Unitree Robotics

Phase A — Understand the business

Lens 1 · Company Overview

Unitree Technology (宇树科技; Hangzhou Unitree Technology Co., Ltd.) designs and manufactures legged robots — quadruped "robot dogs" and bipedal humanoids — plus the actuators, motors, reducers, sensors and control stacks inside them. Founded 2016-08-26 in Binjiang, Hangzhou by Wang Xingxing (b. 1990, Ningbo), on a RMB 2M angel cheque from Yin Fangming. It is the global unit-volume leader in both legged categories and, per its own prospectus, has been GAAP-profitable since 2024.

Product line & 2026 pricing (the whole thesis is in this ladder):

• Quadrupeds — Go2 (consumer, ~$1,600), B2/B2-W (industrial/inspection), the original Go1/A1/Aliengo lineage. ~18,000+ units sold in 2025.
• Humanoids — R1 ($4,150–5,900, launched Jul 2025; 1.20m, 25kg, 26 DOF), G1 ($13,500–16,000 base to $74K EDU; 127cm, 35kg, 23–43 DOF, 3D-LiDAR), H1 ($90–95K; holds the 3.3 m/s humanoid running record), H1-2 (~$150K), H2 ($29,900 commercial / $40,900 EDU; unveiled 2025-10-22, shipping Apr 2026; 182cm, 70kg, 31 DOF). ~5,500+ humanoids sold in 2025.

How it actually makes money today: hardware unit sales, sold direct (shop.unitree.com, Tmall/JD in China) and through distributors/resellers, overwhelmingly to research labs and universities — the buyer that will pay for a capable, hackable, cheap platform without needing it to do autonomous production work. Humanoid revenue was 86.3% research/education in 2024 and still 73.6% in the first nine months of 2025. That single fact frames the entire investment debate: this is not (yet) a company selling labour into factories; it is selling development kits into a research boom. Contracts are one-off hardware sales — no take-or-pay, no meaningful recurring/SaaS line, and (per the prospectus's own risk language) real customer-concentration and payment-cycle exposure on the B-side orders that are starting to matter.

Early marquee customers (NVIDIA, Google research groups by 2020) gave it credibility and a Western research beachhead.

Lens 2 · Supply Chain

Upstream inputs → Unitree → end customer, named:

• Actuators / joint modules — made in-house. This is the crux. Unitree self-develops and self-manufactures BLDC/frameless torque motors, planetary and harmonic reducers, encoders, servo drivers, controllers, battery modules and LiDAR. Externally-sourced components are only 14–18% of bill-of-materials cost, and domestic-content across core parts exceeds 90%. A teardown of the G1 found no supplier logo on the motors — confirming in-house manufacture.
• The QDD design choice. Rather than buy expensive strain-wave/harmonic gearboxes (Harmonic Drive Systems 6324.T, Leader Harmonious 688017.SS, Shenzhen-area suppliers) that took decades to perfect, Unitree standardised on quasi-direct-drive (QDD) actuators using off-the-shelf planetary-gearbox designs — letting it spin a custom motor+gearbox subsystem "within weeks" versus 3+ months for Western peers. In-house motors reportedly cost 30–40% of an equivalent Western motor.
• Where it still buys: some reducer components from Meihu Co. (which has begun mass delivery of joint-module reducers to Unitree); commodity semis, batteries, cameras, structural metals from the dense Yangtze-Delta ecosystem. Compute for the "brain" leans on external AI silicon (historically NVIDIA Jetson-class edge modules on research units) — a genuine external dependency and a geopolitical soft spot (see Lens 10).
• The ecosystem moat around the company: ~3,000 drone-component suppliers left over from DJI's boom, ~200 Chinese humanoid firms creating supplier density, and suppliers "hours away by train" with same/next-day samples. This is the "Unitree chain" that Chinese equity research now tracks alongside the "Tesla chain" and "Huawei chain".

Chokepoints / single-source dependencies: (1) edge AI compute — the one input Unitree does not vertically own and where US export controls bite; (2) its own factories — vertical integration means a fire/quality/labour event at the Hangzhou base has no second source; (3) high-end force/torque sensing and dexterous-hand components at the frontier. This lens is the strongest part of the bull case: the supply chain IS the moat, and it is a China-domestic moat that a US or EU assembler literally cannot buy its way into without years of restructuring.

