PANews has published a long-form article by 0xjacobzhao as part of IOSF’s "Physical AI 2026" series, arguing that embodied intelligence is moving from concept to deployment under a major mismatch between capital enthusiasm and deployment economics. The article’s central point is that the robotics industry is still in the "capitalization" stage, or Phase 3, while capital is already pricing it as if it were in the "deployment economics" stage, or Phase 6. Between those two stages sit "mass-production validation" and "supply-chain profitability," and the article says neither has happened yet.
The author writes that over the next two to three years, the first real signal is unlikely to come from upstream components. It is more likely to show up in a single vertical use case where unit economics work and can be replicated.
Six clocks are moving at different speeds
The article breaks the question of where robotics stands into six separate clocks: technology, commercialization, primary fundraising, primary valuations, IPOs, and secondary-market earnings transmission. Its reading is that technology is still in the early-to-mid stage, commercialization remains early, primary fundraising is accelerating, primary valuations are already late-cycle, IPOs are in a concentrated realization phase, and earnings transmission into public markets has not started.
That distinction matters because, in the author’s view, mixing those clocks leads to serious timing errors. The piece says many of the stock-picking indicators now used in robotics — revenue mix from robots, customer concentration, ASP multiplied by volume, gross margin, capex, and inventory — are mature-stage metrics more suited to Phase 5. Before 2027 to 2028, those metrics often lack real data support. In Phase 3, the article argues, the more useful compass is IPO supply, lockup calendars, channel quality, and unit economics.
Fundraising is accelerating while valuations have already peaked
The article says the common description of the primary market as simply "overheated" and "late stage" is only half right. It argues that the primary market should be split into two lines: the fundraising cycle and the valuation cycle. Right now, those lines are moving in opposite directions.
According to the figures cited in the piece, domestic embodied AI financing reached 935 billion yuan in the first half of 2026, about five times the level in the same period a year earlier. The author says that does not look like a normal late-stage primary market. Instead, it looks like one of the hottest six-month fundraising periods on record. What has actually topped out, the article says, is the valuation cycle, because exit pricing has already been anchored by the secondary market. That combination is described as more dangerous than a standard late-stage market because it means new money is entering at top-of-cycle prices rather than old money exiting.
The distribution of capital is also described as skewed. In the first half of 2026, 169 full-machine companies took in about 33.2 billion yuan, or 72.2% of total industry financing. Downstream applications, by contrast, saw only 12 companies raise about 1.94 billion yuan, or 4.2% of the total. The article says the same dataset views downstream applications as the segment with the smallest valuation bubble and the highest commercialization certainty.
Unitree’s listing and the reset in pricing
The article uses Unitree Technology as a case study. It says Unitree listed on Shanghai’s STAR Market on Aug. 19, 2026. The stock opened up 629% on its first day, taking market capitalization to 444.9 billion yuan. On the same day, however, the STAR 50 Index fell 6.89% and the humanoid robotics sector sold off across the board. By the next day, Unitree’s market value had dropped below 280 billion yuan.
The author challenges the common claim that scarcity premiums in humanoid robotics will disappear only after three years, when around 10 companies have listed. The first two trading days in Unitree, the article says, are consistent with a much earlier timetable: capital may start being drained on day one, and valuation resets may begin on day two. At the same time, the piece is careful not to overstate the point. A single trading day cannot cleanly separate a Unitree-specific drain effect from a broader market decline, so the argument remains a hypothesis to be tested. The proposed method is event-study analysis on future listings such as Yunshenchu and Leju, measuring excess returns of a robotics basket against the STAR 50.
The article also points to float structure. At the start of trading, unrestricted shares accounted for only 7.44% of Unitree’s total share capital. A price formed with a 7.44% free float and 85% turnover, the author says, should not be treated as a valuation anchor. It is more a product of liquidity structure.
The piece then tracks Unitree’s first two months as a public company. On Aug. 19, the intraday high was 1,100 yuan, implying a market value of about 444.9 billion yuan, before the stock closed at 845 yuan. On Sept. 2, it fell intraday to 546 yuan, roughly half the peak. On Sept. 10, it dropped below 500 yuan. By Oct. 13, it was around 430 yuan, with market capitalization near 174 billion yuan, a drawdown of about 61% from the high. On Oct. 19, after policy support, the stock rose about 17% in a single day and returned intraday to around 500 yuan, or about 202 billion yuan in market value, still less than half the peak.
