The robotics market is two markets — built on one supply chain.
China installed 295,000 industrial robots in 2024 — 8.6× the U.S. — and Chinese OEMs now ship the majority of humanoid units worldwide. Public component research sharpens the supply-chain picture: the fight is not only magnets, reducers, and AI compute, but whether suppliers can industrialize planetary roller screws / 行星滚柱丝杠, dexterous hands / 灵巧手, sensors / 传感器, and joint modules / 关节模组 at scale.
Two countries, very different volumes — but a closer dollar gap than the unit gap suggests
China leads installations 8.6×, but Chinese robots are cheaper. The industrial robotics dollar markets are within $250M of each other today, and projected to converge again around $8B by 2033.
Industrial robot installations, 2024
Industrial robot market, USD billions
Humanoid robot market — regional share, 2025
Public-source lens: lower price is a manufacturing-learning loop
The original atlas said Chinese robots are cheaper. Public market and company sources make the mechanism more concrete: cost-down comes from validated suppliers, process yield, tooling, life testing, and assembly feedback across modules such as screws / 丝杠, reducers / 减速器, coreless motors / 空心杯电机, and sensors / 传感器.
New public Chinese-market inputs: more aggressive production scenarios
Public summaries of Donghai Securities and Yongxing Securities reports are more aggressive than some global market-size estimates. Donghai frames humanoids as moving from concept to mass production and cites a 2029 scenario of at least 1 million units of global capacity and a 150B RMB market. Yongxing frames 2025-2030 humanoids as a high-growth hardware cycle tied to AI-enabled manufacturing. The difference is not a clean contradiction: the global chart above tracks market revenue estimates, while the Chinese broker summaries often mix capacity, policy targets, supplier opportunity, and bull-case industrialization scenarios.
Tap any body part — see what's inside, who supplies it, what fails without it
A humanoid robot has roughly 8 critical supplier dependencies, distributed across joints, muscles, hands, and brain. Public component sources add more weight to screws / 丝杠, crossed-roller bearings / 交叉滚子轴承, and hand micro-actuation / 手部微型驱动 than the first-pass atlas did.
The U.S. buys robots for cars. China buys robots for everything.
In the U.S., automotive accounts for 40% of installations and is the only major sector growing. China's mix is led by electronics (28%) — and food & beverage installs jumped 86% YoY as automation moves off the assembly line.
Why sector mix matters for humanoids
Public industrial-robot data does not mean electronics or food robots automatically become humanoids. The implication is subtler: broad factory deployment creates supplier learning, customer validation, and after-sales data for mechatronic modules. That is why China's industrial base matters even when the final robot form factor changes.
Three tiers, seven bottlenecks — plus the manufacturing layer that turns them into robots
Components flow up into integrators; integrators flow out into applications. Public company and market sources shift the question from "which country owns the layer?" to "which supplier can repeatedly build, test, and cost-down the layer?" Both lenses matter.
JPCNUS
JPCNTW
USCN
CNJP
~75% of shipments by Big 4
Warehouse + last-mile
Light assembly + labs
~60% CN by 2025 shipments
Figure BotQ public example
~50% of installs
Amazon, JD, Walmart
Intuitive, iRobot, Pudu
Boston Dynamics, ANYbotics
Bottlenecks — quantified and refined by public sources
Where public sources refine the original supply-chain framing
- Ball screws vs. planetary roller screws: the original chart grouped "Ball screws & LM" as a broad linear-motion market. Public humanoid component sources separate planetary roller screw / 行星滚柱丝杠 and inverted planetary roller screw / 反向式行星滚柱丝杠 because load, lifetime, and compact packaging are harder than generic linear motion.
- Sensor ranking: this atlas still marks force, tactile, and vision sensors as "medium" for supply concentration. Public robot-hand and autonomy sources make them capability-critical for dexterous manipulation. The difference is definition: supply-chain scarcity vs. product capability.
