The Physical AI 2026 series consists of eight in-depth articles that look past the industry's capital boom to the economics of embodied AI as it moves from concept to deployment. From the trade-off between deployment cost and capability ceiling and the value mismatch between robot OEMs and core components, through deep dives into dexterous hands, data infrastructure and world models, to the evolution of geopolitical rules, the series systematically examines the real bottlenecks the industry faces on each dimension and the logic for breaking through them, presenting a full picture of a new era of embodied AI that is turning pragmatic and diverging.

The robotics industry is currently in the "capitalization" stage (Phase 3), yet capital is pricing it as if it were in the "deployment economics" stage (Phase 6), with two stages that have not yet happened, "mass-production validation" and "supply-chain profitability," lying in between. Over the next two to three years, the real signal will not appear first in upstream components, but in a single vertical use case: unit economics that work and can be replicated.

1. The six clocks are out of sync. Technology is in its early-to-middle stage, commercialization is early, primary-market funding is accelerating, primary-market valuations have reached a late stage, IPOs have entered a phase of concentrated realization, and secondary-market earnings pass-through has not yet begun. Mixing these clocks up leads to serious timing errors.

2. The funding cycle is accelerating while the valuation cycle has already peaked. In H1 2026, domestic embodied-AI funding reached RMB 93.5 billion, about 5x the same period last year; this is more dangerous than a "late-stage primary market," because new money is entering at top-of-market prices. After Unitree's listing this judgment began to play out: its share price fell about 61% from the high at one point, the IPO window tightened, and monthly funding in September fell about 70% month on month.

3. Do not wait for "AI-style earnings pass-through." The buyers of AI compute are a handful of hyperscalers, and their spending is defensive; the buyers of robots are thousands of plant managers and production-line owners, and they always have a ready-made alternative: people.

4. Overheating means opposite things for robot OEMs and for components. OEMs bear price wars directly; components first benefit from capacity expansion and later suffer from price wars; compute, tactile sensing, simulation assets and scenario operations are not directly exposed to whole-robot price wars.

5. Three main lines: layers that do not depend on "who wins," narrow scenarios with verifiable ROI, and capitalization cadence. The two things most of the money in the market is doing, broadly investing in humanoid OEMs and back-solving the 2030 market size from the BOM, are in none of these three lines.

Breaking "what phase is robotics in" into six independent clocks is the precondition for understanding every misalignment at present.

Figure 1　The industry is in Phase 3, while capital is pricing for Phase 6　Current readings of the six clocks

This chart reveals the weightiest methodological warning of this series: the mainstream robotics stock-picking indicators in today's market (such as robot revenue share, customer concentration, ASP × volume, gross margin, Capex and inventory) are mostly Phase 5 maturity-stage indicators, and before 2027–2028 they generally lack real readings to support them. In the Phase 3 stage we are in today, what actually applies is an entirely different observation compass: IPO supply cadence, the lock-up calendar, channel quality and unit economics.

The market generally describes the primary market as "overheated, entering its late stage." That is half wrong: the primary market must be split into two lines, the funding cycle and the valuation cycle, and they are now moving in opposite directions.

Figure 2　Funding is accelerating: companies valued at RMB 10 billion or more grew from 3 to 22 in a year and a half

Source: IT Juzi; H1 2025 funding back-calculated from the roughly 5x year-on-year increase

Far from being late-stage, the funding cycle is the hottest half-year in history; what has truly peaked is the valuation cycle, with exit pricing already pinned down by the secondary market. This combination is more dangerous than a "late-stage primary market": the latter means money is exiting, whereas the former means new money is entering at top-of-market prices.

The distribution of funds is equally unbalanced. In H1 2026, 169 complete-robot companies took about RMB 33.2 billion, 72.2% of all funding in the industry, making it the most crowded cell in the entire value chain; downstream applications had only 12 companies raising about RMB 1.94 billion, 4.2%, even though the same dataset considers this segment to have the smallest valuation bubble and the highest commercial certainty.

Figure 3　Seventy percent of funding flows to complete robots; less than 5% flows to downstream applications

Source: IT Juzi (H1 2026)

Unitree Robotics listed on the STAR Market on 19 August 2026 and opened up 629% on its first day (market capitalization reaching RMB 444.9 billion), but on the same day the STAR 50 fell 6.89% and the humanoid-robot sector fell across the board; the next day Unitree fell below RMB 280 billion.

Figure 4　The "drain" of the scarcity premium may begin on the very first day of listing

A common claim is that "the scarcity premium will disappear three years later, when ten humanoid-robot companies are listed." These two days of data are consistent with an earlier timetable: the drain may begin on the first day, and the valuation reset occurs on the second. It should be noted that a single trading day cannot distinguish a "Unitree drain" from a "systematic market decline"; this item is currently a hypothesis awaiting verification, and the way to verify it is to run the same event study on the listings of DEEP Robotics and Leju, computing the excess return of the robot basket relative to the STAR 50.