The article says the decline reflected both positioning and fundamentals. On the positioning side, the initial float was only about 7.4%, making the stock easy to push higher on sentiment and easy to push lower when sentiment faded. On the fundamentals side, the piece says Unitree’s non-recurring-items-adjusted net profit fell about 19.3% year over year in the first half of 2026, while gross margin slipped from about 60% to about 56%. It also says Zhiyuan overtook Unitree in shipment share in the first half, at about 44% versus 31%, and that Unitree was still trading at roughly 300 times earnings at the then-current price. The next lockup events are listed as February 2027 for offline placement shares and August 2027 for the employee stock ownership plan and part of the strategic placement.
The article says the public-market pullback quickly fed into issuance conditions. Citing September reports from Reuters and other media, it says regulators used informal window guidance to slow listing reviews for humanoid robotics companies. The focus, according to the article, was whether valuations were too high and whether revenue tied to local government projects represented real commercial demand. It also says Hong Kong’s market had become noticeably stricter on revenue authenticity and related-party transactions. There has been no public written suspension order, the article notes, and market descriptions range from a slower pace to a de facto temporary freeze. Still, the effects are already visible.
- Zhiyuan started its Hong Kong listing process in July, but a new pre-IPO financing round triggered Hong Kong Exchange rules, pushing the earliest possible listing to early 2027.
- Huandong Technology withdrew its STAR Market application on Sept. 21, becoming the first terminated robotics IPO of 2026. Its parent company said the reason was that the market environment had changed significantly.
- Hikrobot had its review suspended on Sept. 30 because its financial documents had expired.
- Leju’s prospectus showed that data-collection-related revenue accounted for 44.94% of 2025 revenue, and its largest customer was a shareholder affiliate, exactly the kind of revenue structure regulators are watching.
The primary market cooled at the same time. The article says humanoid robotics financing in September was about 10.5 billion yuan, down about 73% from 38.7 billion yuan in August. Capital shifted toward upstream segments such as chips, vision, joints, and dexterous hands, as well as leading companies that had already validated products. Startups with little more than demo prototypes found fundraising much harder. The article says some investors publicly stated that, after Unitree’s listing, their teams had stopped taking embodied AI projects to investment committee meetings. Even so, the author says this is not a full retreat. Zhiyuan still launched a new financing round in October at a valuation of about 50 billion yuan. The better description, the article argues, is that capital is repricing robot OEMs based on revenue quality rather than abandoning the sector.
Why robotics will not get an AI-style earnings pass-through
The article rejects a simple framing that treats AI as capex and robotics as opex replacement. For factories, it says, robots are usually capex as well, or they are deployed through leasing or Robotics-as-a-Service. The real difference is not accounting treatment. It is decision structure.
The key distinction is optionality. AI compute spending carries a FOMO element, so purchases can happen without strict ROI proof. Robots face a standing substitute: human labor. As long as people can still do the job, the burden of proof stays with the robot. Because of that, the article says investors should not wait for EPS transmission. They should wait for a vertical use case where unit economics work and can be repeated, with payback shorter than 18 to 24 months and evidence of repeat orders.
Supply plans are running far ahead of demand
On the question of overheating and duplicated buildout, the article turns to survey data. It cites research by an international investment bank covering nine A-share robotics supply-chain companies. According to the article, not one of those companies confirmed large, certain orders, and not one could provide a clear mass-production timetable. Yet their combined capacity plans corresponded to roughly 100,000 to 1 million robot-equivalent units per year. The same institution’s own forecast for global humanoid robot shipments in 2035 was 1.38 million units.
The article adds that China now has more than 150 humanoid-robot-related companies, with more than half of them startups or cross-sector entrants. Nearly 50 have launched mass-production models, and price competition is now broad-based. Profit on entry-level models has been compressed to below 5%, and some companies are taking orders below cost. The National Development and Reform Commission has publicly called for entry and exit mechanisms.
