The humanoid robotics industry has entered a strange season. On one side of the ledger, private capital is arriving at a pace the sector has never seen: Dealroom counts $8.7 billion of venture investment into humanoid startups in 2026 year-to-date, already double the full-year record set in 2025. On the other side, the industry’s only large, liquid, pure-play public marker has spent six weeks teaching investors what repricing looks like. Unitree Robotics, which closed up 460% on its Shanghai debut in August, has since surrendered roughly half its peak value.
Both facts are true at the same time, and the tension between them is the most important market structure in the industry right now. The humanoid boom is not deflating. It is bifurcating: into assets that public markets will price with growing discipline, and assets that private markets continue to price on narrative. This month delivered three signals in rapid succession that the gap between those two pricing regimes is starting to close: Unitree’s slide, Beijing’s regulatory brake on humanoid IPOs, and Vinod Khosla’s public warning that most robotics startups will be worth less, not more, by 2030. A fourth signal, Figure AI’s $3.5 billion compute commitment, explains why the reckoning will arrive sooner than the industry’s fundraising cadence suggests.
The only clean public mark
Unitree matters for a boring reason: it is the only humanoid company whose financials a retail investor can read without a data-room invitation. Its round trip is therefore the closest thing the sector has to a market clearing price, and the numbers deserve precision.
The Hangzhou company raised about 6.1 billion yuan, roughly $900 million, pricing its IPO at 150.80 yuan per share for a valuation near $9 billion. On August 19 the stock closed at 845 yuan after touching 1,100 yuan intraday, briefly valuing the company at roughly 445 billion yuan, about $66 billion. By the final week of September it traded near 513.93 yuan: down about 39% from the debut close and 53% from the first-day high, a round trip that erased roughly $35 billion in market value from the peak. Even after the fall, the stock remained more than three times its IPO price.
The instructive part is what the market was repricing. Unitree is not a pre-revenue story. It generated 1.70 billion yuan, about $252 million, in 2025 revenue, up from 392.77 million yuan in 2024. Humanoid robots contributed 868 million yuan, or 51.78% of the total, on more than 5,500 units shipped. First-half 2026 guidance implies 36% to 45% year-over-year growth. This is a real business with real unit volume. What cracked was the multiple: at the peak, investors were paying more than 250 times 2025 revenue for it. At the late-September close, about $30 billion, the multiple was still roughly 125 times revenue and more than 350 times adjusted 2025 earnings.
Two disclosures inside the IPO filing explain why enthusiasm cooled once parsed. The Wall Street Journal reported that less than 10% of Unitree’s 2025 revenue came from industrial applications, the deployments that imply recurring labor value. And the Financial Times has documented that China has built more than 90 humanoid training centers, many co-funded by local governments and robot manufacturers, which buy humanoids largely to generate teleoperation training data. Some meaningful share of Chinese humanoid demand is, in other words, demand for data generation rather than demand for labor substitution. The market’s question after the debut was exactly the right one: how much of this revenue is repeatable, and by whom?
Beijing applies the brakes
The Unitree round trip did not happen in a regulatory vacuum. On September 9, Reuters reported, citing The Information, that the China Securities Regulatory Commission had informally raised the bar for humanoid IPO candidates. Prospective listings should now demonstrate recurring revenue, progress toward reducing losses, or significant technological innovation. The Wall Street Journal separately reported the same signal from Chinese regulators.
The context is broader than one listing. Chinese authorities have spent months warning about speculative froth in humanoid robotics, and press accounts describe private-market projects that have already absorbed valuation cuts of 30% to 50%, along with founders who fielded dozens of prospective investors within weeks and refused conventional due diligence. That behavior is the classic late-cycle signature: capital chasing category exposure faster than diligence can be performed. Beijing’s response, slowing the IPO conveyor, has a double effect. It protects retail investors from debut-day manias, and it strands late-stage private capital in companies whose nearest exit window just narrowed.
For the global industry, the Chinese brake matters because China had been the sector’s liquidity engine. The largest 2026 rounds tracked by Dealroom include not only Western names like Apptronik and Neura Robotics but a deep bench of Chinese entrants, and the STAR Market had become the sector’s de facto public proving ground. If that window now admits only companies with recurring revenue and shrinking losses, then the Chinese humanoid cohort faces exactly the discipline that Khosla says the Western cohort is avoiding. The first companies through the tightened gate will define what a “qualified” humanoid listing looks like for years.
Khosla’s arithmetic
Vinod Khosla occupies an unusual position in this debate: he is simultaneously the industry’s loudest bull and its most credible bear. Khosla Ventures backs multiple robotics companies, and he still predicts robotics will have its “ChatGPT moment” within roughly two years. Yet in an interview with The Information late last month, he argued that excessive investment and lofty valuations in the hottest robotics startups portend a major shake-out. His formulation: by 2030, more than half of these companies will carry valuations below their current levels, while the survivors command astonishing multiples.
This is not a contradiction. It is standard power-law arithmetic applied to hardware. If robotics produces one or two platform winners with the economics of an operating system, then a portfolio of forty contenders priced at $39 billion each is collectively mispriced even if the underlying technology thesis is correct. Khosla’s warning is best read as a statement about entry prices, not about the technology. The sector’s problem in 2026 is that entry prices were set during a window when any humanoid company with a credible demo could command a valuation built for the winner-take-most scenario it almost certainly will not win.
