Embodied Intelligence Industry Research: Clear Endgame, Delivery Not Yet Here — Cautiously Positive on Momentum, Anchored on Order Fulfillment

Industry Research · Initiation ｜ 2026-10-13

Core View (Cautiously Positive, conviction 2.5/5, horizon 12-24 months): The endgame of embodied intelligence is clear — labor substitution reinforced by strong policy pushes in both China and the US. Global humanoid robot shipments were approximately 13,000 units in 2025, and announced capacity by leading manufacturers for 2026 totals about 142,000 units; the momentum in shipments and orders is real. But three discount factors dictate caution today: First, the technological leap from "demo-grade to work-grade" (task success rate >90%, 8 consecutive hours without intervention) has been achieved by zero manufacturers, and zero manufacturers publish uptime; second, of the nominal order pipeline of approximately RMB 10.3 billion, about 73% carries policy/state-owned capital attributes, with genuinely recognizable revenue over the next 12 months of approximately RMB 2.7-3.5 billion (recognition rate 35-45%); third, the SW Robotics Index trades at a PE of roughly 130x, at the 82nd historical percentile — valuation compression has already played out in leaders (Unitree fell approximately 62% from peak within about 40 days of listing). Conclusion: the momentum is real, but return risk lies in valuation and delivery — for allocation, use "order/revenue announcement-grade fulfillment" as the anchor for single-stock screening, and avoid purely order-narrative names.

---

I1 Industry Definition and Scope

Embodied Intelligence, as studied in this report, refers to the industry of physical agents driven by AI "brains" with general-purpose task capability as the goal; the current form of core carriers is humanoid and general-task robots. The essential distinction from traditional pre-programmed industrial robotic arms and specialized service robots (floor cleaners, delivery lockers, etc.) is that "tasks are learned rather than pre-defined by engineering" — this is also why VLA (Vision-Language-Action) large models are seen as the industry's "brain transplant" variable.

The value chain has three segments:

• Brain layer: VLA foundation models + training compute and simulation toolchains (NVIDIA Isaac GR00T/Jetson Thor, Physical Intelligence π, AgiBot GO-1, Figure Helix, etc.). Commercial revenue is currently near zero; priced purely on fundraising.
• Body layer: humanoid/general-task robot OEMs (AgiBot, Unitree, UBTech, Tesla Optimus, Figure, Agility, 1X, etc.). Global shipments of approximately 13,000 units in 2025 (Omdia), with Chinese companies accounting for about 95% (IDC).
• Component layer: actuator assemblies (about 45-60% of unit cost), planetary roller screws, harmonic drives, six-axis force/tactile sensors, dexterous hands (about USD 9,500 per hand, roughly 17% of unit cost, per Morgan Stanley), frameless torque motors, rare-earth magnetic materials.

One-line map: the compute/model layer burns cash and sets the capability ceiling → the component layer determines the floor for mass production and cost → the OEM layer carries orders and valuation → the application side (government/enterprise procurement, factories, logistics, research) pays.

I2 Demand: Huge Endgame TAM, Currently in the Late Introduction Phase

2.1 Current Scale and Structure: Shipments Are Real, "Real Work" Is Scarce

Global humanoid robot shipments in 2025 were approximately 13,000 units (Omdia, 2026-01; GGII's figure is approximately 18,800 units, for cross-reference), of which AgiBot shipped approximately 5,168 units (39.8%), Unitree approximately 4,200 units (32.3%), and UBTech approximately 1,000 units (7.7%) — CR3 about 79.8%, all Chinese vendors; the China market was approximately USD 440 million (IDC). But structurally, genuine productive demand accounts for less than 15%: global industrial deployments are estimated at <2,000 units, with the rest being research/education (3,000+ units), commercial exhibitions/cultural tourism, data collection training grounds, and other introduction-phase demand (Analysys figures, third-hand, to be verified). The industry currently resembles "selling shovels to itself": governments and central SOEs buy robots first to build training grounds, producing real-robot data to feed back into models — industrial infrastructure rather than endgame demand.

2.2 Endgame Anchor: A 2.5x Divergence Among Institutions, Must Be Viewed Side by Side

The endgame narrative (labor substitution) has hard data behind it: US manufacturing employs about 12 million people with an annual wage pool of about USD 785 billion; ARK estimates 5.9 million units would fully substitute. MS estimates average unit prices falling from USD 200,000 in 2024 to USD 150,000 in 2028 and USD 50,000 by 2050. But the GS-Omdia divergence for 2035 is 2.5x; this report does not adopt either unilaterally, and the demand bridge uses a bottom-up approach.

