The Global State of Robotics

§00 — Executive summary

1. Robotics is two industries wearing one name. There is the boring one — 542,000 industrial robots installed globally in 2024, a 4.66 million-unit operational stock, roughly 200,000 professional service robots sold, ~20 million consumer units (IFR World Robotics 2025). And there is the loud one — humanoids, which shipped somewhere between 13,000 and 18,000 units in 2025 on roughly $440M of revenue (IDC, Jan 2026). Humanoid revenue is well under 1% of the robotics industry's total and ~95% of its narrative.

2. China won the volume war before the West noticed there was one. Chinese firms shipped roughly 90% of the world's humanoid units in 2025 (The Diplomat, Mar 2026). China installed 54% of all industrial robots in 2024 and, for the first time, domestic Chinese suppliers outsold foreign ones at home — 57% share, up from 47% a year earlier (IFR, Sep 2025).

3. The US won the capital and model war. Figure AI carries a $39B post-money valuation on a still-forming commercial book. Skild AI raised $1.4B at $14B+ (Jan 2026) without building a robot. Physical Intelligence was reported in talks at $11B+ (Bloomberg, Mar 2026). Core VLA architectures — π0, OpenVLA, RDT — originated in North American labs.

4. Both facts can be true, and they describe different bets. China is betting that whoever ships the bodies collects the data and rides the cost curve down. The US is betting that the brain is the durable asset and bodies commoditize. Nobody has been proven right yet.

5. The money is unprecedented and the definitions are unreliable. 2026 year-to-date robotics venture funding: $18.8B (Crunchbase) or $55.8B (Dealroom). Same year, same planet, 3× spread. Any report quoting one number without the other is selling something.

6. Real deployments exist, and they are narrow. Figure 02 logged ~1,250 hours at BMW Spartanburg across 30,000+ vehicles. Agility's Digit moves totes for GXO and Spanx. UBTech (优必选) delivered several hundred Walker S2 units into BYD, Geely, FAW-VW, Foxconn and SF Express plants. Spirit AI's (千寻智能) Moz1 runs a battery-PACK station on a live CATL (宁德时代) line — the strongest productive-deployment claim in China. Every one of these is a structured environment doing a repetitive task.

7. The IFR — the industry's own statistical body — is publicly skeptical. In May 2026 it stated plainly that humanoid capability in real production remains at demonstrator and pilot stage, and that mass adoption as universal factory helpers or household assistants will not happen in the near or medium term.

8. Tesla is behind its own story. Musk conceded in January 2026 that zero Optimus units were doing useful work in Tesla factories. V3 body reveal slipped to late July/August 2026; he declined to give a 2026 production target on the Q1 call.

9. The three Chinese firms you asked about are three different theses. Unitree (宇树科技) = cheap bodies at volume, now profitable, IPO registration approved 2026-07-02. AgiBot (智元机器人) = fastest ramp on earth, 10,000th unit by 2026-03-30, HK IPO targeted Q3 2026. Galbot (银河通用) = brain-first, best-funded, ships fewer than 150 robots a year.

10. The binding constraint is not intelligence. It is physics, data, and magnets. Battery runtime (Digit: ~90 minutes), dexterous manipulation under uncertainty, six-axis force/torque sensor supply, and China's ~90% grip on rare-earth magnet processing (McKinsey, Apr 2026). Removing Chinese suppliers from an Optimus Gen 2 bill of materials was estimated to push it from ~$46K to ~$131K.

§00.5 — Update log (v1.5 · 2026-07-27)

Dated developments since v1.4, each woven into the relevant section below: WAIC 2026 (Jul 17–20) put 400+ humanoid models on the floor and produced this report's best external validation — China's own national-center chief scientist describing the US–CN split exactly as §07's stack map does. The GUARD Act passed the House inside the NDAA (Jul 22), naming Unitree and covering control software — §7.3 rewritten. Hyundai announced the full buyout of Boston Dynamics (Jul 16, ~$325M for SoftBank's residual, implying ~$3.3B) with a 2028 Georgia deployment plan at up to 30,000 Atlas per year. Unitree's registration went effective Jul 6; pricing still pending as of Jul 27, and updated filings show Q1 2026 adjusted net profit down 52.5% YoY. AgiBot's Swancor vehicle is no longer a rumor — AgiBot-system executives chair and run the listco, whose stock ran ~2,000% (§5.3 rewritten). Tesla slipped Optimus production again to "anticipated later this year" (Jul 22 Q2 deck), V3 still unrevealed. The Bot Company added to §04; a seam layer added to §5.6; Matrix Robotics added to §5.5.

§01 — How to read the numbers in this report

This sector has a data-integrity problem, and it is worth being explicit about it before any figure appears.

Market size estimates for humanoids span an order of magnitude. For 2025–2026 alone, published figures include $2.1B, $2.9B, $4.89B, and $6.24B. These are not measuring the same thing. Some count only bipedal full humanoids; some include wheeled dual-arm mobile manipulators; some include the software and services attached. When someone tells you the humanoid market is $X, ask what a humanoid is.

Shipment figures for 2025 disagree by ~40%. IDC (Jan 2026) puts global 2025 humanoid shipments at ~18,000 units on ~$440M revenue, +508% YoY. Omdia implies ~13,000 by crediting AgiBot's 5,100 units with ~39% share. Unitree's own STAR Market prospectus claims 5,500+ humanoid units and a 32.4% global unit share, which implies ~17,000. All three cannot be right. This report uses "13,000–18,000" and names the source wherever precision matters.

Venture funding totals depend entirely on the definition of "robotics." Crunchbase counted $18.8B globally in the first half of 2026; Dealroom, using a broader physical-AI definition, counted $55.8B over the same period. PitchBook counted $27.6B for all of 2025. These are three different universes.

Source tiers used here. Where a figure carries real weight, its provenance is stamped inline:

• Tier 1 — Regulatory filings, IFR World Robotics, IDC, company disclosures, established wire services (Reuters, Caixin, Bloomberg). Use freely.
• Tier 2 — Sell-side and consultancy research (Morgan Stanley, McKinsey, TrendForce, IDTechEx, Omdia), think tanks (MERICS, Jamestown, The Diplomat). Directionally reliable, methodologically opaque.
• Tier 3 — Vendor-published "state of the industry" reports, ETF-sponsor commentary, robot e-commerce blogs. Often the only public source for a specific number, and often commercially motivated. Flagged where used.

A specific caution: several widely-circulated figures in this sector — the ¥1 trillion (~$138B) Chinese state robotics fund, the "$26B in municipal humanoid funds," Galbot's reported $99M 2025 revenue — trace back to single sources that other outlets then cite as established fact. They are reproduced below with that caveat attached, not laundered.

§02 — The taxonomy, and where the money actually is

2.1 Industrial robots (工业机器人) — the real economy

Fixed-base articulated arms, SCARA, delta, and cartesian systems. Welding, painting, palletizing, pick-and-place, machine tending. Boring, profitable, and roughly 100× the size of the humanoid market.

Two structural facts matter more than the headline. First, electrical/electronics (129,000 units) overtook automotive (126,000) as the largest customer industry in 2024 — the auto cycle no longer sets the tempo. Second, Chinese domestic suppliers crossed 50% share of their home market for the first time, hitting 57% (from 47% in 2023, ~30% in 2020). In metal and machinery, domestic Chinese share reached 85%.

IFR forecasts 575,000 installations in 2025 and 700,000+ by 2028.

2.2 Professional service robots — the fast grower nobody covers

~200,000 units sold in 2024, +9% (IFR, sample of 294 suppliers — not a projected total).

Within medical: rehabilitation and non-invasive therapy +106%, surgical +41%, diagnostics/lab automation +610%. The Robot-as-a-Service (RaaS) fleet grew 31% overall and 42% in transport/logistics — the single most under-discussed structural shift in the industry, because it changes robotics from a capex sale to a recurring-revenue business.

2.3 Consumer service robots

~20 million units in 2024, +11%. Overwhelmingly floor cleaning and lawn mowing. Europe and APAC each +16%; the Americas declined 1% on robot-vacuum share loss. This is where robotics is already a mass-market consumer category, and it is almost entirely irrelevant to the humanoid conversation.

2.4 Quadrupeds and other legged platforms

The category Unitree built its business on before humanoids. Boston Dynamics' Spot (~$75K), Unitree's Go2/B2, Deep Robotics (云深处). Inspection, security, defense, research. Unitree's prospectus disclosed that humanoids overtook quadrupeds in its revenue mix during 2025 — 27.6% of revenue in 2024 rising to 51.5% in the first nine months of 2025. That crossover is a useful proxy for where the money moved.