The public markets have already run this logic on Unitree and found even a $30 billion valuation generous. Private markets have run it on almost no one, because private markets can. The structural feature of this cycle is that repricing requires a transaction, and there is no obligation to transact. Late-stage humanoid startups can sit on their 2025 marks for years, raising insider rounds at flat valuations and calling it discipline, while public investors mark the same technology to market daily. The longer that divergence runs, the more violent the eventual reconciliation for whoever is still holding 2026 paper.
The compute bill comes due
The force that will accelerate this reconciliation is not sentiment. It is the capital expenditure requirements of vision-language-action models, the software stack that actually runs the robots. On September 3, AI cloud provider Nscale and Figure AI signed a multi-year agreement worth at least $3.5 billion in GPU infrastructure, with a stated intention to scale beyond $6 billion. It is the largest dedicated compute commitment ever made to a humanoid company, and Nscale took an undisclosed equity stake in Figure as part of the structure.
The number deserves context. Figure has raised approximately $1.9 billion in its entire history, against a reported $39 billion valuation. It has now committed roughly double its lifetime fundraising to compute alone. The deal’s structure, compute now in exchange for equity in the upside, is the market’s first explicit answer to a question the humanoid industry has been dodging: who pays for the training infrastructure, and on what terms? That a vendor felt compelled to take equity to make the deal balance tells you what it thinks of the ability of humanoid balance sheets to pay cash.
The reason the bill is structural rather than one-time lies in how VLA models differ from language models. A large language model trains on internet-scale text that was collected once and effectively for free. A vision-language-action model, the family that includes Figure’s Helix, trains on synchronized triplets of visual observations, language instructions and motor actions, every example of which requires a robot or a teleoperator physically generating trajectories. Helix issues 200 commands per second to the robot’s fingers, wrists and torso, and unlike an LLM that can be pretrained once and fine-tuned cheaply, a deployed VLA keeps accumulating embodied data as its fleet meets new objects and environments. The compute requirement does not end when training ends. It scales with deployment.
Multiply that structure across an industry racing toward fleets of thousands, and the sector’s real capital requirement starts to look less like a hardware industry and more like the AI-lab model: continuous multi-billion-dollar compute burn, funded by rounds that must arrive on schedule forever. Figure’s answer is vendor equity and Jensen Huang’s “physical AI flywheel.” Anthropic, for scale, agreed the same week to pay Nscale roughly $45 billion for West Virginia data-center capacity. The humanoid sector is now competing for the same compute pools as frontier labs, with thinner balance sheets and no cloud-revenue business to absorb the cost.
Two models, one test
Strip away the geographies and two balance-sheet architectures have emerged. The Chinese model, Unitree’s, is hardware-margin-first: sell units, scale the supply chain, accept that profit falls while you scale, and let public markets judge the multiple. Unitree’s first-quarter 2026 reporting showed revenue growing rapidly while profit fell by more than half under the cost of scale, and the market’s response was to compress the multiple rather than abandon the company. That is how price discovery is supposed to work.
The Western model, Figure’s and arguably Tesla’s, is compute-first: absorb enormous training costs before meaningful robot revenue exists, fund it with equity and vendor partnerships, and defer the margin question until the fleet learns enough to be useful. Tesla’s version of the bet is the largest in absolute terms. It is building several hundred Optimus units per week at Fremont, roughly ten times its Q2 output, with a year-end target above 1,000 per week and an eventual goal of 20,000, Electrek reports. Supply-chain data points to orders for roughly 15,000 units this year, against a fleet that mostly generates training data internally, on top of more than 500,000 accumulated training hours. Tesla can carry this because it is Tesla. A startup portfolio cannot.
Both models eventually face the same test, and it is the one Chinese regulators have already written into their IPO guidance: recurring revenue. Not units shipped, not demo videos, not fleet size, but customers who pay again because the robot did the work. Unitree’s disclosure that less than 10% of its 2025 revenue was industrial says the sector’s most advanced commercial player has not passed that test yet. Neither model’s unit economics are proven at the scale the valuations assume. The difference is that public markets are now charging for that uncertainty daily, while the $8.7 billion of 2026 venture inflow is still largely pricing it at zero.
The tests to watch
Four signposts will tell you whether this repricing stays orderly or turns into the shake-out Khosla predicts. First, the next humanoid listings in Shanghai under the tightened CSRC bar: what multiple does the market assign a humanoid company that discloses recurring revenue honestly? Second, whether any Western humanoid startup accepts a flat or down round in the next two quarters, which would mark the private market’s first break with 2026 marks. Third, whether Figure-style vendor-equity compute deals spread, because each one is an implicit admission that cash obligations exceed the balance sheet. Fourth, whether Tesla’s reported leasing structure for Optimus, which keeps fleet economics and data rights on the lessor’s side, becomes the template for how robots are actually sold, since leasing converts robot economics from a hardware sale into a service contract measured precisely in recurring revenue.
None of these outcomes requires the humanoid thesis to be wrong. Unitree at $30 billion is still a bet that general-purpose robots conquer labor markets; it is simply a cheaper one than Unitree at $66 billion. The bull case and the valuation case have decoupled, and an industry that spent 2026 fundraising on the narrative of the former now has to survive the arithmetic of the latter. The robots, as ever, are on schedule. The capital cycle is early.