2.3 S-Curve Position and Inflection Criteria

The current position ≈ EVs in 2015-2016: penetration (2025 shipments ≈ 2.4% of global annual industrial robot installations of 542,000 units), price (industrial-grade full-size at RMB 500k-1M), and demand quality (high share of policy/image projects) all indicate late introduction phase, yet capacity and orders are betting on the ramp phase. Four observable inflection criteria: ① industrial-grade full-size unit price <RMB 200,000 (corresponding to BofA's estimate of BOM falling to USD 13,000-17,000 by 2030-35 plus reasonable gross margin); ② single-scenario ROI payback <2 years (Figure's measured cost at BMW of about USD 25/robot-hour is already near the lower bound of fully burdened US labor hourly wages); ③ global annual shipments >100,000 units with genuine industrial deliveries >50%; ④ leading OEMs' humanoid business gross margin turning positive (UBTech still posted a net loss of RMB 790 million in 2025). Currently, none of the four has been achieved.

The inflection math chain (caveat: qualitative, to be quantified — both ends have hard numbers, but the middle "training capex → task success rate" lacks a quantified mapping, and VLA capability emergence cannot be linearly extrapolated): global AI investment exceeds USD 1 trillion in 2026 (GS, of which the US about USD 581 billion) → NVIDIA open-sources the GR00T N1 series and mass-produces Jetson Thor (per-unit "brain" compute cost drops to about USD 3,000-6,000) → price path USD 200,000 → 150,000 → 50,000 (MS), with China already probing RMB 39,900 (Unitree R1) → leading announced capacity totals about 142,000 units for 2026 → if the ROI<2-year criterion is met in 2027-2028, penetration enters a steep climb. The chain's biggest weakness is the third link: data. The Beijing Humanoid Robot Innovation Center (2026-06) judges that the efficiency of high-quality real-robot data collection is "far from the scale needed for industry takeoff" — 1 million trajectories require roughly 2,976 hours of continuous collection by 100 robots, while the industry currently has only a few thousand hours of high-quality data, a gap of about two orders of magnitude.

2.4 Demand Bridge: 2025→2030E Shipments from 13,000 to ~440,000 Units

This implies approximately 443,000 units in 2030 × blended ASP of about USD 40,000 ≈ USD 17.7 billion (range USD 15-20 billion), a shipment CAGR of about 103%. The bridge was revised down about 22% from the initial draft's 567,000 units: first, the China industrial item moved from top-of-range to midpoint; second, Tesla's 2025 pilot-run fulfillment measured only ~10%, so its 2026-2027 figures are discounted in tiers between the pilot phase and the dedicated-line phase. For reference, GS's 2030 figure is ~890,000 units (second-hand citation, to be verified) and IDC's ~510,000 — this bridge sits in the conservative half of the institutional range.

2.5 Demand-Side Risks

The sustainability of China's policy pulse is the single largest variable: 2025-2026 orders are heavily weighted toward data collection training centers, training programs, and cultural tourism/exhibitions; under the Ministry of Finance's debt-resolution constraints in 2026, if policy-driven procurement tightens under audit in 2027, the "policy scenario +120k units" line in the demand bridge faces a halving risk. The validating counter-signal is the relay from purely commercial orders: Agility has over USD 300 million in multi-year RaaS orders on hand, CATL has had nearly 100 units running on production lines for 3+ months, and Baida Precision has a thousand-unit-scale order — all have appeared but at small volumes (see I4 for details).

I3 Technology Delivery: The Distance from "Can Move" to "Can Work"

3.1 Technology Generations: Gen1→Gen2 Transition, Commercial Delivery Still Dominated by Gen1

Across the industry's three technology generations, we are currently transitioning from the first to the second: Gen1 (teleoperation + rules/pre-programming) is the absolute mainstay of current commercial delivery — 1X Neo publicly relies on "1X Expert" remote human intervention, and most domestic "industrial deployments" are in fact hybrid pre-programmed tasks and teleoperated data collection; Gen2 (end-to-end VLA) has reached the frontier but remains confined to controlled tasks — GO-1 achieves >60% success on complex tasks when trained on AgiBot World (100 robots, 1M+ trajectories), but success on unseen scenarios after domain randomization drops to about 42 (MIT TR China, citing research), and Helix and other vendors do not publish end-to-end numbers; Gen3 (world models + autonomous generalization) starts in 2025-2026. The hallmarks of "commercially viable work-grade" — task success rate >90%, 8 consecutive hours without intervention, and vendors daring to publish uptime — are currently achieved by zero players. Given the magnitude of the data gap and warehouse automation's historical pace (8-10 years from demo to reliable deployment), this report sets the crossover at optimistically 2027-2028, baseline 2028-2030, without making a single-point call earlier.