Lens 3 · Competitive Advantages (moats)

Durable moats (real):

• Cost, via vertical integration — the deepest moat. Humanoid manufacturing cost fell from $10,800 to $9,200 (2022→mid-2025) while humanoid ASP fell from RMB 593K (2023) to RMB 168K (9M2025) and the company stayed profitable. Competitors assembling from third-party actuators + reducers (>70% of industry cost structure) "cannot match that arithmetic without years of supply-chain restructuring". G1 at $16K undercuts even Tesla's long-run target of $20–30K for Optimus, and is 1/10th of Agility's Digit ($250K).
• Scale & volume flywheel — 5,500 humanoids = 32.4% of 2025 global humanoid unit shipments; Chinese firms took >80% of global installs (Counterpoint); Unitree + AgiBot are expected to capture ~80% of a China humanoid output that grows +94% in 2026. Volume feeds unit-cost, which feeds price, which feeds volume — the BYD/DJI playbook: own the bottleneck component, bootstrap through researchers/hobbyists, then scale into mass deployment.
• Brand / mindshare — the 2025 & 2026 CCTV Spring Festival Gala performances (16 H1 robots, Zhang Yimou–directed yangge dance) made Unitree the face of Chinese robotics to a billion+ viewers — a marketing asset no Western rival can buy.

Bargaining power: strong and growing over suppliers (it makes its own critical parts and cites "upstream bargaining power" in the prospectus). Weak-to-improving over customers — a research-lab buyer base is price-sensitive and non-sticky, and the prospectus itself flags B-side customer-concentration/price-cut risk.

Where the moat is thin (this matters):

• The AI "brain," not the body, is the real humanoid moat — and Unitree is behind. "The difference between a Unitree G1 and a Figure 03 is not hardware — it is the AI model". Figure's Helix and Tesla's Optimus stack map vision→action end-to-end; the G1 largely needs explicit programming per task, and real deployments assume ~100% teleoperation. A cost moat on the commoditising layer (hardware) while you are behind on the value-accruing layer (autonomy software) is the classic hardware-commoditisation trap.
• IP is thin — only 262 registered patents (20 domestic invention patents) as of 2026-01-31; the company admits it has historically relied on trade secrecy over patents, which it concedes "may constrain its ability to safeguard core technologies". A know-how moat in a country with ~200 humanoid rivals and high engineer mobility is leakier than a patent estate.

Lens 4 · Segments

Revenue by product line (main-business gross margins in brackets), all ``:

Geography: historically >55% international; in 2025 domestic China overtook exports for the first time, though exports still doubled YoY. The pivot to domestic is partly demand (China robotics boom) and partly a hedge against Western export/security friction (Lens 10).

Why the trend matters: the mix shift into humanoids is the growth story, but note the humanoid gross margin is compressing (87.7%→62.9%) even as volume explodes — the leader is trading margin for share on purpose. Blended GM is still rising only because the quadruped line is improving and the mix is moving to a still-high-margin humanoid line. Watch for the crossover where humanoid GM erosion outruns the mix benefit.

---

Phase B — Measure performance (+private overlay: Lens 5→Funding & valuation trajectory; Lens 7→Cap table + comps; Lens 8→funding/product/virality catalysts; +Traction & unit economics)

Lens 5 · Funding & Valuation Trajectory (private swap for "Earnings Result")

Operating record (prospectus-derived, ``, unaudited-to-me; RMB):

Sources: revenue/2024 net —; FY2025 GAAP net ~288M —; Q1'26 ex-items −52.5% —. H1 2026 revenue guided RMB 1.05–1.13B.

Conflict flagged (do not silently pick one): Humanoids Daily reports FY2025 net "600M+ RMB, +674%" while Caixin reports "288M." These are not the same metric. tanayj.com pins FY2025 at ~18% GAAP net margin (~RMB 308M) and ~35% "adjusted" — i.e. the ~RMB 600M figure is a Western-style non-GAAP adjusted number (adds back SBC / one-offs), while the Chinese 扣非 (deduct non-recurring gains) number is lower than GAAP. Anchor on the conservative GAAP ~RMB 288M (~18%) and treat the Q1'26 ex-items −52.5% collapse as the cleanest read on core operating profitability. ``.

The headline of this lens: revenue is up 10.8× in two years, but core profit is already going backwards (Q1'26 ex-items −52.5% on revenue +68%). The prospectus itself names the cause: price competition from Tesla and domestic automakers/electronics firms. The leader is being squeezed by the very price war it started.