Even under the most conservative reading, the article says, supply planning is still running well ahead of current demand. Which historical analogy fits best depends on one variable: where subsidies are directed. In solar and electric vehicles, subsidies were aimed at the demand side, creating capacity while also pulling through real demand. In robotics, the article says, subsidies are mainly aimed at the supply side through local industrial funds, production-line subsidies, and innovation centers. That creates capacity, but not demand. The problem is compounded by the lack of product standardization. Capacity may be built before anyone even knows what specifications should be produced.
The author says consumer drones are a better primary analogy than solar or EVs. The reasons given are that data flywheels can intensify winner-take-most outcomes, integrated hardware and software mean cost advantage alone is not decisive, and scenario-specific deployment raises switching costs for early entrants. The expected outcome, in the article’s words, is that the industry may succeed while many companies die. Concentration could end up higher than in solar or EVs. Of the more than 20 companies now valued above 10 billion yuan, only a single-digit number may survive to Phase 5.
Overheating means different things at different layers
The article says overheating has opposite implications for robot OEMs and components. OEMs take the price war directly. Components benefit first from capacity expansion and then get hurt by the same price war. Compute, tactile sensing, simulation assets, and scenario operations are not directly exposed to whole-machine price competition.
To make that point, the author breaks the robotics stack into nine layers, from L0 to L8.
- L0 edge compute: The article calls this the only layer structurally similar to the AI compute chain. Demand does not depend on a "million humanoids" assumption because all robot forms need low-power, on-device multimodal inference. The disqualifier is vertical integration by OEMs. Tesla, the article notes, has already developed its own robot chip.
- L1 embodied foundation models: This is described as the highest-value technical layer and also the most expensive. Its valuation assumes winner-take-all dynamics, but the article says robotics lacks the unified benchmarks and interfaces that helped create that outcome in large language models.
- L2 data, simulation, and evaluation: The article says the real distinction is not real versus simulated data, but service versus asset. The question is whether the data can be sold a second time. Customized data-collection services are close to non-reusable, while simulation asset libraries and evaluation standards can be reused across customers.
- L3 deployment software: The author says this is the most attractive long-term layer and also the one most likely to die by arriving too early. Platform value depends on reuse across robot bodies, but every OEM is still building a full stack in-house. The article gives three entry signals and says two out of three would be enough: a de facto hardware interface standard, second-tier OEMs willing to adopt third-party stacks, and more than 1,000 robots deployed by a single customer. Teleoperation is described as the only investable subsegment right now.
- L4 robot OEMs: This is called the most crowded layer in the entire chain. Public-market pricing is driven by float structure and IPO supply, not by fundamentals.
- L5 components: The article says this layer must be split into five subgroups. The risk in actuators is not waiting for orders, but waiting for orders and still not getting margin. Dexterous hands are overheated. Force control and tactile sensing are defined by bottlenecks. Precision parts with demand independent of humanoids have real cash flow plus free humanoid optionality. Batteries and thermal management could move from ignored to bottleneck status when industrial use cases require 24-hour operation.
- L6 motion control and industrial robots: This is described as real cash flow plus humanoid optionality. In the two- to three-year gap between Phase 3 and Phase 5, the article says this is the only robotics exposure backed by actual earnings.
- L7 vertical deployment and RaaS: The article calls this the best non-consensus direction. The first real signal for the whole industry is expected to appear here. To tell the difference between a segment that has been unfairly sold off and one with a low ceiling, the author suggests checking whether a second company in the same use case can also raise funding.
- L8 safety, certification, standards, and insurance: This is described as the real gate to commercialization. The likely triggers, in order, are mandatory regulatory certification, implementation of ISO and Chinese national standards, insurance underwriting requirements, and, least predictable of all, a major safety accident.
The article also revises a common entry rule for components. A single threshold such as "robotics revenue above 10% of company revenue" can generate false signals, it says. If actuator revenue enters during a price war at margins below the company average, then a 10% robotics revenue mix is a dilution event, not a rerating event. The correct threshold should be two-part: robotics revenue at or above 10% of total revenue, and segment gross margin at or above the company’s blended gross margin.
Durable value appears only after deployment reaches customer economics
The article compresses the nine layers into four broader groups and reaches a direct conclusion: value becomes durable only when capability passes through the deployment layer and enters customer economics.