3.2 Key Module TRL Matrix

Structural conclusion: the "cerebellum" (locomotion) is good enough; the "brain" (generalization) and the "hands" (dexterous manipulation) are the bottlenecks — this determines the order of industry delivery: commercialize first in structured scenarios (warehousing, fixed production-line stations), then extend to semi-structured and unstructured settings.

3.3 Production Ramp: Nominal Capacity vs. Effective Output, an Order of Magnitude Apart

Industry nominal capacity in 2025 was ~120,000 units vs. actual shipments of 13,000 — capacity utilization ~11%; announced capacity for 2026 of ~142,000 units vs. absorbable demand of ~75,000 units — announcement fulfillment rate ~53%. The root of "paper capacity" is not assembly (line conversion takes months) but three upstream points: ① rare-earth magnetic materials — ~3.5-4kg per unit, China accounts for ~94% of global permanent magnet production, and the export licensing controls since April 2025 have already materially slowed Optimus's ramp (confirmed by Musk on an earnings call); ② planetary roller screws — fewer than 6 companies globally can mass-produce, and precision-grinding capacity expansion takes years; ③ the real-robot data loop — the ultimate constraint on whether the product "can actually be sold."

3.4 Cost Decline Curve

Average unit prices have fallen steeply along the learning curve: Unitree's humanoid ASP fell from RMB 593,400 in 2023 to RMB 260,700 in 2024 (-56%); Unitree R1 dropped from RMB 39,900 to RMB 29,900 (2026-06, -25%); research/demo-grade has broken below RMB 30,000. But two segments must be distinguished: research/demo robots (RMB 29,900-99,000) and industrial full-size work-capable robots (BOM ~RMB 250k-450k, price RMB 500k+) are two different markets — the former has already reached "tens of thousands of dollars," while the latter is estimated to need 2 more years to that milestone (GGII implies a global average price of ~RMB 190,000 by 2030; Tesla's V3 target price of USD 20,000-30,000 is a post-ramp vision). Three major cost levers: in-house joint development (Unitree's module self-sufficiency >90%), component domestic substitution (rapid localization gains in screws/coreless motors/six-axis force sensors), and scale; one major offset: rare-earth magnetic materials account for nearly half of motor cost and cannot be diluted by scale (China holds 91% of refining/separation; overseas NdFeB capacity totals ~53,000 tons by 2030 vs. China's 290,000 tons).

I4 Commercialization and Orders: The Industry's Ceiling and Floor (Headline of This Report)

4.1 Nominal Order Pipeline of ~RMB 10.3 Billion: 73% Carries Policy Attributes

Discounting to revenue: genuinely recognizable revenue over the next 12 months is approximately RMB 2.7-3.5 billion (China nominal ~RMB 7.8 billion × recognition rate 35-45%). The discount is grounded in hard receivables evidence: UBTech's total accounts receivable at 2026H1 were ~RMB 2.2 billion with bad debt provisions of ~RMB 545 million (a provision ratio of ~24.8%), and credit impairments up two orders of magnitude YoY — government/enterprise orders "growing revenue without growing cash" has moved from risk to fact. Even more notable is the regulatory characterization: in September 2026, the CSRC's window guidance on humanoid robot IPOs requires applicants to have "sustainable revenue and commercial orders" (of ~24 filing companies, only a few — possibly zero — are expected to qualify, per Reuters/CNBC), and the NDRC has previously publicly warned against industry overheating — regulators have already concluded that "hype is running ahead of reality."