Funding history (``):

• 2016 angel: RMB 2M (Yin Fangming) → early VCs HongShan/Sequoia China, Matrix Partners, Shunwei Capital.
• Series C, June 2025: pre-money ~RMB 12–12.7B ($1.7–1.75B), co-led by Alibaba, Tencent, China Mobile Capital, Ant Group, Geely Capital, Jinqiu Capital. The strategic-heavy syndicate (three of China's largest platforms + an automaker + the state telco) is a China-domestic-champion signal, not a crossover-fund/IPO-proximity signal in the Western sense.
• STAR Market IPO (2026): raising ~RMB 4.2B ($610–618M) for ≥10% (≥40.4M shares) → implied ~RMB 42B ($6.2B) valuation, with post-listing chatter toward ~RMB 50B ($7.4B). ~3.4× step-up from the June-2025 Series C mark in ~13 months.

Lens 6 · Founder Comms & Sentiment (private swap for "Earnings Calls")

No earnings calls exist. Proxy: Wang Xingxing's public posture across People's Daily, 36Kr, ChinaTalk and gala/CES appearances.

• Consistent message: cost-obsession ("cost maniac"), embodied-AI-is-the-next-platform, and — increasingly in 2025–26 — a pivot narrative from hardware to "robot brains." The IPO's flagship use-of-proceeds (RMB 2.02B / ~48% to an intelligent-robot-model project) is Wang publicly conceding the AI layer is where the company must now win.
• Tone shift over time: 2023–24 = triumphant hardware-cost story (viral demos, gala, records). 2025–26 = a subtly more defensive register — the prospectus and Wang's IPO-eve comments openly acknowledge margin pressure, Tesla/automaker competition, thin IP and R&D-intensity questions. That the company is pre-empting the bear case is a maturity signal — but also confirms the bears are pointing at something real.
• What he stopped saying: the pure "cheapest robot wins" framing is being quietly replaced by "cheapest robot with a capable brain wins" — an admission the hardware-only moat isn't enough.

Lens 7 · Cap Table, Secondary Marks & Listed Comps (private swap for "Comps")

Cap table / control (`):

• Wang Xingxing: ~23.82% economic, 68.78% voting via a dual-class structure (Class A = 10 votes/share; equal-voting restored for charter changes, independent-director appointments). Founder-controlled through listing.
• Meituan ~9.6%, Sequoia China/HongShan ~7.1%, Matrix Partners ~5.4%, plus the Series-C strategics (Alibaba, Tencent, Ant, China Mobile, Geely). "Half the investment circle owes Unitree gratitude" — an unusually broad blue-chip cap table.
• Secondary marks: pre-IPO A-share lock-ups + the ~3.4× Series-C→IPO step-up are the only marks; no active Western secondary market (A-share listing).

Listed comps — provenance-critical, multiples are `` or n/a:

Sources: UBTech/Ecovacs caps & fundamentals —. Read: at ~24× trailing sales and ~146× trailing GAAP earnings ``, Unitree would list at a steep premium to the profitable consumer comp (Ecovacs ~19× P/E) and at a valuation comparable to loss-making UBTech — justified only if you underwrite the 2026–28 volume/margin ramp. It is priced as the category-defining growth leader, not as a hardware manufacturer.

Lens 8 · Catalysts — Funding / Product / Virality (private swap for "Stock-Price Catalysts")

Events that moved perception/valuation (no ticker yet; these are the private-name equivalents of >5% moves):

• Jan 2025 CCTV Spring Festival Gala (16 H1 robots, Zhang Yimou) — global virality; widely called a "DeepSeek moment for robots"; the demand-generation engine.
• June 2025 Series C at ~$1.7B — institutional validation.
• Jul 2025 R1 launch at sub-$6K — broke the humanoid price barrier again.
• Oct 2025 H2 unveil; recurring kickboxing-tournament / kung-fu demos (teleoperated) — kept mindshare, but note teleoperated (ties to the autonomy gap).
• 2026 Spring Festival Gala encore — sustained brand.
• 2026-03-20 IPO filing → 2026-07-03 CSRC approval (a 73-day speed record) → listing as early as late July 2026 — the value-crystallising catalyst.

Pattern: the market reacts to (1) viral capability demos and (2) funding/IPO milestones — not yet to commercial-deployment revenue, because there barely is any. The re-rating risk cuts both ways: a credible autonomous commercial deployment would be the first catalyst of a genuinely new kind; a failed-demo or security headline is the downside mirror.

