of 0.018%, the lowest in STAR Market history. When retail demand is compressed to that level against that supply, price formation on day one reflects scarcity, not valuation.

The third amplifying factor was political signal. Wang Xingxing had earned a seat alongside China's most prominent technology executives at a high-profile meeting with President Xi Jinping earlier in the year — a visible indicator of state-strategic priority that domestic institutional investors interpret as a directional signal. Unitree's CSRC approval from filing to listing cleared in a record 104 days, far faster than comparable high-profile applications, which is itself a form of government prioritization that market participants read and price.

The second trading day removed the scarcity premium with characteristic speed: Unitree shares fell more than 18% as the initial crowd of buyers with no alternative allocation disappeared and the ordinary dynamics of secondary trading took over. By the close of August 21, the stock had reached ¥672.41, erasing more than ¥170 billion ($25.3 billion) in market capitalization from its peak in just three sessions. Today's move to new all-time lows continues that trajectory.

The Fundamentals the Opening Price Skipped

The valuation the market briefly assigned at ¥1,100 per share was approximately 1,200 to 1,300 times Unitree's trailing earnings — a multiple that no equipment manufacturer and very few software companies anywhere have ever sustained. Even the IPO price of ¥150.80, which priced Unitree at roughly 219 times 2025 earnings and 36 times revenue, would have attracted significant skepticism had it arrived without the STAR Market's structural mechanics. The numbers behind those multiples are worth examining, because they are what is now being repriced.

Unitree's 2025 financial record is genuinely exceptional for its sector. The company reported ¥1.708 billion (approximately $253 million) in revenue, up 335% year-over-year, with net profit of ¥278.21 million ($41 million) and a gross margin of 60.27%. Those figures make Unitree the only profitable, publicly traded pure-play humanoid robotics company at scale anywhere in the world — a real competitive moat in a sector where almost every other player is burning cash. The same vertical integration that produces those margins, however, also contains the company's most visible structural constraint.

The story changed in the first quarter of 2026. Revenue grew 68.5% year-over-year to ¥422.84 million — an impressive number on its face, and a sharp deceleration from the 335% growth rate of 2025. More significantly, adjusted net profit fell 52.55% in the same period as marketing spend for the Spring Festival Gala performance and research and development investment accelerated ahead of the listing. For the first half of 2026, Unitree reported revenue of ¥1.152 billion (up 48.54% year-over-year), but with profit under continued pressure. The pattern — revenue growth decelerating while investment costs accelerate — is characteristic of a company pivoting from a hardware business toward what Wang himself has described as a platform for embodied AI. That pivot requires capital. It also requires time.

The deeper constraint is not on the income statement — it is in the revenue mix. According to Unitree's own prospectus, more than 70% of its humanoid robot revenue through the first three quarters of 2025 came from research institutions and universities. Industrial deployments, the category that justifies trillion-dollar long-run forecasts, accounted for less than 10% of revenue in the same period. Independent industry estimates place the number of active Unitree humanoid units in real-world labor applications at roughly 250 to 500 units as of mid-2026, out of more than 18,000 cumulative bipedal robots shipped across all models. Corporate demonstrations and guided tours made up the majority of what passed for industrial business. HSBC analysts, in a report cited before the listing, warned that the surge in humanoid robot shipments in China could prove illusory without significant improvements in the AI models that actually direct robot behavior in uncontrolled environments.

What QDD Actuators Enable and What They Cannot Do

The engineering choices that made Unitree the world's most recognizable robotics company are also what define its commercial ceiling, at least in its current product generation. Unitree's robots run on quasi-direct-drive (QDD) actuators — electric joints built around a low gear ratio (roughly 6:1 to 9:1) paired with high-torque-density rare-earth permanent magnet synchronous motors. The low gear ratio gives QDD joints a property called backdrivability: the joint can be moved by external force with low resistance, allowing the robot to absorb impacts and produce compliant, human-like motion. That is what enables the backflips, the martial arts demonstrations, and the Spring Festival Gala performance that made Unitree's videos viral. It is also what drives the 60% gross margins: Unitree designs and manufactures these actuators in-house, capturing margin that an external supplier would otherwise extract.