2.5 Humanoids (人形机器人)

Bipedal full humanoids, wheeled dual-arm mobile manipulators (Galbot G1, AgiBot G-series), and upper-body torsos. As of April 2026, roughly 12 commercial humanoid platforms were available for purchase or structured lease — up from 3 in 2024 — priced $28K (torso-only) to $245K (full biped with onboard compute), with lease programs at $3,500–$8,000/month. (Source: SVRC "State of Robotics 2026" — Tier 3, vendor-published; the firm sells robots.)

Price floor as of mid-2026: Unitree R1 at ~$5,900; Unitree G1 at ~$13,500–16,000; 1X NEO at $20,000 or $499/month; Galbot G1 at ~¥700,000 (~$87K–97K); Agility Digit and full industrial bipeds at $150K–320K.

2.6 The brain layer — the newest and most contested category

Companies that build no robot at all. Physical Intelligence (π0, π0.7), Skild AI, Google DeepMind's Gemini Robotics, NVIDIA's Isaac GR00T. Their valuations now rival or exceed the hardware makers'. This is the crux of the entire investment thesis in §07.

§03 — Inside the humanoid: what the stack actually is

3.1 Actuation — 30–60% of the bill of materials

The dominant cost and the dominant chokepoint (McKinsey, Apr 2026).

• Harmonic / strain-wave drives — compact, high reduction ratio, zero backlash. Dominant in humanoid joints. Vulnerable to backdriving damage when a robot falls.
• Cycloidal (RV) drives — high torque, high reliability, higher cost. Used in large joints. Nabtesco (Japan) is global #1.
• Planetary, belt, cable/tendon — cheaper, shorter lifespan, common in low-cost platforms. 1X's NEO uses tendon drive specifically for compliant, human-safe motion.
• Roller screws and linear guides — the linear-motion analogue, and arguably a tighter constraint than rotary reducers because qualification cycles are longer.

Cost split within an actuator, per McKinsey: reducer is the largest single element; driver electronics 15–20%; motor 10–20%.

Supply concentration: Harmonic Drive (Japan) holds ~20–25% of strain-wave gearing globally. Leader Harmonious Drive (绿的谐波, 688017.SS) is the Chinese incumbent and a Tesla supplier. Sanhua (三花智控) supplies Tesla thermal/actuator subsystems. Schaeffler (Germany) has signed three humanoid actuator partnerships in five months — Neura Robotics, UK-based Humanoid, and China's Leju (乐聚) — and in May 2026 committed to supplying over 1 million joint actuators to Humanoid through 2031.

3.2 The magnet problem

High-torque actuators require neodymium-iron-boron (NdFeB) permanent magnets. China controls ~69% of global rare-earth mining and ~90% of magnet processing and refining capacity (McKinsey, Apr 2026, citing IDTechEx). Musk has publicly attributed Optimus production constraints to magnet supply.

MP Materials CEO Jim Litinsky's framing is the one to remember: a humanoid carries double to triple the magnetic content of an EV, because it has dozens of actuators rather than one or two traction motors. Humanoid mass production is, mechanically, a rare-earth demand event.

The SCMP estimate (Dec 2025) that building Optimus Gen 2 without Chinese suppliers would move the BOM from ~$46K to ~$131K is the single most important number in the US–China robotics debate. It is a Tier 2 estimate. It has not been independently replicated.

3.3 Hands, tactile sensing, and the dexterity wall

Degrees of freedom in the hand is the field's vanity metric. Tesla's V3 claims 22 DoF per hand and 50 actuators across both hands, up from 11. XPeng IRON claims 82 total DoF. Figure 03 has 16 DoF hands.

DoF is not dexterity. Dexterity is closed-loop force control with tactile feedback, and the sensor supply chain is thin: six-axis force/torque sensors are supplied by a handful of vendors (ATI Industrial Automation, OnRobot), are calibration-intensive, and remain largely un-automated in production. IDTechEx (Feb 2026) names dexterous hands and tactile sensing as the primary barrier to expanding humanoid task scope beyond simple industrial operations. Durability compounds the problem: The Information reported that in Tesla's package-sorting trials, hands lasted roughly six weeks at a unit cost above $6,000. The hand has consequently become its own investable layer — see §5.6.

3.4 Power

The unglamorous ceiling. Published runtimes:

Note the pattern: the wheeled platform gets 5× the runtime of the best biped. Bipedal balance consumes energy before any work is done. This is why the mobile-manipulator form factor (Galbot, Kinisi, AgiBot G-series) is quietly winning the deployments that need shift-length uptime.

Workarounds are architectural, not chemical: UBTech's Walker S2 performs autonomous hot-swap of its own battery in under 3 minutes using dual-arm coordination. Samsung SDI showed humanoid-specific solid-state cells at InterBattery 2026, targeting mass production H2 2027. LG Energy Solution is confirmed as battery supplier for Boston Dynamics' Atlas.

3.5 The brain — VLA models and the data scaling law

A vision-language-action (VLA) model ingests camera pixels and a natural-language instruction and emits motor commands through one network, replacing the classical perception → planning → control pipeline. Its economic significance: a VLA can be pre-trained across multiple robot bodies and fine-tuned per platform, which turns the robot brain into a software product with network effects rather than a hardware feature. That single architectural fact is why Skild AI can carry a $14B valuation without manufacturing anything.

The state of the art as of mid-2026:

• NVIDIA Isaac GR00T N1.7 (early access, 2026-04-17): 3B parameters, Cosmos-Reason2-2B backbone, Apache 2.0. Its central result is EgoScale — pretraining on 20,854 hours of human egocentric video. NVIDIA claims this establishes the first scaling law for robot dexterity: going from 1,000 to 20,000 hours of human egocentric data more than doubles average task completion. GR00T N2, previewed at GTC 2026 and slated for end-2026, uses a world-action-model architecture and claims >2× the new-task success rate of leading VLAs.
• Physical Intelligence π0.7 (2026-04-16): compositional behavior — recombining skills, transferring across robots and scenes via multimodal prompting.
• Figure Helix 02 (2026-01-27): full-body control; two robots running a single shared policy coordinating without a central planner or explicit messaging. Project Go-Big collects first-person human video across Brookfield's 100,000 residential units; Figure reports zero-shot transfer from 100% human video to robot navigation.
• Google DeepMind Gemini Robotics / ER 1.6: adapts to new tasks with 50–100 demonstrations. Apptronik's Apollo is its designated hardware platform.
• AgiBot GO-2 → GO-3 (GO-3 expected Q3 2026): behavioural foundation model plus GE-2 action world model, trained on the open AgiBot World dataset (1M+ trajectories, 217 tasks).
• Galbot AstraBrain (银河星脑): end-to-end integration of task planning, motor control and neural coordination in one architecture, trained on synthetic data from AstraSynth (银河星坊).

The data cost curve is the story here. Teleoperation data collection ran ~$340/hour in 2024 and ~$118/hour in 2026 — a 60% decline (SVRC, Tier 3). That puts a $50K–150K enterprise pilot data budget within reach for the first time. VLA models appeared in ~40% of new deployments in 2026, and imitation learning overtook reinforcement learning as the primary manipulation training method.

The counter-argument, from the research literature: a physical-AI study covered in January 2026 found the remaining bottlenecks are physical, not cognitive. Real-world trials are slow, expensive and damage hardware. Simulation is the substitute, but simulators fail to reproduce contact dynamics, friction, compliance and timing — so policies that pass in sim degrade sharply on hardware. The paper describes this gap as structural rather than incremental.

§04 — The United States

4.1 Figure AI — the capital-markets outlier

$39B post-money (Series C, Sep 2025, >$1B committed; Parkway led, with Brookfield, NVIDIA, Intel Capital, Qualcomm, Salesforce, T-Mobile, LG, Macquarie). Valuation path: $500M (May 2023) → $2.6B (Feb 2024) → $39B (Sep 2025). A 78× step-up in 28 months.

What is real: Figure 02 ran ~11 months at BMW Spartanburg, contributing to 30,000+ X3 vehicles and loading 90,000+ sheet-metal parts across ~1,250 runtime hours. The fleet was retired in Nov 2025 and Figure 03 deployed at the same plant (~40 units). Hyundai and Google DeepMind reserved 2026 BotQ capacity.