4.2 Commercialization Maturity Score: Industry 38/100, "Demo-Grade Commercialization, Work-Grade Unproven"

Scored on order quality 30% + repurchase 25% + collections 25% + gross margin 20%:

Two extremes deserve elaboration: Agility (76) is the only player to have run a complete commercial loop of "pay-per-hour billing + repeat customer expansion"; its flaws are single-customer concentration and milestone clauses. Unitree (72) is the industry's only OEM with scale profitability (2025 revenue RMB 1.708 billion, +335%, gross margin 60.4%, adjusted net profit ~RMB 600 million), but its customers are mainly research/education and consumer entertainment — a genuinely cash-generative business, yet one that has not validated "work-grade" value. UBTech's 47 best illustrates the industry's condition: No. 1 in China by order value, but its collections score was pierced by the 24.8% bad-debt ratio.

4.3 Price Ladder and the 12-Month ASP Path

The current three-tier price ladder: research/demo units at RMB 30,000–100,000 (China) → industrial full-size winning bid/contract prices at RMB 470,000–800,000 (representative value ~RMB 600,000) → implied overseas full-unit pricing of USD 150,000–250,000. Three pricing mechanisms coexist: SOE centralized procurement tender prices (budget-based bidding, currently ~3/4 of order value), list-price direct sales (cost-plus + learning curve), and overseas commercial contracts (RaaS hourly billing benchmarked against human labor rates). Chinese OEMs hold pricing power — they sit at the far left of the cost curve, pricing at 1/3–1/5 of Western levels; the FCC ban keeps Chinese OEMs out of the US market, which objectively slows overseas price declines relative to China.

Baseline ASP path (China industrial full-size average winning bid): Price discovery is already ahead of the curve — the rental market is the most sensitive thermometer. Unitree G1 daily rental rates fell from ~RMB 15,000 in spring 2025 to ~RMB 3,899 on platforms by October 2026, and as low as RMB 289 on Xianyu (-98%). The "payback in two weeks" performance-rental economy has collapsed. Transmission to industrial full-size units:

Over the next 12 months, the full-size average winning bid will decline from ~RMB 600,000 to around RMB 450,000 (cumulative ~-22%, range -20% to -30%): State Grid's batch tendering, volume ramp from Unitree/Agibot squeezing out smaller players, and BOM declines along the learning curve will jointly pressure prices; commercial repeat orders rising to ~30% of the mix provide partial price resilience. Blended ASP will be roughly flat as industrial share rises (~USD 12,000–23,000 → USD 15,000–20,000). Rising volume with falling prices is the baseline picture: of the USD 17.7 billion value pool in 2030, "price" contributes negatively — all growth comes from "volume."

I5 Competitive Landscape and Profit Pools: Who Gets the Money

5.1 Concentration: OEM CR3 ≈ 80%, All Chinese Vendors

(Cross-reference: per China Electronics News, Unitree ranks first at 32.4% and Agibot second at 23.5% — rankings swapped, but the conclusion of "top three all Chinese" holds.) Signs of an unsettled landscape are equally evident: in 2025 alone, over 300 humanoid models were released in China; OEM offerings are heavily homogenized and market share can reshuffle at any time. True stability monopoly occurs at the component layer — Leaderdrive holds roughly 80–90% domestic share of harmonic reducers installed in humanoid robots.

5.2 Three-Layer Profit Pool: OEMs All Loss-Making, Real Money Elsewhere

• OEM layer (global ~USD 440 million, 2025, IDC): The entire industry is loss-making except Unitree — UBTech posted a net loss of RMB 790 million in 2025, Agility had revenue of only USD 1.8 million against an operating loss of USD 140 million, and Figure/Tesla Optimus generate no external revenue. OEMs are squeezed from both sides: weak bargaining power over downstream (government/SOE concentrated customers) with long receivable cycles; rigid costs upstream for core components and compute.
• Component layer: Robot-specific revenue is still small (Leaderdrive 2025 revenue RMB 571 million / net profit RMB 124 million, +121%; Tuopu's robot actuator revenue only RMB 13.59 million, 0.05% of total revenue), but the associated Tier-1 core businesses are massive cash cows (Sanhua 2025 net profit RMB 4.06 billion; Tuopu RMB 2.78 billion) — the only large-scale real profit today comes from the "traditional core business + robot option" structure.
• Brain layer: Commercial revenue ≈ 0, priced purely on fundraising (Figure at USD 39 billion post-money, Physical Intelligence ~USD 5.6 billion → rumored USD 11 billion+, Skild above USD 14 billion; compare Galaxy General's RMB 20 billion valuation against 150 units sold per year — a clear inversion).