Lens 8b · Traction & Unit Economics (+private add)

• Deployments: ~250 G1s reportedly in industrial pilots in 2025 (one firm ~30 units, several with 5–6) — early, real, but tiny.
• Unit-economics claim: under full teleoperation, 15% service contracts, 2-yr life, 50–67% utilisation and zero residual value, a G1 pencils below the ~$30/hr cost of a human for light tote-handling (2–4kg) — i.e. economically viable today for a narrow task band.
• Hard ceiling: payload 2–3kg sustained (5kg for 10–15 min); the G1 overheats on strenuous work — ~10–15 min then 5–10 min cool-down; no autonomous task-learning at scale. This is the difference between "viral" and "deployed": the economics work only where the task is light, teleoperated and duty-cycle-limited.

---

Phase C — Judge people & books

Lens 9 · Management

• Wang Xingxing (Founder/CEO/CTO) — the archetypal founder-engineer. Built his first biped as a student for ~¥200; his quadruped master's work (Shanghai University, 2013) and the XDog viral video seeded the company. Track record: quantified and real — took Unitree from a 50m² office to global unit-share leadership and sustained GAAP profitability in a field where every Western peer burns cash. That is the single most impressive operating fact about this company: it is profitable while UBTech loses ~RMB 790M/yr and Figure/Agility/1X are pre-profit.
• Skin in the game / control: ~23.82% economic, 68.78% voting — total founder control through listing. Aligns incentives, but concentrates key-man risk (see Lens 10/13).
• Capital allocation: disciplined and cost-obsessed to a fault — the "cost maniac." The worry is the mirror image: R&D as a share of revenue collapsed 31.4%→17.8%→7.7% (2023→24→25) even as absolute R&D rose only RMB 30M→70M→~90M. For a company raising $300M to build an "AI brain," a 7.7% R&D ratio is startlingly thin — critics (TMTPost) explicitly ask whether Unitree "deserves" the raise on that spend profile. The IPO proceeds are, in effect, an admission that organic R&D wasn't enough.
• Red flags (management): thin patent/IP posture by choice; a demo-forward, PR-heavy culture (galas, kickboxing) that can outrun substance; and the Go1 security-backdoor episode (Lens 10) reflects a security-as-afterthought engineering culture — dangerous for a company whose entire export thesis depends on Western trust.
• Archetype implication: a brilliant hardware-cost founder now being asked to win a software/AI war — a different game than the one he has dominated. The bet is whether a hardware culture can build a frontier robotics-foundation-model culture fast, with 175 R&D staff, against Tesla and Figure.

Lens 10 · Forensic Red Flags & Regulatory

Accounting / disclosure flags (``, prospectus-derived, unaudited to me):

• Core-profit deceleration masked by a growth headline — Q1'26 ex-non-recurring net −52.5% while revenue +68%. The gap between the flattering GAAP/adjusted numbers and the ex-items number is the thing to watch; reconcile GAAP ~288M vs "adjusted 600M" carefully (Lens 5).
• Government support flatters the P&L — RMB 76M tax incentives (9M2025 alone) + RMB 32M direct grants (2022–9M2025). On a GAAP net base of ~RMB 288M, subsidies/incentives are not immaterial — a chunk of "profit" is state support that could normalise.
• Humanoid gross margin sliding (87.7%→62.9%) as ASP falls RMB 593K→168K — sustainable only if unit cost keeps falling faster than price; the Q1'26 print says it currently isn't.
• Revenue-quality / concentration — 73.6% of humanoid revenue is research/education; the prospectus flags B-side customer-concentration and lengthening payment cycles. Watch receivables/DSO post-listing.
• R&D-capitalisation & the 7.7% ratio — verify how much R&D is expensed vs capitalised once audited statements are public.

Regulatory findings (required sub-section):