Capability evidence beyond BMW: in May 2026 Figure livestreamed three Helix-02 robots autonomously sorting 30,000 packages in 24 hours with zero failures — the run was extended for days, passing 100,000 packages — and staged a 10-hour head-to-head in which a Figure 03 sorted at 98.5% of a human's throughput. Staged conditions, single task, but sustained multi-day autonomy is not something any competitor has publicly demonstrated.

The production ramp is the most striking number in Western humanoids: Adcock disclosed monthly output of ~60 (Feb 2026) → ~120 (Mar) → ~240 (Apr), reaching one robot per hour at BotQ by May 2026 — a 24× throughput gain in under 120 days. BotQ's first-generation line targets 12,000/year, scaling to 100,000 over four years, with a supply chain built for 3 million actuators.

Figure exited its OpenAI partnership in Feb 2025 to vertically integrate AI. Helix is now developed entirely in-house.

The bear case, stated plainly: $39B is priced on a commercial book that consists of one automotive customer, two reserved-capacity partners, and a manufacturing ramp. A former head of product safety sued in Nov 2025, alleging termination after raising that the robots were strong enough to fracture a human skull. Figure 03 is not consumer-available; Adcock's stated goal is "select homes" by late 2026.

4.2 Tesla Optimus — the ambition-to-delivery gap

The clearest case of narrative running ahead of hardware.

• Jan 2025: Musk predicts ~10,000 Optimus built in 2025. Actual: a few hundred (independent reporting, Jul 2025).
• Jan 2026 (Q4 2025 call): Musk concedes units are "primarily for learning" — zero doing useful work in Tesla factories.
• Mar 2026 (Abundance Summit): V3 in "final stages," production start Summer 2026.
• Apr 2026 (Q1 call): production begins at Fremont late July or August. Output "quite slow," rate "literally impossible to predict" — 10,000 unique parts on an entirely new line. Musk declined to give a 2026 production target.
• May 2026: Model S and Model X discontinued after 14 and 11 years; their shared Fremont line is being dismantled and converted to Optimus assembly.
• Jul 2026 (Q2 shareholder deck, Jul 22): Model S/X lines decommissioned; first-generation Optimus lines being installed; production "anticipated later this year" — another slip from April's "late July or August." Musk: the ramp's S-curve "will be quite flat and long." V3 remained unrevealed as of late July, and initial builds go to an internal "Optimus Academy," not customers.
• Giga Texas: permit documents show a 5.2M sq ft Optimus facility, $5–10B, targeting 10M units/year from 2027. A 10-million-unit humanoid factory would be larger than any car factory ever built.

Tesla's structural advantages are real: automotive-scale manufacturing, in-house actuators, the FSD neural stack, Cortex 2 training compute at Giga Texas (>100k H100e installed, doubling in H1 2026), and the AI5 chip (taped out Apr 15, 2026; Optimus gets it first). Its structural weakness is equally real: no external customer. Figure has BMW. Boston Dynamics has Hyundai. Apptronik has Mercedes. Agility has GXO. Tesla has Tesla.

4.3 Agility Robotics — the only one with a boring revenue line

Digit, at GXO since June 2024 — widely considered the first humanoid deployment at a commercial customer site — plus Spanx, Mercado Libre (San Antonio), and Amazon testing. Hundreds of thousands of cumulative operating hours moving totes at ~35 lb payload. RoboFab targets 10,000 units/year. $641M+ raised. Agility targets sub-2-year payback against $30/hour labour, and was the first commercial humanoid to adopt NVIDIA's Halos for Robotics safety stack (announced 2026-06-22).

Agility's position is the least glamorous and the most defensible: it is the only humanoid company generating consistent revenue from productive commercial work.

4.4 Apptronik — the industrial pragmatist

Apollo. $520M Series A-X (Feb 2026) brought the Series A above $935M and total capital near $1B, at a reported ~$5.5B valuation — the round closed at 3× the original valuation. Backers include Google, Mercedes-Benz, B Capital, AT&T Ventures, John Deere, QIA, ARK.

Partnerships: Mercedes-Benz, GXO, Jabil (worldwide production partner), and Google DeepMind — Apollo is the designated hardware platform for Gemini Robotics. Modular field-replaceable joints, designed around MTTF and total cost of ownership rather than demo spectacle.

CEO Jeff Cardenas has been unusually explicit that the funding is intended to beat Chinese humanoids to market, and has publicly called for a coordinated US national robotics strategy.

4.5 Boston Dynamics — the incumbent, now fully Hyundai's

Electric Atlas. The entire 2026 production allocation is committed — to Hyundai's Robotics Metaplant Application Center (RMAC) in Georgia and to Google DeepMind, with additional customers slated for 2027. 50 kg payload. LG Energy Solution supplies the battery.

On July 16, 2026 — after SoftBank exercised its put — Hyundai announced the buyout of the residual 9.65% for ~$325M, making Boston Dynamics wholly owned for the first time at an implied ~$3.3B, essentially the 2021 price. The plan attached to full ownership: Atlas into the Georgia Metaplant from 2028 (parts sequencing first, component assembly by 2030), scaling toward 30,000 units/year — while Hyundai's Korean union struck partially in July over robot-driven job security, a preview of the labor politics the whole sector will import. Boston Dynamics also announced a strategic partnership with Google DeepMind at CES 2026.

4.6 1X Technologies — the consumer bet

NEO. Norwegian-founded (Moss), now Palo Alto-headquartered with a Hayward, CA manufacturing facility opened April 2026. $20,000 or $499/month; first US early-access deliveries in 2026. 66 lb, tendon-drive actuation, soft 3D-lattice polymer body.

1X is candid that NEO ships with limited autonomy and relies on teleoperation for complex tasks — a human-in-the-loop model in which early adopters share home data to train the policy. That is either the most honest data-acquisition strategy in the industry or the largest consumer-privacy exposure in it. Probably both.

4.7 The brain layer

• Skild AI — $1.4B Series C (2026-01-14) at $14B+, led by SoftBank with NVentures, Bezos Expeditions, Macquarie. Tripled valuation in seven months. Acquired Zebra Technologies' robotics division (Apr 2026). Thesis: one "omni-bodied" brain for any robot.
• Physical Intelligence — $600M at $5.6B (Nov 2025, CapitalG-led); reported in talks for ~$1B at $11B+ (Bloomberg, 2026-03-27). π0.7 released 2026-04-16. Explicitly the "Android of robotics."
• Google DeepMind — Gemini Robotics + Gemini Robotics-ER 1.6; hardware partners Apptronik and Boston Dynamics.
• NVIDIA — not a competitor so much as the substrate: Isaac Sim, Cosmos world models, GR00T N-series, Jetson Thor/T4000, Halos safety stack, and an Isaac GR00T Reference Humanoid (announced 2026-06-01) built from a Unitree H2 Plus body, Sharpa hands and Jetson Thor. NVIDIA claims 110 robot-brain developers in its ecosystem and has integrated Isaac/GR00T into Hugging Face's LeRobot.
• Meta — acquired Assured Robot Intelligence (May 2026), folded into Superintelligence Labs to train a foundational physical-AI model.

Note what NVIDIA's reference humanoid means: the canonical American robot-AI platform ships on a Chinese body.

4.8 The Bot Company — the anti-humanoid home bet

Kyle Vogt (Twitch, Cruise) is running the largest rejection of the humanoid form factor in US robotics: a non-humanoid household robot, ~$552M raised across three rounds in ~18 months to a $4B+ post-money (Eclipse-led $250M, Oct 2025), ~94 employees, and — deliberately — no public product. Vogt has cited bipedal balance, mass, and safety certification as fundamental barriers to consumer adoption; the company is, in effect, scenario B (§11) organized as a cap table. Reports of larger raise ambitions circulate secondhand and are unverified.

§05 — China

5.1 The policy architecture

Embodied intelligence (具身智能) barely existed in Chinese policy documents before 2023. In the 15th Five-Year Plan (十五五规划, 2026–2030) it occupies its own dedicated inset box among the plan's top ten "new industry tracks" (新兴产业赛道), alongside integrated circuits, biomanufacturing, and commercial space. "Artificial intelligence" (人工智能) appears more than 50 times in the 141-page outline.

The plan's language (Box 3, Item 02) is operational rather than aspirational: coordinate the layout of embodied-intelligence training grounds; promote virtual–real fusion collaborative training; develop integrated "big-brain / small-brain" (大脑/小脑) embodied models; tackle key technologies in the body and core components; accelerate deployment of humanoid robots. The Diplomat's reading — that this is a procurement directive, not a wish list — is the correct one.