Direction of profit migration: Over the next 2–3 years, profits will nominally migrate toward high-value components (actuator assemblies account for 45–60% of unit cost, dexterous hands 17%; if the Tesla chain ramps in 2027, components will be first in line to monetize). But two ongoing forms of erosion must be hedged: ① OEM in-house development — Unitree's module self-development rate is >90%, reducing reliance on Leaderdrive for upper-limb harmonic procurement; Laifual's per-unit harmonic cost fell 27% in one year (RMB 482 → 351); Leaderdrive's gross margin has slipped from 34.77% to 32.26%, and its 1-million-unit capacity start date has been pushed back from end-2026 to end-2028; ② Order authenticity — the Tesla chain's "~5,000-unit production orders / Sanhua's ~RMB 1.3 billion orders" remain entirely supply-chain hearsay with no company announcement confirmation (Musk personally clarified in September 2025 that similar "10,000-unit orders" were false), and Tesla's 2026 output will be primarily for internal deployment. External component supply is a "transition state," not an "end state" — the migration thesis holds, but the magnitude should be discounted by 40–50%.

5.3 Ecosystem War: Who Gets the "Android Moment"

The open-source camp (NVIDIA GR00T + π0/π0.5 + Agibot's "model + dataset + full unit" trinity) versus the closed camp (Tesla's FSD-style vertical integration, Figure Helix, and OpenAI confirming a restart of in-house hardware development in 2026). Judgment: the model layer will not converge within 2–3 years; what decides the outcome is not model open-sourcing itself but the real-robot data flywheel — ~95% of installed units are in China, meaning the data ecosystem naturally favors the "Chinese OEMs + NVIDIA toolchain" combination; NVIDIA, with its "compute + simulation + open-source models" full stack, is currently the player closest to the Android position. Implication for investors: winner-take-all uncertainty is highest in the brain layer, and certainty highest in the toolchain/compute layer.

5.4 Player Profiles

5.5 Valuation Position: 82nd Percentile + Contraction Already Materializing

The SW Robotics Index trades at 129.92x PE / 82nd historical percentile and 5.90x PB / 75th percentile (window since June 2014, as of 2026-10-12). The index is still diluted by auto parts; purer names are more extreme: Unitree at PE(TTM) ~292x, PB 59x, P/S ~100x; ~40 days after listing, market cap fell from a peak of ~RMB 444.9 billion to ~RMB 170.4 billion (~-62%) — valuation contraction is not a "risk disclaimer" but a reality in progress. The primary market is tightening in tandem: ~24 humanoid companies' HK IPO filings are stuck due to window guidance, and Figure's USD 39 billion valuation has lost its secondary-market reference. The combination of an upcycle × 82nd-percentile valuation is a fundamentally different investment proposition from an upcycle × 20th percentile: the sector's main tension in returns lies not in the direction of the cycle but in valuation digestion and the pace of realization. Key supply calendar to watch: Unitree's free float is only ~7.44%; the 2027-02-19 offline placement unlock and the 2027-08-19 IPO lockup expiry both fall within the 12–24 month window — supply shocks that can be timed to the day.

I6 Policy and Geopolitics: "Policy Content" Under Two-Way Interlocking

6.1 China: Unusual Intensity, but Clear Circular-Funding Characteristics

Three central-level pillars: ① The MIIT-led "Guiding Opinions on Innovative Development of Humanoid Robots" (November 2023) established the 2025/2027 milestones and the "brain–cerebellum–body" roadmap; ② The State Council's "AI+" action opinions (August 2025) and the eight-ministry "AI + Manufacturing" special program (January 2026) elevated embodied intelligence into national action; ③ The 2026 MIIT + SASAC real-scenario training special program — ≥20 key scenarios per province and ≥10 per central SOE, distilling one hundred high-value scenarios and driving ten-thousand-unit-level deployment by end-2026. Local subsidy map: Beijing's RMB 100-billion-scale government investment fund cluster (robotics sub-fund target RMB 10 billion); Shanghai's RMB 50 billion core industry target for 2027 (rewards of up to 5% of contract value, capped at RMB 5 million, for sales/leasing); Shenzhen's RMB 100 billion related-industry target for 2027; Hangzhou's per-project rewards up to RMB 1 million. Standards side: the first "Humanoid Robot and Embodied Intelligence Standards System (2026 Edition)" establishes L1–L5 intelligence grading and mandatory safety requirements, with the system to be completed by 2028.