Also note the structure of the free float: at the early stage of Unitree's listing, unrestricted tradable shares were only 7.44% of total share capital. A price formed by a 7.44% free float and 85% turnover is not a valuation anchor, but a product of liquidity structure.

Two months of trading since Unitree's listing have confirmed the judgment that "the drain comes before the valuation reset." On 19 August, the first-day intraday high was RMB 1,100, market capitalization about RMB 444.9 billion, closing at RMB 845; on 2 September, the stock hit RMB 546 intraday, half of the high; on 10 September, it fell below RMB 500; on 13 October, it stood at about RMB 430, market capitalization about RMB 174 billion, a drawdown of about 61% from the high. On 19 October, on the back of favorable policy, it rose about 17% in a single day, returning intraday to about RMB 500 and a market capitalization of about RMB 202 billion, still less than half of the high.

Figure 5　Two months after Unitree's listing: about 61% drawdown from the high, and still less than half of the high after the rebound

The decline has both share-supply causes and fundamental causes. On the share-supply side, the free float at the early stage of listing was only about 7.4%, which makes the stock very easy to push up on sentiment and equally easy to fall when sentiment ebbs. On the fundamentals side, net profit excluding non-recurring items in H1 2026 fell about 19.3% year on year, consolidated gross margin fell from about 60% to about 56%, H1 shipment share was overtaken by AgiBot (about 44% versus 31%), and at the current price the P/E is still about 300x. The subsequent lock-up expiry points are February 2027 (offline-placement restricted shares) and August 2027 (employee stock ownership plan and some strategic placements).

The sell-off in the secondary market was quickly transmitted to the issuance side. According to media reports in September, including Reuters, regulators slowed the listing reviews of humanoid-robot companies through informal window guidance, focusing on whether valuations are too high and whether revenue tied to local-government projects represents genuine commercial demand; on the Hong Kong side, scrutiny of revenue authenticity and related-party transactions has also clearly tightened. To date there is no public written suspension order, and market characterizations range from "slowing pace" to "effectively a temporary freeze," but the results are already observable:

• AgiBot launched its Hong Kong listing process in July; because a new pre-IPO financing round triggered HKEX rules, it cannot list until early 2027 at the earliest.

• Huandong Technology (环动科技) withdrew its STAR Market application on 21 September, the first terminated robot IPO of 2026, with the parent company explaining that "the market environment has changed considerably"; Hikrobot's review was suspended on 30 September because its financial information had expired.

• Leju's prospectus shows that data-collection-related revenue accounted for 44.94% in 2025, with its largest customer being a related party of a shareholder, which is exactly the form of revenue that regulators are watching.

The primary market cooled in tandem. In September, humanoid-robot funding was about RMB 10.5 billion, down about 73% from RMB 38.7 billion in August; capital concentrated in upstream segments such as chips, vision, joints and dexterous hands, and in leading companies whose products have already been validated, while OEM start-ups relying only on demo prototypes found fundraising markedly harder. One investor stated publicly that, after Unitree's listing, the investor's team no longer takes embodied-AI projects to its investment committee. But this is not a full retreat: in October, AgiBot still launched a new round at a valuation of about RMB 50 billion. A more accurate description is that capital is repricing OEMs on revenue quality, not abandoning the track.

This implies that the pricing anchor in the capitalization stage is switching from "scarcity" to "revenue quality." Over the next 12 months, what determines OEM valuations is no longer shipment volume and funding amounts, but whether they can prove independent customers, replicable revenue and cash collection, which is exactly the evidence framework of Parts 02 and 03 of this series.

First, a common misstatement needs correcting: defining AI as Capex and robots as Opex substitution. This dichotomy is not accurate; for a factory, robots are usually also Capex (or leasing, or RaaS). The real difference lies not in the accounting line but in the decision structure.

Table 2　Decision-structure comparison: AI compute chain vs. robots

The key row is "optionality." AI capital expenditure carries a FOMO character, so it can do without ROI justification; for robots, by contrast, a ready-made alternative is always sitting right there: people. As long as people can still do the job, the burden of proof will always remain on the robot's side. So the anchor to wait for is not EPS pass-through, but unit economics working in a single vertical scenario and being replicable: a payback period shorter than 18–24 months, together with repeat purchases.

The question of "overheating and duplicated construction" needs to be answered with data, not impressions.