The supporting stack:

Municipal subsidy race. Beijing's Yizhuang / E-Town (北京亦庄) aims to be a "global first-class embodied intelligent robotics industrial city." Shenzhen launched a ¥10B AI & Robotics Industry Fund. Shanghai published an Embodied Intelligence Industry Development Implementation Plan (2025–2028). Cities offer free land, subsidised rent, and buyer rebates of roughly 10% of each robot's sale price (Tier 2 — etcjournal / policy aggregation).

Q1 2026 financing (The Paper, via Jamestown): 210 embodied-intelligence financing events totalling >¥30B (~$4.2B). Shenzhen 44 deals, Beijing 40, Shanghai 38, Hangzhou 24.

What subsidy actually looks like at company level. Unitree's prospectus is the only audited window: ¥76M (~$11M) in tax incentives in the first nine months of 2025 alone — reduced corporate income tax rate, R&D super-deductions, VAT refunds — plus ¥32M (~$4.7M) in direct government grants between 2022 and September 2025. That is meaningful but not transformative against ¥1.7B revenue. The often-cited "¥1 trillion (~$138B) state-backed robotics fund" traces to a single secondary source and should be treated as unverified.

Patents. China registered 5,688 humanoid-robotics patents 2020–2025 against the United States' 1,483 (WEF, Jun 2025) — a ~3.8× gap. Patent counts measure activity, not quality.

Military-civil fusion. MIIT's standardization framework has been read by US analysts as explicitly linking civilian robotics to PLA priorities. US lawmakers have cited this alongside rare-earth dependence. Treat as a live and consequential policy claim rather than a settled empirical one.

5.2 Unitree (宇树科技) — cheap bodies, real profit, first mover to the public market

Founded 2016 in Hangzhou by Wang Xingxing (王兴兴). Started in quadrupeds (Go2, B2); pivoted to humanoids.

The financials, from the STAR Market prospectus:

Gross margin ~60.1–60.3%. Adjusted net profit (ex-non-recurring) ¥590–600M for 2025. Humanoids rose from 27.6% of revenue (2024) to 51.5% (9M 2025), overtaking quadrupeds. Shipped 5,500+ humanoid units in 2025, which the prospectus claims is a 32.4% global unit share.

The IPO, and what it signals:

Raise: ¥4.202B (~$580–620M) via ≥40.4464M new shares. Implied initial market cap ~¥42B (~$5.9–6.2B). CITIC Securities sponsoring. Meituan-backed entities hold 9.65% (largest external holder), HongShan/Sequoia China 7.11%; Tencent, Alibaba and Ant are also shareholders. Wang Xingxing controls 68.78% of voting rights.

This will be the first embodied-AI listing on China's A-share market, and the first pure-play humanoid company anywhere with an audited, profitable P&L. If it prices well it resets valuation anchors for the entire supply chain.

Status as of 2026-07-27: registration went effective July 6 (valid 12 months); pricing and listing date are still unset, with market chatter pointing at a late-July/early-August debut. And a new number the roadshow will have to explain: updated filings show Q1 2026 adjusted net profit down 52.5% year-on-year — the first crack in the growth story, disclosed at the worst possible moment. And on July 22 — the same week the House voted the other way — Unitree launched commercially in Europe, the first Chinese-made humanoid to enter Western commercial markets, with a stated 2026 target of 20,000 units. Scenario D, playing out in the same news cycle.

The bear case: two years of audited profitability. ¥1.7B revenue against an implied ¥42–50B valuation. Revenue quality is unclear — a large share of Unitree's humanoid units go to research institutions, developers and entertainment/performance use rather than to customers doing productive work. The Spring Festival Gala (春晚) appearances and the Beijing E-Town half-marathon are extraordinary marketing and weak evidence of industrial utility.

Unitree was also selected by NVIDIA as the body for its first reference robotics design sold to research institutions including Stanford and ETH Zurich.

5.3 AgiBot / Zhiyuan Robotics (智元机器人) — the fastest ramp on earth

Founded Feb 2023 in Shanghai by two ex-Huawei engineers: Deng Taihua (邓泰华), former VP, as CEO, and Peng Zhihui (彭志辉), known online as 稚晖君 and recruited through Huawei's "Genius Youth" (天才少年) programme, as CTO.

Production trajectory:

Doubling in three months. TrendForce attributes this to a standardised embodied-AI supply chain with flexible order-driven manufacturing and dedicated supply arrangements — in other words, AgiBot industrialised the process, not just the robot.

Product and software. A-series full-size bipeds (Expedition/远征 A2, A3), G-series wheeled humanoids (G2, G2 Air) for precision industrial work, X-series compact interactive (Lingxi/灵犀 X1, X2 — X1 open-sourced), D-series quadrupeds. Software: GO-1 → GO-2 → GO-3 (expected Q3 2026, ViLLA + world model), GE-2 action world model, Genie Sim 3.0 (built on NVIDIA Isaac Sim), AimRT middleware, and the open-source AgiBot World dataset. NVIDIA Isaac GR00T ecosystem partner (GTC 2026). Multiple Best of CES 2026 awards. At WAIC 2026 it unveiled four products in one pass — A3 Ultra, X2 Edu, G2 Max, and the OmniHand 3 Ultra-M direct-drive dexterous hand — with 30+ robots working the venues and the Yuanzheng A3 debuting as the first full-size humanoid to play table tennis autonomously. Note the hand: the body-makers-go-in-house trend §5.6 warns about, now shipping.

At its April 2026 partner conference AgiBot declared 2026 "Deployment Year One" and announced the "358 Grand Plan" — ¥10B (~$1.4B) revenue within five years, ¥100B (~$14B) within eight. It also launched RaaS rental at €899/day across 17 countries.

Capital structure is unusual and worth watching. Backers: Tencent, HongShan, Hillhouse, BYD, LG Electronics, Mirae Asset, Warburg Pincus, Baidu. The Swancor (上纬新材, 688585.SS) play — which both sides once soft-pedaled — is now explicit: on 2025-11-25 five AgiBot-system executives were elected to Swancor's board, Peng Zhihui became chairman, and Zhou Bin (周斌) serves as the listco's co-CEO and CTO with, per the company's own regulatory reply, full charge of all R&D including embodied robotics. The stock ran from ~¥7.8 (Jul 2025) to ~¥156 by mid-January 2026 — roughly 2,000%, market cap peaking above ¥65B — while fundamentals barely moved (Q3 2025 revenue ¥1.28B, +16.6%; net profit −6.9%) and the embodied business remained, in the company's words, in product development. The exchange formally questioned the arrangement in January. It is the single cleanest exhibit for §08: a wind-blade materials firm at a ¥60B valuation because robotics executives joined the board. Separately, AgiBot is targeting a Hong Kong IPO in Q3 2026 at HK$40–50B (~$5.1–6.4B), with CICC, CITIC and Morgan Stanley as sponsors, selling 15–25%.

AgiBot has also directly invested in 40+ companies, concentrating on supply-chain control: tactile sensors (Qianjue/千觉), linear motors and vision components (Linghou/灵猴), servo motors (Fuxing/伏兴). This is vertical integration by cheque book.

The bear case: executive turnover in Aug 2025 (Lingxi division president Wei Qiang, co-founder Yan Weixin), a 2024 reorganisation that created competing internal R&D divisions, and a shipment mix that skews toward data collection, exhibition, retail and hospitality rather than heavy industry. Shipping 10,000 robots and deploying 10,000 robots are different claims.

5.4 Galbot (银河通用机器人) — brain-first, best-funded, lowest volume

Founded May 2023 in Beijing. The clearest strategic contrast in the sector.

Capital. Nearly ¥7B (~$1B) raised in three years — the most of any Chinese embodied-AI startup. The March 2026 round of ¥2.5B (~$362M) set the sector's single-round record for the third time and brought post-money above ¥20B (~$3B), making Galbot the world's highest-valued unlisted humanoid company. Hong Kong IPO reportedly under consideration.

The investor list is the point. It includes the third phase of China's "Big Fund" (国家集成电路产业投资基金三期) — a state vehicle previously reserved for semiconductors, investing in embodied AI for the first time — alongside SMIC-affiliated Zhongxin Juyuan (中芯聚源), the National AI Industry Investment Fund, CITIC, Sinopec, China Mobile's industry-chain fund, CDB, SAIC (上汽), and Yizhuang State Investment. This is not venture capital. It is industrial policy with a cap table.