The net stimulus is real (State Grid's RMB 6.8 billion / 8,500 units is the single largest order), but structurally it exhibits clear circular-funding characteristics of "funds invest in OEMs → state-capital platforms procure → scenario subsidies" — nominal orders propped up by fiscal money passing from one state hand to another; their sustainability depends on how fast purely commercial orders can take the baton.

6.2 United States: Dual Chokepoints on Market Access + Compute

On July 29, 2026, the FCC banned imports of new foreign-made humanoid robots on cybersecurity grounds (aimed at China); the FY2027 NDAA prohibits Pentagon procurement of humanoid systems from adversary nations; the DoD's Section 1260H list includes Unitree (extending from June 30, 2027 to products containing its components); the Section 232 robotics investigation remains pending. On the compute side, advanced GPU export controls continue, forcing Ascend 950-series domestic substitution and raising training costs for China's "brains." The actual impact on Chinese OEMs is limited (near-zero US export exposure; revenue and capital circulate domestically), so symbolism outweighs substance; the real tail risk is controls extending to the components and model/design-software layers.

6.3 Rare Earth Retaliation: The November 10, 2026 Pause Expiry Is a Bidirectional Volatility Source with Undetermined Direction

China imposed export licensing on 7 categories of medium/heavy rare earth items in April 2025 (still in effect); the October 2025 escalation measures (magnet technology export ban, extraterritorial controls) were suspended for one year from November 10, 2025. The 2×2 at expiry: If escalated controls resume — short-term reinforcement of Chinese OEMs' relative advantage (overseas magnet capacity cannot be filled before 2030; the DoD's equity stake in MP Materials with a price floor represents an accelerating de-China effort, though China holds ~94% of permanent magnet processing), medium-term acceleration of US supply chain decoupling; If the pause extends — overseas OEM ramps accelerate, benefiting the global volume bridge and weakening the "self-reliance" premium. The two directions have opposite implications for the judgment that Chinese components will "capture global incremental volume" — do not bet one-sided.

I7 Key Debates and Scenarios

D1: When Does VLA Cross the "Task-Level" Threshold? (The Technical Decisive Battleground)

• Bulls: GR00T N1.x open-source + Jetson Thor mass production drive down the marginal cost of the "brain"; State Grid/training-program investments of tens of billions are accumulating real-robot data; success rates on controlled tasks have reached 60–90%.
• Bears: No vendor has published uptime or end-to-end success rates; the crossing is "from unmeasurable to 90%+"; the real-robot data gap is roughly two orders of magnitude, implying data capacity must scale 100x over two years for a 2027–2028 crossing; the ICLR 2026 survey of 164 VLA submissions concluded "most VLAs still struggle with zero-shot generalization."
• Metrics to track: First published end-to-end unattended success rates and uptime by leading vendors; cross-scenario benchmark (EAI-type) scores; the pace of teleoperation data collection cost declines.

D2: Can Commercial Repeat Orders Take Over from Policy Orders? (The Demand Decisive Battleground)

• Bulls: State Grid's RMB 6.8 billion in batch tenders + central SOE training programs extended through end-2026; commercial repeat orders already emerging from Agility (USD 300 million), CATL, Baida Precision, Fulin Precision, etc.; carmaker pilot verification is advancing.
• Bears: Policy orders account for ~73% of nominal value — essentially a one-time infrastructure pulse, not recurring demand; UBTech's 24.8% bad debt rate exposes the "orders → cash" break; CSRC window guidance and NDRC overheat warnings signal an official attitude shift; under 2027 debt-resolution constraints, policy orders face tightening audits.
• Metrics to track: YoY policy-order winning bid amounts in 2027; pure commercial order share; changes in UBTech receivables/bad debt; State Grid tender fulfillment rates.

D3: Who Gets the Profit — OEMs, Components, or Toolchains? (The Landscape Decisive Battleground)

• Bulls: Actuators/dexterous hands account for ~60% of BOM; if the Tesla chain ramps in 2027, components are first in line to monetize; Chinese supply chain costs are 1/3–1/5 of Western levels; NVIDIA's toolchain position is the most secure.
• Bears: Unitree's >90% self-development rate and Laifual's 27% price cut show external components being eroded on both fronts by "in-house development + price wars"; the Tesla chain's 5,000-unit order lacks announcement-grade evidence; the FCC ban and potential component-level controls shrink the global addressable market for the Chinese supply chain.
• Metrics to track: Whether Sanhua/Tuopu orders receive announcement confirmation; Leaderdrive's share retention and gross margin; Unitree's in-house reducer progress; whether US controls extend to the components/model layer.