The trade-off is payload. Low gear ratios reduce the continuous torque available at the joint, capping practical payload capacity at roughly 3 to 5 kilograms (6.6 to 11 pounds) for H1-class humanoid platforms. Boston Dynamics' Atlas, the primary Western benchmark for full-size humanoid capability, uses a different actuator philosophy and is designed to lift approximately 23 kilograms (about 50 pounds). These are not directly competing products — they reflect different engineering priorities for different use cases — but the comparison matters when evaluating what kind of industrial work Unitree's current platform can actually perform.

Unitree's new "Superman" robot, unveiled days before the IPO listing and coordinated with the opening of the World Robot Conference in Beijing, made the tradeoff explicit. By building a platform with no hands at all, Unitree achieved a claimed running speed of 12.66 meters per second (28.3 mph, faster than Usain Bolt's world record) and a standing jump of two meters (6.6 feet). It is an impressive locomotion result. It is not a commercial product. Speed without manipulation capability has no direct application in the industrial scenarios — logistics, assembly, service operations — that justify the market's forward projections. The timing of the announcement, which maximized media coverage during the IPO window, illustrates the distinction between Unitree's marketing cycle and its product roadmap.

The Last Millimeter Problem and the VLA Bottleneck

The commercialization timeline Wang Xingxing offered at the World Robot Conference in Beijing on August 20, one day after the IPO, was more cautious than the market's enthusiasm had implied. According to Reuters and CNBC, Wang said the humanoid industry's equivalent of a "ChatGPT moment" — a capability threshold at which robots can complete roughly 80% of tasks across 80% of unfamiliar scenarios based on ordinary voice or text instructions — was achievable within two to three years on an optimistic path, or five to ten years if progress slows.

The specific technical failure Wang identified is what researchers in the field call the "last millimeter problem." Current humanoid robots can understand a task description, navigate to an object, and execute an approach. They routinely fail at the physical contact interface: the final few centimeters or millimeters where force feedback, tactile sensing, and unpredictable object geometry overwhelm the robot's current control stack. The problem is not the robot's legs or its ability to traverse the environment. It is the robot's hands, and more precisely the AI system directing those hands.

The technical framework for addressing this is called a Vision-Language-Action (VLA) model: a neural network that takes visual input (camera feed) and language instructions ("pick up the blue bottle") and outputs motor control signals for the robot's joints. Current VLA models can produce correct high-level behavior — approach the right object, reach toward it — but lack the ability to correct errors at the contact surface when the object does not behave exactly as the training data predicted. Every deployment of Unitree robots in research environments is partly a data-collection exercise: the contact-level failures those robots experience become training examples for the next generation of VLA models.

That is why Unitree's strategic partnership with DeepSeek, formalized in a public stock-exchange filing disclosed on August 6, 2026, identified the VLA bottleneck as the specific obstacle preventing humanoid robots from performing autonomous industrial work at scale. DeepSeek, whose open-weight language models have already reshaped the AI industry's cost assumptions, brings large-model training capability to a problem that requires it. Nvidia made the same diagnostic from a different direction: in naming Unitree's H2 Plus body as the hardware foundation for its open GR00T humanoid reference platform, Nvidia specified that the platform pairs the H2 Plus with Sharpa Wave tactile hands containing more than 1,000 tactile pixels per fingertip — a configuration that signals where frontier researchers believe the manipulation gap actually lives. Not in the motors. In the fingertips and the models that interpret what they feel.

AgiBot Has Already Overtaken Unitree by Volume

The competitive landscape that greeted Unitree's stock market debut contained a development that received less attention than the opening pop: Unitree was no longer the top humanoid robot shipper in its most recent reporting period. According to Smart Analytics Global (SAG), AgiBot — Unitree's primary domestic rival — commanded 44% of the global humanoid market in the first half of 2026 with approximately 8,400 units shipped, against Unitree's 31% share and approximately 5,900 units in the same period. The South China Morning Post and Bloomberg separately confirmed the SAG data.

That shift matters for how the market should interpret Unitree's current ¥272 billion ($40 billion) valuation. The bull case for Unitree's premium rests substantially on its position as the world's leading humanoid robot shipper, its verified profitability, and its 60% gross margins. If AgiBot maintains its lead on volume through the second half of 2026, and then proceeds with its planned Hong Kong IPO — for which Unitree's debut price will serve as the primary benchmark — the comparison will become a direct test of whether Unitree's profitability premium or AgiBot's volume leadership commands the larger multiple. AgiBot is not yet profitable, but it is growing faster by units.