Corporate partners: CATL (宁德时代) — Galbot is the only embodied-AI company CATL has invested in, and the only humanoid running normalised operations in a CATL battery plant. Also Bosch (JV with Bosch Capital / 博原资本), Toyota, Hyundai, BAIC (北汽), SAIC, and Bada Precision (>1,000 units committed).

Technology. AstraBrain (银河星脑) — an end-to-end model unifying task planning, motor control and neural coordination in a single architecture, rather than as separate modules. Most Chinese competitors run on foundation models from Alibaba, ByteDance or Huawei; Galbot argues integrated proprietary intelligence is the value layer. Training uses simulated synthetic data for pre-training (AstraSynth / 银河星坊) and real data for post-training.

Product. G1 — wheeled, dual-arm, foldable mobile manipulator, ~85 kg, ~10-hour battery, ~¥700,000 (~$87K–97K), NVIDIA Jetson Thor in the premium configuration. S1 — industrial heavy-load, 50 kg dual-arm payload, autonomous battery swap, rated for dust/vibration/temperature extremes. Galbot Store: autonomous retail across 30+ Chinese cities. Hospital deployments at Xuanwu Hospital (宣武医院).

The reality check — and this is the most useful single data point in this report.

Galbot operates a FamilyMart (全家) in Beijing's Zhongguancun (中关村), on the ground floor of its own headquarters, with engineers stationed inside collecting training data. Reported by HelloChinaTech (July 2026):

• The store is 63 m², roughly 40% smaller than a standard outlet, redesigned around the robot.
• It carries ~1,000 SKUs, ~60% of a normal range. The robot handles fewer than 10 of them.
• Galbot's own figure: grippers fail on a bag of potato chips roughly 4 times in 10.
• Four full-time humans staff the store — the same headcount as a comparable FamilyMart without a robot.
• The robot cannot complete a transaction. A human handles payment.
• Galbot ships fewer than 150 robots a year.

Hold that against the ¥20B valuation and the CCTV Spring Festival Gala appearance, where a Galbot G1 became the first working robot on that stage — rolling walnuts, picking up broken glass, retrieving items from shelves, folding clothes, skewering sausages, live, under stage lighting, with celebrity improvisation, and (the company says) entirely under AstraBrain's real-time decision-making rather than a pre-programmed routine.

Both things are true. The Gala performance is a genuine technical achievement. The convenience store is what deployment actually looks like. The distance between them is the distance between where this industry is and where its valuations assume it will be.

5.5 The rest of the Chinese field

More than 150 humanoid robot companies operate in China; six of eleven major makers planned >1,000 units in 2026 (TrendForce). At least 25 domestic embodied-intelligence unicorns carry valuations above ¥10B, of which 15 were minted in H1 2026 alone (Zhidx, 2026-06-29). The trajectory of that club is the single best measure of the 2026 capital frenzy: 3 members at end-2025 (Unitree, AgiBot, Galbot) → at least 13 by April → at least 25 by end-June. The industry's own framing has turned openly Darwinian — sub-¥10B players, in the words of one practitioner, no longer get a seat at the table.

• UBTech (优必选, HKEX 9880) — the first listed humanoid pure-play, and the one with the most credible industrial deployment. Walker S2, made in Liuzhou (柳州). Began mass delivery Nov 2025 ("hundreds" of units), crossed 1,000 units by end-2025, order book >¥800M (~$113M), later reported ~$195M. Customers: BYD, Geely, FAW-Volkswagen Qingdao, Dongfeng Liuzhou, Audi FAW, BAIC, Foxconn, SF Express, and Airbus (aviation manufacturing, Jan 2026). Capacity target 5,000/year in 2026 → 10,000 in 2027. Autonomous 3-minute battery hot-swap. Humanoids went from ~10% of UBTech sales in 2024 to ~30% in 2025. Its U1 consumer companion robot took 13,000+ pre-orders at launch. UBTech trades at a P/S ratio that is not defensible on any conventional basis.
• XPeng (小鹏) IRON — 82 DoF including 22 per hand, three Turing AI chips at 3,000 TOPS, solid-state battery, deliberately human-like gait. Mass production targeted end-2026. Deployed at Baosteel (宝钢).
• EngineAI (众擎机器人) — Shenzhen, founded 2023. PM01, SE01. Sub-$10K target. Fast iteration, locomotion-first.
• Fourier Intelligence (傅利叶智能) — GR-1/GR-3, rehabilitation and eldercare focus.
• Leju (乐聚机器人), LimX Dynamics (逐际动力), RobotEra (星动纪元), Kepler (开普勒), Astribot (星尘智能), Deep Robotics (云深处) — the long tail, mostly thousand-unit scale or below.
• X Square Robot (自变量机器人) — Series C above ¥20B (~$2.94B), July 2026. WALL-B foundation model on a "World Unified Model" architecture; open-sourced WALL-OSS-0.5 and WALL-WM. Backed by Meituan, Alibaba, ByteDance, HongShan, Xiaomi, IDG.
• Xinghaitu (星海图) — ¥1B Series B (Feb 2026), then ~¥2B Series B+ (Apr 2026) at ¥20B. Fastest valuation increase in the track post-Spring Festival 2026. Fast-WAM world model, open-sourced G0.5 VLA.
• TARS (它石智航) — founded Feb 2025 by Chen Yilun (陈亦伦) — ex-DJI machine-vision chief engineer, architect of Huawei's first-generation ADS autonomous-driving stack, then Tsinghua AIR chief scientist — with chairman Li Zhenyu (李震宇, ex-Baidu SVP who built Apollo) and chief scientist Ding Wenchao (丁文超, first-batch Huawei "Genius Youth"). A $120M record angel, then a $455M Pre-A (April 2026) — the largest single round in Chinese embodied-AI history — at ~¥13B, roughly 14 months after founding. AWE 3.0 world-action model; its A1 holds a Guinness record for sub-millimeter wire-harness assembly cycles per hour.
• Spirit AI (千寻智能) — Hangzhou, Jan 2024. Han Fengtao (韩峰涛, ex-ROKAE co-founder/CTO, ~20,000 industrial robots delivered) + Gao Yang (高阳, Tsinghua IIIS assistant professor, Berkeley PhD). Six rounds, ~¥3.3B in under two years; past ¥10B (Feb 2026, Yunfeng-led). The claim that matters most: its Moz1 (小墨) robots run a live battery-PACK station at CATL's Zhongzhou plant — billed as the world's first humanoid production line, ~1,000 batteries with a zero-fault claim at human-or-better takt. If independently verified, this is the strongest productive-deployment datapoint anywhere. Open-source Spirit v1.5 claims to outperform π0.5.
• AI2 Robotics (智平方) — ¥1B+ B-series (Baidu strategic, CRRC Capital); signed a ¥500M, thousand-unit order with an HKC (惠科) subsidiary — one of the sector's first large binding commercial orders, claimed 2–3-year payback.
• Booster Robotics (加速进化) — near-¥1B round (April 2026; Beijing state funds, Legend Star); education and developer-platform bipeds.
• Pudu (普渡) and Keenon (擎朗) — the profitable commercial-service incumbents crossing over: Pudu raised ~¥1B in April 2026, entering the ¥10B club on its delivery/cleaning fleets; Keenon debuted the XMAN humanoid at WAIC 2025. Real revenue, global distribution, now bolting on humanoid form factors — the quiet consolidators.
• GigaAI (极佳视界) — world-model pure-play; entered the ¥10B club in April 2026.
• Matrix Robotics — Shanghai; broke out at WAIC 2026 with MATRIX-3 (1.7 m, 65 kg, 27-DoF cable-driven hands, fingertip force sensing to 0.1 N, ~4 h continuous runtime) priced from ¥580K, cumulative orders approaching 1,000, a Zhangjiang base targeting 5,000 units/yr in 2026 → 10,000 in 2027, and its WAVE physical-world foundation model plus a global partner program. The commercial-service dark horse of the show.
• Noetix / Songyan Dynamics (松延动力) — Beijing, founded Sep 2023; Tsinghua-background team led by Jiang Zheyuan (姜哲源). The purest consumer price-floor test in the industry: Bumi (小布米), ¥9,998 — billed as the world's first sub-¥10,000 high-performance humanoid (94 cm, 12 kg, 21+ DoF), first deliveries March 2026, cumulative orders in the thousands. The only maker running both bipedal (N2, E1) and bionic (Hobbs) product lines; the 2026 CCTV Gala's humanoid partner. Unusually transparent production: 105 units delivered in July 2025 alone, Beijing capacity >200/month, Changzhou phase one targeting 300/month; order book >¥100M by mid-2025. Nine funding rounds, capped by a near-¥1B cumulative B round (March 2026) led by Chendao Capital (晨道资本) — CATL's industrial investment platform — after Pre-B (~¥300M, Fangguang) and Pre-B+ (~¥200M, CICC Capital) rounds in late 2025.
• Automaker programmes — BYD, GAC (GoMate), Chery, Geely, Xiaomi. The EV-to-humanoid pivot is structural: shared actuators, batteries, sensors, and assembly discipline. Huawei invested ~$413M into its Jimu (极目) robotics subsidiary.