That is also where the current investment asymmetry shows up. The strongest narratives and the most crowded capital are in embodied foundation models, world models, and general-purpose humanoid hardware. The article says those areas top out at DE3 on the deployment ladder and show the worst mismatch between valuation and evidence. More attractive areas, in the author’s view, include evaluation and observability, deployment infrastructure, precision parts protected by time barriers, and vertical deployments with proven economics. Their shared feature is that value does not disappear depending on which OEM wins, and their duration can be matched to the right investment instrument.
In Phase 3 to Phase 4, the article says, deployment intensity and reliability explain cash flow better than model capability does. Yet the market is paying the highest premium for model capability.
How the article says investors should re-anchor
The piece lays out three investable lines: layers that do not depend on picking the eventual winner, narrow scenarios where ROI can be verified, and the rhythm of capitalization itself. It says two of the market’s most common approaches — broad bets on humanoid OEMs and 2030 market-size projections built bottom-up from BOM assumptions — do not fit any of those three lines.
The author then sums up the current setup in three statements. On timing, the primary market is trading 2030, the secondary market is waiting for 2027 earnings reports, and the variables that will actually move prices over the next 12 months are lockup calendars and IPO supply. On valuation, using Unitree as an example, the article says a segmented valuation based on the prospectus’s real products and end-market mix suggests that optionality accounts for about 95% of market value. The question is not whether 200 times sales is expensive. It is whether investors are willing to pay about 270 billion yuan for an option with an uncertain maturity and exercise conditions that may take two to 10 years to meet. On industry fundamentals, the article says humanoid robots have already proved they are useful. Success rates above 99% in narrow tasks are not stage performances. But the scale is still tiny. For the global shipment leader, revenue from humanoid robots actually used in smart manufacturing and inspection was 15.7 million yuan, about 2.6% of humanoid revenue.
The article also says investment opportunity is not the same thing as commercial opportunity. Precision parts are best suited to public-market allocation. Certification and compliance services look more like private equity. Transitional capacity and dual-source validation cash flow look more like bonds than options. If a fund mandate allows only early-stage equity, then the reachable opportunity set is smaller than the identified opportunity set. The author says that should be treated as an investment-committee mandate decision, not a footnote.
Seven false signals and the evidence the author wants instead
The article closes with an investment discipline for what it calls the capitalization phase of Physical AI: invest before earnings transmission becomes obvious, but only after deployment evidence becomes factual. It argues for a negative framework that refuses to pay for unsupported signals and instead waits for quantifiable, falsifiable evidence that would justify changing course.
- Reject shipment numbers without deployment; ask for human intervention rates. Shipment volume detached from real industrial and logistics usage has little value, the article says. A first disclosure of human intervention rates, such as under the YD/T 6770 standard, is a more useful measure of practical usability.
- Reject leaderboard chasing; ask for utilization and MTBF. The article says leaderboard performance in physical understanding has little to do with real visual experience and can change quickly. What matters is customer-confirmed utilization, mean time between failures, and SLA operating hours.
- Reject piling on degrees of freedom; ask for cross-year repeat orders. The article notes that the top seller with six degrees of freedom and the top-valued company with 42 degrees of freedom are serving very different markets. Reliability is proven only when the same customer places independent, incremental orders across multiple years.
- Reject low-quality government or data-collection orders; ask for buyer-calculated ROI. If more than half of orders come from government-funded data-collection and testing centers, then only ROI and payback figures published by the actual buyer, not the seller, count as commercial proof.
- Reject cosmetic offshore structures; ask for cross-site deployment cost. Offshore structures can hide risk but not remove it, the article says. Replicability is only proven when the deployment cost at the Nth site falls below 30% of the first site.
- Reject component narratives detached from earnings; ask for segment-level financials. The article says real earnings transmission begins only when leading component makers separately disclose robotics revenue and segment gross margin.
- Reject OEM and model valuations supported only by long-dated TAM; ask for software marginal effects. Market-size forecasts alone are still narrative, the author says. Scale effects are only validated when growth in reusable software revenue outpaces growth in delivery headcount.
The article’s overall conclusion is not that embodied intelligence lacks industrial value. It is that capital pricing and industrial stage are badly misaligned. Over the next 12 months, the author says, the key variables will not be shipment volume or financing totals alone, but lockup schedules, IPO supply, revenue quality, and whether a repeatable unit-economics model appears in a real vertical use case.