Beyond the domestic competition, Unitree faces a shrinking addressable market in the United States. The FCC added all foreign-produced advanced robotic devices to its Covered List on July 28, 2026, under public notice DA-26-786, citing security vulnerabilities confirmed in Unitree hardware. The action blocks new or previously unauthorized foreign robot models from receiving the FCC equipment authorization required for import, marketing, or sale in the United States. Unitree's six best-selling existing models — the G1, H2, and R1 humanoids and the Go2, B2, and A2 quadrupeds — cleared FCC authorization before July 28 and remain legal for sale from existing inventory. But the pipeline of new US products is effectively closed until Unitree can obtain conditional approval through a process the Department of War oversees, with no clear timeline.

Two specific security vulnerabilities drove the FCC's designation and remain unpatched as of the stock's all-time low today. CVE-2025-2894 (the CloudSail backdoor), cataloged under CWE-912 (Hidden Functionality), describes a firmware service that auto-connects on boot to a cloud tunnel operated by a Chinese company called Zhexi Technology, granting anyone with the correct API key complete remote control of the robot. Researchers confirmed vulnerable units were operating inside networks at MIT, Princeton, Carnegie Mellon University, and the University of Waterloo. The second vulnerability, dubbed UniPwn and documented by IEEE Spectrum, is a wormable Bluetooth Low Energy exploit affecting Unitree's Go2, B2, G1, and H1 models that yields root-level system access and can automatically propagate to nearby Unitree units — a botnet mechanism built into the hardware by design flaw. The Department of War separately listed Unitree under Section 1260H on June 8, 2026, barring US defense contracts immediately and extending broader supply chain restrictions to June 2027.

The United States accounted for roughly 13% of Unitree's revenue in 2025. Overseas markets in total contributed about 44%. Wang Xingxing has publicly acknowledged that geographic diversification toward Europe, Japan, South Korea, and Southeast Asia is now a strategic priority — a direct response to the structural ceiling that US policy has created. Japan Airlines' operational trial of Unitree G1 units at Tokyo's Haneda Airport for cargo handling and cabin cleaning is the clearest current example of that pivot in practice.

What ¥603 Per Share Is Actually Pricing

At Monday's close of approximately ¥603, Unitree trades at roughly 360 times its 2025 adjusted earnings and about 160 times its trailing twelve-month revenue. Those are still historically elevated multiples — well above any conventional equipment manufacturer and most mature technology companies. The single analyst price target confirmed by Investing.com as of August 24 sits at ¥370, implying approximately 38% further downside from current levels.

The market is not, however, pricing Unitree as a conventional equipment company. It is pricing an option on a future industry — specifically, the scenario in which the VLA bottleneck yields to the next generation of AI models and tactile sensing hardware, industrial deployment rates accelerate from their current low single-digit share of revenue, and Unitree's manufacturing capacity expansion to 190,000 units annually (75,000 humanoids, 115,000 quadrupeds) generates the throughput that a ¥272 billion market cap implies. Morgan Stanley has projected China's humanoid robot market reaching $15 billion by 2030 and global annual unit shipments approaching 446,000 by the same date. If even a fraction of those projections materialize and Unitree retains meaningful market share, the current multiple is defensible. If the VLA breakthrough takes ten years rather than two, if AgiBot's volume lead compounds, or if US market closure proves more economically significant than anticipated, the current multiple is not.

Wang Xingxing himself was precise about where the uncertainty lies. At the World Robot Conference, according to Reuters and Fortune, he described robots that can currently approach objects correctly but fail at the final contact point. That admission — from the founder, on the day after a $66 billion debut — is the most accurate summary of where the sector's defining technical problem actually sits. The next several quarters will determine whether the industrial deployment data — reorder rates after pilots, productive hours per robot per shift, task completion rates in uncontrolled environments — begins to narrow that gap at a pace consistent with current valuations. If it does, ¥603 will look like a discount. If the "last millimeter" gap proves as stubborn as Wang's more conservative timeline suggests, the market's price discovery process has more work to do.