A pattern worth naming: CATL (宁德时代) capital now touches three of the sector's Gala-adjacent names — Galbot (direct investment; the only embodied-AI company in its portfolio), Spirit AI (the live Zhongzhou production line), and Noetix (B round led by CATL's Chendao platform). The battery incumbent is quietly assembling an embodied-AI position spanning brains, deployment, and consumer hardware — and, unlike the municipal funds, it brings a factory to train in.

5.6 The picks-and-shovels layer — hands, touch, and data

In the first half of 2026 the most violent repricing in robotics did not happen at the integrator layer at all. It happened one level down, in components and data — the picks and shovels. Three sub-markets matter.

Dexterous hands, and the Linkerbot phenomenon (灵心巧手). The context is §3.3's dexterity wall, plus one datapoint that concentrated minds: Tesla's sorting-trial hands lasting ~six weeks at $6,000+ apiece, with Musk tying V3's delays to the hand. More than 80 companies worldwide now build dexterous hands, over half of them Chinese, across three unconverged technical routes — tendon-driven (腱绳), direct-drive (直驱), and linkage (连杆).

Linkerbot is the runaway. Founded by Zhou Yong (周永), a Huazhong UST alumnus on his third startup (games/community, then outdoor autonomous vehicles; incubation from 2019, independent from 2023), it claims to be the only maker mass-producing high-DoF hands at four-digit monthly volume — ~4,000–5,000 units/month by spring 2026 — and 80%+ of the global high-DoF dexterous-hand market, with its Linker Hand L20/L30/O30 lines deliberately covering all three drive routes. Around the hardware sits a data play: LinkerSkillNet (500+ human manual skills captured as standardized robot capabilities) and an announced "million-robot data collection plan." ~400 staff, five plants across Beijing and Shenzhen, and a "hands making hands" automated line.

The capital curve is the steepest in the industry: seven rounds since early 2025 (Ant Group, Sequoia China, Leju, AUX, GF Securities in the seed-to-A+ stretch), a ¥1.5B B round in February 2026 that broke ¥10B valuation — a reported ~10× step-up in seven months — then a B+ on April 29 at ~$3B (Zhongguancun Science City Fund, BOC Asset Management, Fosun Chuangfu), and within a week Reuters reported it seeking $6B for the next round. From a ~¥200M-class valuation in early 2025, that is roughly a 100× re-rating in about 14 months, against a 2026 delivery target of 50,000–100,000 hands.

The bear case for the hand trade deserves equal airtime: one report puts current orders at ~1,000 units/month against a 10,000/month capacity target, meaning the valuation prices the entire industry's ramp arriving on schedule; a hand business cannot out-scale the bodies it attaches to; hedging all three drive routes is capital-intensive; and sell-side coverage (Soochow Securities) expects body-makers to bring hands in-house (Tesla and Unitree already do) while component makers — Zhaowei (兆威机电), Leadshine (雷赛), Paxini (帕西尼) — climb upward into complete hands. Squeezed from both ends, the independent integrator's moat is manufacturing yield and data, not design. Founder sentiment is its own gauge: Zhou predicts a ¥100B+ Chinese embodied-AI company within two years and a ¥1T one within five. Other hand players: Dexterous Intelligence (灵巧智能), Sharpa (supplies NVIDIA's reference humanoid), Inspire (因时机器人, the incumbent research supplier), BrainCo (强脑), OYMotion (傲意).

Tactile sensing. Paxini (帕西尼感知) — a sensor maker climbing upward into full hands — entered the ¥10B+ club in April 2026; Daimon (戴盟机器人) carries the HKUST visuotactile lineage; Qianjue (千觉) is AgiBot's captive tactile bet. This is §3.3's six-axis force/torque constraint, now with venture-scale capital attached.

Data infrastructure. The §09 data bottleneck has become an investable category in its own right. AgiBot runs a 4,000 m² dedicated data-collection factory in Shanghai producing 30,000–50,000 trajectories a day. Spirit AI's fifth-generation wearable rigs cut collection cost to ~1/10 of teleoperation — 200,000+ real-interaction hours banked, a 1-million-hour 2026 target, and an explicitly contrarian thesis that "dirty data is the key to scaling VLA models." TARS built SenseHub wearables and the WIYH multimodal dataset on the same logic. Linkerbot's million-robot plan is a data play wearing a hardware costume. Add 40+ state-funded training grounds and the Kupas (库帕思) corpus-standards partnership, and the picture is clear: the industry has converged on §3.5's scaling-law bet, and the new competition is over who can manufacture data cheapest. For calibration on how far there is to go — TARS's Chen Yilun puts the usable threshold at 10 million+ hours of embodied data; Galbot's team argues trillion-scale.

And a fourth sub-market is forming at the geopolitical seam. Between Chinese hardware supply and Western demand, an integrator layer is emerging: US-based distributors that sell, lease, and service Chinese-made humanoids (San Jose's IntBot, which resells and rents AgiBot units, is an early example), campus resellers moving Unitrees into American universities, and — still nascent — insurance products for deployed fleets. The insurance angle deserves attention beyond its premium volume: an underwriter is the one commercial actor positioned to require the MTBF and teleoperation-ratio disclosures this report keeps flagging as absent (§10, §11), which would make underwriting data the sector's de facto certification layer. The whole seam, however, is hostage to §7.3's regulatory trajectory.

§06 — The rest of the world

6.1 Japan — components, not platforms

Japan's strategy is legible and deliberate: own the parts, cede the assembly. Harmonic Drive dominates strain-wave gearing; Nabtesco is #1 globally in RV cycloidal reducers; Nidec in motors; THK in linear guides. At iREX 2025, Harmonic Drive showed reducers designed specifically for humanoid joints — flat high-torque units for neck and arms, ultra-compact units for fingers.

TrendForce's framing (Dec 2025): Japan is raising the entry barrier at the component layer while the US and China race on full systems.

Japan's demand-side driver is unique and urgent: the working-age population is projected to fall ~31% between 2023 and 2060, and the eldercare labour gap is the country's most acute. TrendForce expects eldercare to be Japan's fastest-growing humanoid application — Kawasaki's Nyokkey, Fourier's GR-3. Kawasaki also showed Kaleido 9 (30 kg lift, VR teleoperation, disaster response). Toyota Research Institute runs a $1B+ robotics R&D programme (T-HR3, Punyo).

Japan installed 44,500 industrial robots in 2024 and holds 450,500 in stock — second globally. It also remains the world's predominant robot manufacturing country.

And: SoftBank agreed to acquire ABB's robotics division for $5.4B, explicitly framing physical AI as its next frontier. That is the largest single robotics transaction of the cycle and it moves a top-tier industrial arm business into a Japanese AI-first owner.

6.2 South Korea — density leader, vertically integrated

Korea has the world's highest robot density: 1,220 robots per 10,000 manufacturing employees, growing ~7% annually since 2019 — more than 4× Germany. 30,600 installations in 2024.

• Hyundai Motor Group — owns Boston Dynamics outright as of mid-2026. Announced an AI Robotics Strategy at CES 2026; Atlas deploying into the Georgia Metaplant. Codeveloping robot-dedicated batteries with Samsung SDI.
• Rainbow Robotics (레인보우로보틱스) — RB-Y1 dual-arm, HUBO lineage from KAIST. Samsung Electronics holds ~35%. The clearest pure-play listed humanoid stock in Asia.
• Samsung SDI / LG Energy Solution — humanoid-specific cell programmes; solid-state targeted H2 2027.
• LG Electronics — CLoiD home robot at CES 2026; adopting GR00T N1.7.

6.3 Europe — capital arrived late, and loudly

Europe's humanoid position was negligible until 2026. It is not any more.