• SEC (EDGAR EFTS — LR + AAER): Zero. Unitree has no CIK — it is private and not an SEC filer, so no EDGAR enforcement search is possible.
• US national-security / export track (material): In April 2025, security researchers Andreas Makris & Kevin Finisterre disclosed an undocumented remote-access backdoor in the Unitree Go1 — a pre-installed CloudSail tunnel (Zhexi Technology) with default creds pi/123 auto-connecting to Unitree servers; ~1,919 devices exposed, including units at MIT, Princeton, CMU; IEEE Spectrum warned of botnet risk. In response, a bipartisan group of ~24 members of Congress and the House Select Committee on the CCP (Chair John Moolenaar) called for Unitree to be investigated and potentially designated under NDAA FY21 §1260H (Chinese military company), added to the Commerce Entity List and the FCC Covered List; AUVSI backs the probe. This is a live, unresolved geopolitical tail risk — an Entity-List/1260H designation would sever US research/commercial demand and edge-AI-chip supply.
• Non-SEC (FTC/DOJ/FDA/CFPB): web search surfaced no US civil-enforcement action to date; the exposure is national-security/export, not consumer-protection.
• Item 3 (Legal Proceedings) equivalent: n/a — no 10-K exists; the prospectus's risk factors (price war, IP thinness, customer concentration, security/export) are the closest disclosure.
• Verdict: No accounting-fraud findings surfaced; the material red flags are (a) subsidy-supported, decelerating core profit and (b) an unresolved US security/export-designation threat stemming from a documented product backdoor.

---

Phase D — Project & stress-test

Lens 11 · IPO-Readiness & Path-to-Tradeable (private swap for "Forward Projection")

Readiness: 5/5 — S-1-equivalent filed, approval in hand, listing imminent. This is as close to tradeable as a private gets:

• Milestones cleared: IPO tutoring w/ CITIC (Jul 2025) → STAR filing (2026-03-20) → SSE listing-committee approval (early June) → CSRC registration approved 2026-07-03 → pricing/subscription now → debut as early as late July 2026.
• Vehicle: Shanghai STAR Market (科创板) — an A-share listing. Practical caveat for a Western book: not directly buyable. Foreign access is via Stock Connect eligibility (post-inclusion, not day one) or thematic ETFs (KraneShares KSTR / KOID) that may hold it — Unitree is a watch-and-express-indirectly name, not a place-an-order name.
• What unlocks the next re-rating (post-listing milestones to track): (1) research→commercial revenue-mix inflection (the 73.6% coming down); (2) blended & humanoid gross-margin stabilisation vs the price war; (3) a credible autonomous (non-teleoperated) commercial deployment at scale; (4) resolution (either way) of the US §1260H/Entity-List overhang; (5) the RMB 2.02B AI-model spend producing a shippable "brain."

No EPS forecast logged (forecast.ts skipped per --watchlist/private rules — no reliable share count or audited EPS pre-listing; a fabricated multiple would violate provenance). Directional base case ``: if 2026 humanoid units hit the 10,000–20,000 target and quadrupeds scale, 2026 revenue plausibly RMB ~2.7–3.5B on continued mix-shift — but with core margin compression, so profit growth likely lags revenue growth. The falsifiable claim: the ex-non-recurring net margin does not re-expand above its 2025 level through 2026 unless the commercial-mix and AI-brain milestones land.

Lens 12 · Bull vs Bear

Bull case. Unitree is the BYD/DJI of robots: it owns the bottleneck (actuators/motors/reducers), has an unassailable China-domestic cost moat (>90% domestic content, in-house motors at 30–40% of Western cost), is already profitable and the unit-volume leader (32% global humanoid share) while every Western peer burns cash, and it is riding a state-backed China robotics wave with a blue-chip cap table and a viral consumer brand. As embodied-AI models mature, a company that can ship capable bodies at $16K in volume is positioned to convert the coming "robot-labour" TAM the way BYD converted EVs. The $6.2B IPO could look cheap in five years if the 75,000-humanoid/yr capacity fills.

Bear case (2–3 things that could permanently impair it). (1) The moat is on the wrong layer. Hardware commoditises; the durable humanoid moat is the autonomy brain, where Unitree is behind Figure/Tesla and its R&D intensity has collapsed to 7.7%. It could win the body and lose the war. (2) A self-inflicted margin death-spiral. It started a price war it cannot fully control — Q1'26 core profit −52.5% — and China output +94% with ~200 rivals means structural oversupply; a hardware maker in a commoditising, oversupplied category earns hardware multiples, not AI multiples. (3) Geopolitical severance. A US §1260H/Entity-List designation (live, post the Go1 backdoor) would cut off Western research/commercial demand and edge-AI-chip supply in one stroke.

Pre-mortem (18 months out, thesis broke): it's early 2028. Unitree still ships the most humanoids, but they are still teleoperated demo/research units; the promised "brain" underdelivered against Figure/Tesla; the price war compressed blended GM below 45%; core profit went negative ex-subsidy; and a US Entity-List listing killed the export/research beachhead. The stock de-rated from an AI multiple to a Chinese-industrials multiple. Body: won. War: lost.