Scenarios (12–24 Months)

Catalyst Calendar

I8 Investment Implications and Beneficiaries

8.1 Cycle Direction × Business Quality Decision Matrix

Cutting through the value trap logic: a bullish cycle ≠ worth buying. The three points in the lower-right quadrant share one trait — "the cycle is up but profit/cash never reaches shareholders" — Tuopu's robot revenue is 0.05% of total while the stock has already priced in the 2027 ramp, and order rumors lack announcement-grade evidence; UBTech ranks first in order value but its 24.8% bad debt rate proves the "orders → cash" break; Agibot is #1 in shipments but loss-making, with IPO pricing front-running fundamentals. Unitree sits in the upper right (real cash flow + 60.4% gross margin) — genuinely high business quality — but its 292x PE and two unlock waves in 2027 mean "a good business ≠ a good price today"; it is a watchlist name during valuation digestion, not a chase target. Leaderdrive is in a quality downtrend (customer in-house development erosion + price wars), positioned near the midline, and is classified as watch-level.

8.2 Beneficiaries and Names to Avoid

Preferred allocations (balancing probability of realization and margin of safety under the base case):

• Sanhua Intelligent Controls: RMB 4.06 billion of net profit from its core business provides a cash-cow cushion, while robotics is a "free option" backed by the strongest positioning in the Tesla supply chain—the discipline being: do not treat the robotics thesis as the primary basis for holding the position until confirmed by order announcements;
• NVIDIA (global allocation): Compute + simulation + the full GR00T open-source stack—currently the closest positioning to an "Android moment," benefiting regardless of which OEM wins.

Watchlist (thesis valid, timing not yet ripe): Unitree (reassess after lockup expiry and valuation digestion), Leaderdrive Harmonic (track retention rate of customers' in-house sourcing share), Zhiyuan Robotics (after a reasonably priced IPO window emerges).

Avoid:

• UBTech: Before receivables/bad-debt issues are fully cleared, the revenue-growth-without-profit model continues to erode the financials (if bad debt is already fully priced in, upside would come from improving order mix, but maintain avoidance until verifiable improvement in collections appears);
• Theme plays driven purely by order rumors (exemplified by Topgroup, whose current robotics revenue is 0.05% vs. valuation pricing in 2027 volume ramp): until announcement-level orders materialize, realization risk and valuation elasticity are equally high;
• Long-tail homogeneous OEMs and rental/exhibition-dependent players: the price war across 300+ homogeneous models plus the collapse of G1 daily rental rates (-98%) mark the first casualties of industry consolidation.

8.3 Scenarios × Names

8.4 Key Risks

① VLA generalization fails to cross the "task-level" threshold for an extended period, leaving the industry long-term dependent on research/exhibition/policy procurement demand, with the S-curve flattening; ② Audit tightening on China's 2027 policy procurement, with policy orders falling off a cliff faster than commercial orders can take over; ③ Tesla Optimus V3 delayed again or the "5,000-unit order" falsified, reversing A-share supply chain expectations; ④ Escalation of two-way restrictions on rare earths/compute, pressuring supply chain costs and training capability simultaneously; ⑤ Downward shift of the sector's valuation anchor triggered by Unitree's lockup expiry; ⑥ UBTech-style bad debt becoming an industry-wide phenomenon, prompting a repricing of government-enterprise credit risk. For tracking of any of the above risks, anchor to the I7 catalyst calendar.

---

Conclusion reiterated: Embodied AI is a nascent-stage industry with "a clear endgame but unrealized payoffs"—cautiously bullish on the booming segments (cautiously bullish, conviction 2.5/5, 12–24 months), while remaining restrained on the sector's overall returns: only three areas currently generate real money—Unitree's C-end business, scarce components, and NVIDIA's toolchain—while the "boom" in other links is still mainly sustained by fiscal and financing lifelines. Allocation discipline: announcement-level realization of orders/revenue is the only credible signal for adding positions. For further interest in individual names (Sanhua Intelligent Controls, Leaderdrive Harmonic, Topgroup, UBTech, etc.), standalone dossiers with target prices and buy/sell-point analysis are available on request.