• Neura Robotics (Germany) — raised ~€1B / $1.4B at a ~€4B valuation, backed by Amazon and Qualcomm. 4NE1 humanoid, from €19,999 at CES 2026. Partnerships with Schaeffler (actuators) and Bosch (components, motor production).
• Hexagon (Sweden) — AEON humanoid, in testing at industrial clients including BMW.
• Schaeffler, Bosch, Infineon, NXP, STMicro — the component incumbents are all formalising humanoid-specific product lines. Schaeffler expects up to 10% of group sales by 2035 from new sectors including humanoids.
• Humanoid (UK) — HMND 01; the Schaeffler deal (May 2026) covers up to 1,000–2,000 robots at Schaeffler plants by 2032 plus 1M+ joint actuators through 2031.
• PAL Robotics (Spain), Engineered Arts (UK), Wandercraft (France), Agile Robots (Germany), Oversonic (Italy) — specialist and premium niches.

Europe's structural asset is regulation, not hardware. The EU AI Act and the updated Machinery Regulation will force the first commercial humanoid operators to demonstrate systematic safety cases — SVRC projects by Q3 2027. Whoever solves European safety certification first gets a moat that cheap actuators cannot cross.

6.4 Elsewhere

Canada (Sanctuary AI — Phoenix, Magna equity stake, tactile in-hand manipulation), Israel, India (record 9,100 industrial installations in 2024, +7%, 6th globally), Vietnam (VinMotion, Vingroup).

§07 — US vs China: the structural comparison

7.1 The scorecard

External validation for the scorecard's shape arrived at WAIC 2026 — from the Chinese side. Jiang Lei (江磊), chief scientist of the national humanoid robotics innovation center, told the conference that China is already ahead in smart manufacturing, datasets and training grounds, while the US retains the edge in large-scale foundation models and computing (Global Times, 2026-07-18). That is this report's stack map, endorsed by the counterparty.

7.2 The three arguments that actually matter

Argument 1 — "The body commoditises, the brain doesn't." This is the American thesis and the reason Skild is worth $14B without a factory. If VLA models transfer across embodiments, hardware becomes a low-margin substrate and the policy layer captures the value, exactly as Android captured value over handset OEMs. Evidence for: GR00T runs on Unitree, AgiBot and Figure bodies. π0.7 transfers across robots. Evidence against: Figure exited OpenAI specifically because Adcock concluded embodied AI must be vertically integrated to work at all. Both cannot be right.

Argument 2 — "Whoever ships the bodies gets the data, and data compounds." This is the Chinese thesis, and NVIDIA's EgoScale result is the strongest evidence for it: 1,000 → 20,000 hours of egocentric data more than doubled task completion. If robot dexterity follows a data scaling law, then 10,000 fielded AgiBot units generating trajectories is a compounding asset that no amount of San Francisco capital can buy. Counter-evidence: most Chinese fielded units are in exhibitions, retail greeting, education, and data-collection farms — high volume, low task diversity, questionable trajectory quality. Also, China's own state-funded robot training centres (40+) exist precisely because natural deployment is not generating enough data.

Argument 3 — "Supply chain beats everything." The magnet number is not a talking point; it is a physical constraint. ~90% of magnet processing sits in one jurisdiction that has already demonstrated willingness to use export licensing as leverage. A humanoid carries 2–3× the magnetic content of an EV. Every Western scale-up plan — Figure's 3M actuators, Tesla's 10M units/year, Apptronik with Jabil — runs through that chokepoint or through a several-year, several-billion-dollar effort to build around it. Counter-argument: magnets are a cost problem, not an existence problem. MP Materials, Lynas and others are scaling; the cost penalty is real but not prohibitive at the volumes anyone is actually building in 2026–2028.

7.3 The US policy response, such as it is

There is no national robotics strategy. There are bills and coalitions:

• National Commission on Robotics Act (Sens. McCormick, Hickenlooper) — advancing.
• S.3275, "Humanoid Robotics Oversight and Blocking of Obtainment from Totalitarians Act" (Sens. Cassidy, Coons) — would bar federal acquisition of AI-integrated humanoids from military suppliers to China, Iran, North Korea, Russia.
• The GUARD Act (H.R., Moolenaar/Obernolte/McClellan, introduced 2026-06-03) — the escalation. It directs national-security agencies to review humanoid and quadruped robots from adversary nations including the software and communications systems that control them; unacceptable-risk products go on the FCC Covered List (import and operation ban), and — the forcing mechanism — anything not affirmatively reviewed within one year is listed automatically. Moolenaar names Unitree as "flooding the market with artificially cheap products"; Agility has endorsed the bill. On 2026-07-22 the House passed the NDAA carrying the GUARD Act's core provision. Watch the Senate conference, and whether the control-software language survives — it determines whether "Chinese body + US brain" is a legal architecture or a banned one, which is to say it prices scenario D. Late-July reporting points to further moves barring Chinese-made humanoids from US military use specifically — the procurement wall going up ahead of the commercial one.
• Robots for America — industry coalition launched in response to a call from OSTP, Commerce, SBA and the Senate. Founding members include Formic, Machina Labs, Standard Bots, Dexterity, Path Robotics, Chef Robotics, GrayMatter, Mujin, Viam.
• Partnership for Robotics Competitiveness (PfRC) — AUVSI-organised.
• A presidential executive order on robotics has been reported as under consideration; Commerce Secretary Lutnick has been meeting robotics CEOs.
• Rep. Zoe Lofgren and others have flagged that Chinese dominance extends into the supply chain American robot makers depend on.

The Special Competitive Studies Project's framing is the sharpest: Beijing has R&D goals, financial incentives, and adoption quotas. The United States has no strategy. SCSP's specific recommendation — a national Data Foundry for Robotics — recognises that no internet-scale corpus exists for embodied AI, and that whoever builds one first sets the terms.

§08 — Signal versus marketing

Read this section before quoting any number from this sector.

The spectacle metrics are spectacle. In April 2026, a fully autonomous humanoid named "Lightning," built by Honor (荣耀), won the Beijing E-Town Half-Marathon in 50:26 — faster than the human half-marathon world record. More than 300 robots from 100+ teams competed. This is a real engineering accomplishment in locomotion endurance and thermal management. It is not a like-for-like comparison with a human athlete, and it says nothing about whether a robot can pick up a bag of crisps. Similarly: the CCTV Spring Festival Gala (春晚) performances, the Shanghai embodied-intelligence experience stores, the Seoul fashion-show runway, the 15 humanoid "police officers" in Hangzhou. Every one of these is a marketing artefact that Western coverage reports as a capability datapoint.

The IFR — the industry's own statistics body, with no incentive to be bearish — says this out loud. Its May 2026 assessment: despite the staged presentations, actual capability in real-world production is limited to demonstrators and pilot projects. It invokes form follows function: a human body is not well-suited to most industrial tasks; conventional industrial robots have fewer joints, simpler control, higher speed and better reliability, and will remain the backbone of precision manufacturing. IFR's conclusion: mass adoption as universal factory helpers or household assistants will not happen in the near or medium term.

"Shipped" ≠ "deployed" ≠ "productive." AgiBot's 10,000 units are shipped. UBTech's Walker S2 fleet is deployed. Agility's Digit — and, per company claims, Spirit AI's Moz1 on CATL's line — are productive. These are three different claims and the industry uses one word for all of them.

Payback-period math is doing enormous unexamined work. IDTechEx (May 2026) found that under high-utilisation industrial conditions, humanoid payback can fall to ~6 months. The load-bearing phrase is "high-utilisation." A Digit that works in 30-minute intervals around 9-minute charges is not high-utilisation. Against $30/hour labour and a $50K robot, a 12–18 month payback is defensible; the sub-6-month claims assume 2028 prices and 24/7 duty cycles that current batteries cannot support.

Where the hype cycle actually sits. Rare Earth Exchanges (July 2026) places humanoids at roughly 70–80% of the way to the Peak of Inflated Expectations, with the peak likely landing between now and 2028 — when production targets, coverage and investor expectations outrun deployment reality. The trough follows when buyers discover that wheeled robots, fixed automation and conventional arms are cheaper and more reliable for most tasks. That is a Tier 3 source with a stake in rare-earth demand. It is also, on the evidence assembled here, approximately correct.

A record fundraising year guarantees a record washout. Crunchbase's own framing. Physical products are unforgiving. Over-funded humanoid plays will collide with manufacturing, reliability and margin.

§09 — The bottlenecks, ranked

1. Data. The core constraint. No internet-scale corpus of embodied experience exists. Physical trials are slow, expensive, and damage hardware. Simulation is the substitute, and simulators do not reproduce contact dynamics, friction, compliance and timing — so sim-trained policies degrade sharply on hardware. The January 2026 literature calls this gap structural, not incremental. The counterweight is EgoScale: if human egocentric video substitutes for robot teleoperation at scale, the constraint loosens dramatically. This is the single most important open question in robotics.