The Float Mechanics That Will Define the Next Twelve Months

One aspect of the post-IPO decline that has received insufficient attention in coverage focused on the crash magnitude is the lock-up structure that still constrains Unitree's price discovery. Of the shares Unitree sold in its IPO, 20% went to nine strategic investors who received allocations on terms that bar them from selling for between 12 and 36 months. These investors — which include DeepSeek's parent company (36-month lock), Tencent's Shanghai Qishan investment vehicle, CNPC's Kunlun Capital, China Southern Power Grid's finance arm, China Telecom's Tianyi Capital, and the National Council for Social Security Fund — collectively hold a position worth approximately ¥54 billion (approximately $8 billion) at today's price, locked away from the secondary market.

Management and core employees took an additional ¥271.5 million through two employee asset management plans, subject to their own holding requirements. Wang Xingxing himself, who controls 68.78% of voting rights despite holding approximately one-third of the economic interest, is subject to the standard founder lock-up of 36 months from listing. Meituan, the food-delivery and technology conglomerate that is Unitree's largest outside shareholder with an 8.7% post-IPO stake — a position worth roughly ¥23.7 billion ($3.5 billion) at today's price — is subject to a lock-up as a pre-IPO strategic investor.

The implication is that the freely tradable supply of Unitree shares is a small fraction of the total share count, and it will remain so for the better part of the next three years. When the 12-month lock-up window opens in August 2027, investors who acquired allocations at prices well below the IPO price will face a decision about whether to sell into whatever market exists at that point. That event horizon is likely to be a source of structural overhang and volatility, and it is already visible to sophisticated market participants pricing the stock today. The lock-up schedule means that today's correction, while steep from the peak, is happening with the vast majority of Unitree shares still legally untradeable. Price discovery has further to run in both directions.

The Benchmark Effect on the Humanoid Robotics Sector

The most significant consequence of Unitree's IPO — separate from what it says about Unitree specifically — is the benchmark it has created for an entire industry that previously had no daily-priced public valuation. Before August 19, 2026, the only data points for humanoid robotics valuations were private funding round prices, which are disclosed infrequently, at round prices that reflect negotiated terms rather than continuous market clearing, and with a selection bias toward successful rounds. Figure AI's $39 billion private valuation, widely cited as the sector's prior high-water mark, was set at a moment of maximum venture capital enthusiasm and carries no public market accountability.

Unitree's STAR Market listing has changed that, even if the lessons are complicated by the structural mechanics described above. At today's corrected price, Unitree trades at approximately 160 times trailing revenue and 360 times trailing earnings. Those ratios, embedded in a publicly available daily price, are now the reference point for every valuation model in the sector. UBTECH Robotics, which has traded on the Hong Kong Stock Exchange since December 2023 and is the prior benchmark for listed humanoid robotics companies, trades at substantially lower revenue multiples on far lower shipment volumes and no profitability. Agility Robotics is pursuing a US public listing via SPAC at an implied valuation of approximately $4 billion — a figure that now looks conservative relative to Unitree's multiple even at today's reduced price, which has prompted analysts to scrutinize whether Agility's pre-order pipeline and deployment track record at Schaeffler, GXO, Toyota, and Mercado Libre can support a premium to Unitree's current market-implied revenue multiple.

AgiBot's Hong Kong IPO, which was widely expected to launch in the second half of 2026, will now be priced in explicit comparison to Unitree's trading history. That comparison will be structurally complicated: AgiBot has surpassed Unitree on H1 2026 unit volume but is unprofitable, while Unitree maintains its profitability advantage and the manufacturing depth that comes from owning its actuator supply chain. The market will have to decide which it values more — the revenue mix that currently favors research over industrial deployment is identical in its structural weakness for both companies. The fact that China's broader investor base poured more than ¥100 billion ($14.9 billion) into the humanoid robotics sector in 2026 alone, according to Bloomberg, suggests that demand for the benchmark is far from exhausted.