Figure 6　Planned supply-chain capacity is already close to the scale of global annual demand ten years from now

A survey by an international investment bank of 9 A-share robot value-chain companies shows that none confirmed a firm large order and none could provide a clear mass-production timetable; yet their capacity plans combined correspond to about 100,000–1,000,000 robot-equivalents per year, while the institution's own forecast for 2035 global humanoid-robot shipments is 1.38 million units. Domestically, there are more than 150 humanoid-related companies, over half of them start-ups or cross-industry entrants; nearly 50 have concentrated on launching mass-production models and have launched an all-out price war, squeezing entry-level model profit to within 5%, with some orders taken below cost; the NDRC has publicly proposed establishing admission and exit mechanisms.

Even under the most conservative reading, supply planning remains well ahead of current demand. Which historical analogy to use for judging the outcome depends on a single variable: which end the subsidy lands on.

Table 3　Historical analogies

Subsidies for solar PV and new-energy vehicles were applied on the demand side, creating capacity while genuinely pulling demand; robot subsidies are mainly applied on the supply side (local industry funds, production-line subsidies, innovation centers), which only creates capacity and does not create demand. Combined with the lack of product standardization, when capacity is built it is not even known what specification to produce. This report tends to use consumer drones as the primary analogy: data flywheels increase the degree of winner-take-all, hardware–software integration means a cost advantage is not enough to decide the winner, and scenario binding raises the switching cost for early entrants. The expected outcome: the industry succeeds, companies die in large numbers, and concentration is higher than in solar PV and new-energy vehicles. Of the 20-plus companies currently valued at over RMB 10 billion, those that can last to Phase 5 may be in the single digits.

This yields the most important inference: overheating means opposite things for different layers of the value chain.

Table 4　What overheating means for different layers

Breaking the robot value chain into nine layers and judging each layer in turn.

Figure 7　Crowding is concentrated in OEMs, actuators and dexterous hands; structural opportunity lies in compute, precision parts, the industrial installed base and vertical scenarios

Source: Assessment in this report; colors are shaded by primary-market judgment, represent only structural judgments of each segment, and do not constitute securities advice

Core logic of each layer:

• L0 On-device compute: The only layer in the value chain that is structurally similar to the AI compute chain; demand does not depend on the "million humanoids" assumption, since every form factor needs low-power on-device multimodal inference. The veto condition is being absorbed by OEMs' full-stack in-house development; Tesla has already developed its own robot-specific chip.

• L1 Embodied foundation models: Highest technical value, and equally the highest valuation. Its valuation assumption is "winner takes all," but robots, unlike large language models, lack unified evaluation standards and unified interfaces, so the mechanism by which winner-take-all forms is not yet in place. Consider only targets that combine full-stack capability with real-scenario data feedback.

• L2 Data, simulation and evaluation: Splitting into "real vs. simulated" is wrong; the true dividing dimension is "service vs. asset," that is, whether the data can be sold a second time. Custom data-collection services are almost non-reusable; simulation asset libraries and evaluation standards can be reused across customers. See Part 06 for details.

• L3 Deployment software: The most promising over the long term, and also the layer most likely to die from "entering too early." Platform value comes from reuse across robot bodies, yet every OEM is now doing full-stack in-house development. Entry signals, two out of three: a de facto hardware interface standard emerges; second-tier OEMs willing to adopt a third-party stack appear; a single customer's installed robot base exceeds 1,000 units. Teleoperation is currently the only investable sub-cell.

• L4 Robot OEM: The most crowded cell in the entire value chain. Secondary-market prices are dominated by free float and IPO supply and do not constitute a fundamental anchor. See Part 03 for details.

• L5 Components: Must be viewed as five cells. The risk in the actuator chain is not "waiting for orders that never come" but "waiting for orders and not waiting for margin"; dexterous hands are overheated; force control and tactile sensing are defined by bottlenecks; precision parts whose demand is independent of humanoids have real cash flow plus a free humanoid option; battery and thermal management go from "unnoticed" to a bottleneck when industrial scenarios require 24-hour operation. See Parts 04 and 05 for details.

• L6 Motion control and industrial robots: Real cash flow plus a humanoid option. In the two-to-three-year window between Phase 3 and Phase 5, this is the only robotics exposure with real profit support.

• L7 Vertical deployment and RaaS: The best non-consensus direction. The first real signal for the entire industry will appear here. The way to distinguish "wrongly sold off" from "limited ceiling" is to see whether a second funded company appears in the same scenario. See Part 02 for details.

• L8 Safety, certification, standards and insurance: The real gate to commercialization. The trigger paths, in order of probability, are mandatory regulatory certification, the landing of ISO and national standards, insurer underwriting requirements, and, least predictable, a major safety incident.