2. Dexterous manipulation under uncertainty. Deformable fabrics, reflective packaging, random clutter, occlusion. Galbot's own gripper failure rate on a chip bag: ~40%. Six-axis force/torque sensing remains supply-constrained and calibration-intensive. High-DoF biomimetic hands are mechanically complex, expensive, and historically failure-prone — Tesla's 22-DoF-per-hand bet is not obviously the right one.

3. Energy. 90 minutes to 5 hours against an 8-hour shift. Bipedal balance burns power before any work happens. The architectural answers (hot-swap, wheeled bases, battery-swap stations) are conceding the point.

4. Reliability and MTTF. Enterprise buyers demand mean-time-between-failure guarantees and zero-downtime safety metrics. Current generative robot policies exhibit stochastic failure modes in unstructured environments. Until modular execution is standardised and self-calibration protocols cut field-service cost, commercial contracts scale slowly regardless of funding. Boston Dynamics' own CEO sets the bar plainly: Atlas must learn new factory tasks within a day or two and reach 99.9% reliability before being broadly useful on production floors — a number no humanoid publicly demonstrates today.

5. Safety, certification, liability. ISO humanoid safety standards are expected around 2028. EU AI Act + Machinery Regulation safety cases by Q3 2027 (projected). US OSHA guidance on autonomous robot co-workers expected H1 2027. The industry's own pioneers make the sharpest point: when a chatbot hallucinates, it confuses; when a 70 kg robot does, it injures. NVIDIA's Halos for Robotics (June 2026) is the first full-stack open safety architecture, drawing on 18,600 engineering-years of autonomous-vehicle safety work, with Agility's Digit as first adopter.

6. Cost curve. Manufacturing cost fell ~40% in 2024 against an expected 15–20%. BOM for a useful humanoid in 2026 is roughly half its 2024 level. Bank of America models average BOM falling from ~$35K (2024) to ~$17K (2030) at million-unit scale with full Chinese sourcing. Every credible sub-$20K forecast assumes Chinese sourcing. That is a policy problem disguised as an engineering forecast.

7. Cybersecurity. At GEEKCon 2025 in Shanghai, researchers from DARKNAVY demonstrated seizing control of a commercially available humanoid with a single verbal command. A machine with cameras, microphones, network access and physical manipulators is an attack surface with no precedent. Recorded Future has flagged hijacking, data exfiltration and botnet formation. Consumer home robots that ship with mandatory teleoperation and data-sharing (1X NEO) compound this.

8. The form-factor question itself. IFR's form follows function critique is the one nobody in the funding cycle wants to engage. If wheeled dual-arm mobile manipulators get 10-hour runtimes, better stability, and lower cost, and can do 90% of warehouse and retail tasks — why is anyone building legs? The honest answer is: stairs, unmodified human environments, and the home. Whether that residual is worth the energy and reliability penalty is an empirical question that 2027–2029 will settle.

§10 — What to watch, next 12–24 months

Near-term, dated triggers:

Leading indicators worth tracking continuously:

1. Teleoperation ratio. The single most diagnostic and least disclosed metric. What fraction of a fielded robot's task-hours are autonomous? 1X is honest about it. Nobody else publishes it. Demand it.
2. Repeat orders. Pilots are marketing. Second and third purchase orders from the same customer are the only evidence of ROI.
3. Actuator and magnet pricing. Leader Harmonious Drive, Sanhua, Northern Rare Earth, MP Materials, Harmonic Drive, Nabtesco, Schaeffler. The component cascade re-rates before the robot makers do.
4. RaaS attach rate. AgiBot at €899/day, Agility and Apptronik subscription, 1X at $499/month. Whether robotics becomes a recurring-revenue business determines the terminal multiple.
5. Whether anyone publishes an MTBF number. As of July 2026, effectively no one has.
6. Export controls in both directions. Rare-earth licensing from Beijing; Entity List and CFIUS from Washington. Either can reprice the sector in a week.

§11 — Scenarios to 2030

A. "Automation 3.0" (bull, ~25%). Data scaling laws hold. Egocentric video substitutes for teleoperation. Costs reach $17–20K BOM at volume. Automotive and logistics scale first, as IDTechEx expects, then general warehousing. Global installed base reaches low millions. Morgan Stanley's $5T-by-2050 framing looks conservative. China supplies the bodies; the US and Google/NVIDIA supply the brains; Japan and Germany supply the joints. Everyone wins except workers doing repetitive physical labour.

B. "The wheeled correction" (base case, ~45%). Bipedalism turns out to be a solved research problem and an unsolved commercial one. Wheeled dual-arm mobile manipulators — Galbot G1, AgiBot G-series, Kinisi KR1 — take the deployable market on runtime, stability and cost. Legged humanoids remain a prestige and R&D category through 2030, with real but bounded use in unmodified human environments. The market is real, ~$15–25B by 2030, and looks nothing like the renderings. Several well-funded biped startups fail or are acquired for talent.

C. "Peak, trough, plateau" (bear, ~25%). The Gartner peak arrives 2027. Reliability and safety certification prove harder than modelled; a serious workplace injury involving a humanoid triggers regulatory freeze in at least one major market. Funding stops. The washout Crunchbase predicts arrives. The technology survives, the companies mostly do not, and the eventual winners are 2032-vintage firms building on the corpse of 2026 capex — the exact pattern of autonomous vehicles.

D. "Bifurcation" (tail, ~5%). Export controls and procurement bans harden into two non-interoperable robotics stacks: a Chinese stack (Chinese bodies, Chinese components, Chinese models, exported across the Global South) and a Western stack (US models, allied components, 2–3× cost, deployed in North America, Europe, Japan, Korea). Standards diverge. This is the semiconductor playbook, and China's HEIS 2026 standard-setting push suggests Beijing is at least prepared for it.

The single fact that would move me most: an independently audited teleoperation ratio and MTBF figure from any of Figure, AgiBot, UBTech or Agility. Nobody publishes one. Until someone does, every valuation in this report is a bet on an undisclosed number.

§12 — Appendix

12.1 Company quick reference

12.2 Bilingual glossary — 术语对照

12.3 Principal sources

Tier 1 — IFR World Robotics 2025 (Industrial Robots, Service Robots, Robot Density); IFR press statements Sep 2025, Apr 2026, May 2026. Unitree STAR Market prospectus (2026-03-20, updated 2026-05-25). Tesla Q4 2025 and Q1 2026 earnings calls; Tesla 8-K FY2026. IDC Worldwide Humanoid Robotics Market Analysis 2026 (Jan 2026). Caixin (2026-03-03, 2026-05-26, 2026-07-03). Reuters. Bloomberg (2026-03-27). SSE / CSRC filings. Company disclosures: Figure, Apptronik, Agility, NVIDIA, UBTech, AgiBot, Galbot.

Tier 2 — McKinsey, Scaling the humanoid robotics supply chain (Apr 2026). Morgan Stanley, The Humanoid 100 and May 2026 humanoid market note. TrendForce (Dec 2025, Apr 2026). IDTechEx, Humanoid Robots 2026–2036 (Feb 2026). Omdia. MERICS, Embodied AI: China's ambitious path (May 2026). Jamestown Foundation (Jan 2026, Apr 2026). The Diplomat (Mar 2026). Special Competitive Studies Project, National Robotics Strategy memo. Crunchbase News. Dealroom. PitchBook.

Tier 3 — used, and flagged where used — SVRC State of Robotics 2026 (vendor-published; the firm sells robots and datasets). KraneShares / KOID commentary (ETF sponsor). RoboZaps, humanoid.press, botinfo.ai, optimusk.blog (specialist trackers, useful for dates and specs, commercially motivated). HelloChinaTech (Jul 2026) for the Galbot FamilyMart field observation — single-source but first-hand and specific.

12.4 What this report does not cover

Defence and autonomous maritime/aerial systems (Saronic's $1.75B Series D at $9.25B is the largest robotics round of 2026 and sits outside this scope). Autonomous vehicles (Wayve, $1.2B at $8.6B). Surgical robotics beyond the IFR category data. Agricultural robotics. Warehouse fixed automation (Symbotic, Amazon's 750,000+ mobile robots). Any of these would materially change a "state of robotics" picture framed by revenue rather than by attention.

Report ends · Prepared 2026-07-09 · All figures date-stamped inline · Neutral market intelligence — no positions held, none recommended