Are multiples too high? At ~24× sales / ~146× GAAP earnings, yes — unless the commercial-autonomy inflection is real and near. You are paying an AI-platform price for what is, today, a profitable-but-price-warring hardware company.

Contrarian view (what the market is refusing to see): the bull consensus fixates on the $16K price as the moat. The market is under-weighting that cheap hardware is exactly what commoditises, that 73.6% research/education revenue is not a labour market, and that Unitree's own R&D-intensity collapse is a tell that it may be harvesting the hardware lead rather than funding the software leap. The most valuable line in this dossier: the price that made Unitree famous is the same price that caps what it can earn.

Lens 13 · Devil's Advocate (short-seller)

Dismantling the bull case:

• Revenue concentration: 73.6% research/education — a grant-funded, cyclical, non-sticky buyer base. If China's robotics-subsidy/venture wave cools (it is already cooling — "hype cools, profit plunges," SCMP), the core customer evaporates. This is closer to selling dev kits into a hype cycle than selling labour into production.
• The moat is weaker than bulls think: cost leadership on commoditising hardware + a 262-patent, trade-secret-reliant IP posture in a country with ~200 humanoid firms and high engineer churn = a moat that leaks. The parts of the value chain that will hold value (autonomy models, dexterous manipulation, safety-certified OS) are exactly where Unitree is behind (SemiAnalysis: "not the full package"; Agility has the better OS/functional safety).
• Most dangerous competitor bulls underestimate: not Tesla — the Chinese fast-followers (AgiBot/Zhiyuan, UBTech, Xiaomi, the automakers) who share the same domestic supply chain and can copy the cost structure while a cash-rich automaker eats the price war. Unitree's cost moat is a China-industry moat, not a Unitree moat.
• Worst capital-allocation / governance signals: 68.78% founder voting dual-class (minority-holder risk), subsidy-dependent profit, and a 7.7% R&D ratio while marketing an AI pivot — you are asked to fund the R&D the company chose not to self-fund.
• What must hold for the price: ~24× sales requires 2026–28 volume and margin to compound. Q1'26 (−52.5% core profit) says margin is already breaking. If 2026 revenue disappoints 20–30% or GM keeps sliding, the equity halves.
• Single permanent-impairment scenario, and plausibility: US §1260H/Entity-List designation — plausible (active bipartisan push, documented backdoor), and it would sever demand + edge-compute simultaneously. Second: a safety incident with a teleoperated/autonomous humanoid in a public/industrial setting triggering a regulatory clampdown.

Lens 14 · Management Questions (ordered by information value)

1. Ex-subsidy and ex-non-recurring, what was 2025 core operating margin, and given Q1'26's −52.5%, at what humanoid unit price do you stop cutting?
2. What share of your 2025 humanoid revenue was autonomous (non-teleoperated) commercial deployment vs research/education/demo — and what is the 2026 plan to move it?
3. You are raising ~$300M for a robot-model "brain." What is your concrete plan to close the gap to Figure's Helix / Tesla's end-to-end stack, and why is 175 R&D staff enough?
4. R&D fell to 7.7% of revenue. Was that harvesting a mature hardware lead, and what is the target ratio post-IPO?
5. How do you defend cost leadership when AgiBot, UBTech, Xiaomi and the automakers draw on the same domestic supply chain?
6. What is your response plan if the US designates Unitree under §1260H / Entity List — and how exposed are you to non-Chinese edge-AI compute?
7. The Go1 backdoor: what changed in your security engineering and governance so it cannot recur on G1/H2?
8. Humanoid gross margin fell 87.7%→62.9%. Where does it bottom, and what is the floor unit cost given thermal/DOF limits?
9. With only 262 patents and a trade-secret posture, how do you protect core tech as engineers move to ~200 rivals?
10. What is real, repeat-order commercial demand (not one-off research sales) — name your largest non-research customers and their re-order rate.
11. On the thermal/duty-cycle ceiling (10–15 min work / 5–10 min cool-down), what is the roadmap to a full-shift industrial robot?
12. How much of 2026–28 volume is captive/affiliated (Alibaba, Geely, China Mobile ecosystem) vs open-market?
13. What is the dual-class sunset (if any), and what protects minority holders given 68.78% voting concentration?
14. Post-listing, what capacity utilisation do you need on the 75,000-humanoid line to stay GAAP-profitable through a price war?
15. What is your succession / key-man plan given the founder is CEO, CTO and controlling shareholder?

---