Wang Xingxing's Strategic Pivot and Its Cost

The financial profile of Unitree's first quarter of 2026 — revenue growing at 68% while adjusted profit fell 52.55% — is not a coincidence or an accident. It is the visible signature of a deliberate strategic pivot that Wang has been telegraphing since at least early 2026 and that the company's IPO prospectus makes explicit. Unitree is in the process of redefining itself from a hardware manufacturer — the company that builds the world's best-selling robot bodies — to a platform company for embodied intelligence. The largest single allocation of IPO proceeds (¥4.2 billion / $625 million) is directed toward AI model research, exceeding the allocation to robot body hardware development. That prioritization tells investors what management believes is the company's forward value driver: not the actuators, but the AI that makes them commercially useful.

The cost of that pivot is landing on the income statement in real time. Research and development headcount has grown faster than revenue. Marketing spend on the Spring Festival Gala performance — which cost more than its immediate revenue return justified on a short-term basis — was an investment in brand visibility and training data, not a quarterly profit optimization. Wang's challenge is that he is trying to make that investment case to a public market that bought into Unitree at 219 times earnings and is now watching the earnings it paid for fall further by the quarter. The IPO gave Wang the capital to accelerate the AI investment. The crash reflects the market's uncertainty about how long it will take before that investment produces earnings that justify even the reduced multiple the stock carries today.

Nomura analysts, in a research note on Unitree cited in the days following the IPO, identified the company's scale of real-world robot deployment as a structural advantage in building the training data required for next-generation AI models. The argument is that 5,500 humanoids deployed in research environments, generating contact-level failure data that Unitree can use for reinforcement learning and VLA model improvement, creates a data flywheel that pure software companies cannot replicate. If that flywheel produces a capable enough VLA model — one that closes the "last millimeter" gap Wang described at the World Robot Conference — the return on the current R&D investment could be substantial. If the data flywheel proves slower or less efficient than expected, the investment cycle will have compressed margins during a valuation correction, a combination that makes the path to justifying current multiples longer.

The Regulatory Ceiling and the Reorientation East

The US market constraint deserves treatment as an economic reality rather than a geopolitical abstraction, because the numbers are concrete. The United States accounted for roughly 13% of Unitree's approximately $253 million in 2025 revenue — approximately $33 million. That figure is not recoverable through new product sales under the FCC's Covered List designation. Future models require conditional approval through a process that has no established timeline and that Unitree's own legal structure makes difficult: China's National Intelligence Law (2017), Article 7, requires all Chinese companies to support, assist, and cooperate with national intelligence work. No US regulatory process for conditional approval operates independently of that legal reality, which is a structural feature of Unitree's Chinese incorporation rather than a negotiable aspect of its product design.

The FCC cited two confirmed vulnerabilities in its designation rationale, and those vulnerabilities add a technical dimension to the regulatory barrier that is separate from the geopolitical one. CVE-2025-2894 describes a hidden firmware service that automatically establishes a cloud tunnel on boot to a server operated by Zhexi Technology in China. UniPwn describes a Bluetooth Low Energy exploit that can propagate automatically between nearby Unitree units without user intervention, a wormable mechanism that IEEE Spectrum characterized as a category-level security problem rather than a product-specific one. The combination of architectural vulnerability and governing-law obligation creates a US market ceiling that cannot be resolved by software update alone.

The strategic response Wang has described — diversification toward Europe, Japan, South Korea, and Southeast Asia — reflects a realistic reading of where the growth must come from. Japan Airlines' operational trial at Haneda Airport, Japanese startup ZEALS' hospital-service D1 humanoid built on a Unitree G1 body, and the commercial launch in Europe in June 2026 are the early evidence of that pivot. Whether these markets can absorb the volume that a planned 190,000-unit annual manufacturing capacity would produce, and whether they can do so without the pricing leverage that Unitree's US market access once provided, is the deployment question that will define the company's revenue trajectory over the next three to five years.

The correction from ¥1,100 to ¥603 has been steep, fast, and, at these fundamentals, probably not finished. But the company it is correcting toward — the one that is profitable, that has shipped more humanoid robots than any other manufacturer in history by year-end 2025, and that is making the AI investment required to close the manipulation gap — is a real company with a real technology base. The question the market will spend the next several years answering is not whether Unitree deserves to exist at a premium valuation. It is what that premium should be for a company where the gap between what the robots can demonstrate and what they can reliably produce on a factory floor still measures in millimeters, and where the timeline to closing that gap belongs to the same founder who just admitted it could take a decade.