The component layer has an entry criterion that needs special correction. The common single-condition threshold, "robot revenue exceeds 10% of company revenue," gives false signals: if the robot actuator business comes in at a gross margin below the company average under price wars, "robot revenue at 10%" is a dilution event, not a re-rating event. The correct threshold must be dual-condition: robot revenue share ≥ 10%, and that segment's gross margin ≥ the company's consolidated gross margin.

Compressing the nine layers into four gives a more direct conclusion: value becomes durable only when capability passes through the deployment layer and enters customer economics.

Figure 8　Value becomes durable only when capability passes through the deployment layer and enters customer economics

The current investment asymmetry shows up here as well. The strongest narratives and the most crowded capital sit in embodied foundation models, world models and general-purpose humanoid complete robots; this layer's deployment-ladder ceiling is DE3, and the mismatch between valuation and evidence is the most severe. Relatively more attractive are evaluation and observability, deployment infrastructure, time-barrier precision parts, and verticals with validated economics, which have this in common: their value does not disappear depending on which OEM wins, and their duration can match the investment vehicle. In Phases 3–4, deployment intensity and reliability explain cash flow better than model capability, yet the market is paying the highest premium for model capability.

Three sentences to summarize. On timing: the primary market is trading 2030, the secondary market is waiting for the 2027 financial reports, and what truly determines prices over the next 12 months is the lock-up calendar and the IPO supply cadence. On valuation: taking Unitree as an example, a sum-of-the-parts valuation using the real products and end-market structure disclosed in the prospectus puts the option portion at about 95% of market value, so the question is not "is a P/S of 200x expensive," but "is one willing to pay about RMB 270 billion for an option whose expiry date is uncertain and whose exercise conditions require 2–10 years." On the industry: humanoid robots have already proven "useful," and a success rate above 99% on narrow tasks is not a stage performance; but the scale is extremely small: for the global shipment leader, revenue from humanoids actually used in intelligent manufacturing and inspection is RMB 15.7 million, about 2.6% of humanoid revenue.

Table 6　Four investment directions and vehicle forms

An investment opportunity is not the same as a business opportunity. The optimal form for precision parts is a secondary-market allocation, certification and compliance services are a PE form, and the cash flows from transitional capacity and dual-source qualification are more like bonds than options. If the fund's mandate allows only early-stage equity, the reachable opportunity set is much smaller than the identified opportunity set; this is a mandate decision that should be put in front of the investment committee, not a footnote.

In the capitalization stage of Physical AI, "invest before earnings pass-through becomes obvious, but only after deployment evidence has become fact" is the core rule for avoiding becoming the one who pays for the bubble. Investors should build a rigorous negative framework, firmly refuse to pay for seven kinds of "pseudo-indicators" that lack factual support, and take quantifiable, falsifiable counter-evidence as the sole basis for changing course:

1. Reject shipments without deployment; ask about the manual-intervention rate instead. Shipment volumes detached from their real industrial and logistics share carry no reference value; only a first-ever disclosed manual-intervention rate (e.g. under the YD/T 6770 standard) is the key measure of actual usability.

2. Reject leaderboard gaming; ask about real utilization and MTBF instead. On leaderboards, physical understanding and visual realism are essentially unrelated, and the top spot changes hands easily; what really matters is utilization confirmed by the customer, mean time between failures (MTBF) and SLA operating hours.

3. Reject stacked degrees of freedom; ask about cross-year repeat purchases instead. The shipment leader (6 degrees of freedom) and the valuation leader (42 degrees of freedom) face completely different markets; only independently delivered orders from the same customer, across years and growing batch by batch, can prove reliability.

4. Reject low-quality government and data-collection orders; ask about buyer-calculated ROI instead. When more than half of orders come from government-funded data-collection and testing centers, only ROI and payback periods published by the actual buyer (not the seller) carry commercial credibility.

5. Reject illusory offshore structures; ask about cross-site deployment cost instead. An offshore structure can only hide risk, not eliminate it; only proof that deployment cost at site N has fallen below 30% of the first site validates replicability.

6. Reject component narratives detached from earnings; ask for segment financial breakdowns instead. Stay wary of concept hype; only when component leaders report robotics revenue and segment gross margin separately has earnings pass-through truly landed.

7. Reject whole-robot and model valuations propped up only by long-range TAM; ask about software marginal effects instead. All pure market-size forecasts are narrative, not fact; only reusable software revenue growing faster than delivery headcount validates economies of scale.

Disclaimer: This article was written with the assistance of AI tools such as Claude Opus 5, ChatGPT-5.6 and Gemini 3.6 Flash. The author has made every effort to proofread it and ensure that the information is true and accurate, but omissions are still unavoidable, and readers' understanding is appreciated. Please note in particular that the content of this article is for information integration and academic/research exchange only; it does not constitute any investment advice and should not be regarded as a recommendation to buy or sell anything.