- Country-share precision: the country-share chart mixes broad component markets with humanoid-specific BOM estimates. Public sources are better for part-priority hierarchy than exact global market shares, so the chart should be read directionally.
New public Chinese supplier map
Donghai's public report summary reinforces the same component hierarchy and gives public examples of domestic supplier categories: harmonic reducers / 谐波减速器, planetary roller screws / 行星滚柱丝杠, frameless torque motors / 无框力矩电机, dexterous hands / 灵巧手, and six-axis torque sensors / 六维力矩传感器. The site uses those as category signals, not as verified design wins for Tesla, Figure, Unitree, or UBTech.
Country share of each layer
The hardware lesson: cost-down manufacturing, not just better robot demos
Public company posts, public market-research abstracts, and public robotics coverage point to the same commercialization hurdle: tightly integrated suppliers must turn custom mechatronic parts into reliable, lower-cost, repeatable production. Where public sources differ from the original atlas, the explanation is usually scope, robot version, or definition.
1. Linear joints are a hard mechanical bottleneck
- Planetary roller screws / 行星滚柱丝杠 are a public market category for humanoids because robots need compact, load-bearing linear motion at knees, waist, and lift axes.
- The practical hierarchy is inverted planetary roller screw / 反向式行星滚柱丝杠 and planetary roller screw above ordinary ball screw when miniaturization and load density dominate.
- The manufacturing barrier is less the design concept and more grinding, process control, life testing, and cost reduction at volume.
2. Hands move the BOM toward micro-motors and sensing
- Dexterous hands / 灵巧手 create demand for small, high-response drives rather than only large hip and knee actuators. Public TrendForce coverage also flags robot-hand design as a production bottleneck for Optimus.
- Coreless motors / 空心杯电机 matter because narrow fingers need compact, low-inertia, high-precision motion.
- Force and tactile sensors / 力觉与触觉传感器 are the difference between a robot that grips and a robot that manipulates fragile objects.
3. Manufacturing platforms may capture more value than parts vendors
- Figure's public BotQ posts make the same layer visible: in-house assembly, MES/PLM/ERP/WMS software, reliability testing, and external piece-part vendors.
- This points to a separate value pool: companies that can convert lab robots into manufacturable SKUs may become sticky even before the final market standardizes.
- For the atlas, this is now grounded only in public company and market sources, not private clipping material.
Public 2026 machinery strategy lens
Yongxing's public 2026 machinery strategy summary widens the atlas from humanoid-only hardware to AI-enabled manufacturing equipment / AI赋能制造装备. It highlights intelligent manufacturing equipment, industrial internet, 3D printing / 3D打印 for some robot sensor structures, intelligent logistics, and unmanned forklifts. The implication is that humanoids are one expression of a broader factory-intelligence cycle, not the whole market.
Public-source reconciliations
Ball screw market vs. humanoid PRS bottleneck
The original atlas used broad linear-motion shares. Public humanoid component sources focus on compact high-load joints, where planetary roller screws / 行星滚柱丝杠 and inverted planetary roller screws / 反向式行星滚柱丝杠 are harder and higher value than ordinary ball screws / 滚珠丝杠.
28 body actuators vs. hand-generation counts
The atlas uses 28 body actuators for Optimus Gen 2: 14 rotary and 14 linear. Public sources also discuss hand generations separately, including 11 DoF per hand for Gen 2 and higher Gen 3 hand counts. That is a counting-scope issue, not a direct contradiction.
Medium sensor chokepoint vs. critical manipulation layer
Supply concentration makes sensors look less severe than magnets or AI compute. Product capability makes six-axis force / 六维力传感器 and tactile sensing / 触觉传感器 indispensable. Public sources are arguing capability criticality, not necessarily single-country dependency.
Figure supplier mix is not validated by the Chinese pack
Public Figure sources support in-house core-module assembly plus external piece-part vendors, but they do not name the precision-component supplier list. The atlas therefore treats Figure's supplier mix as undisclosed.
Bilingual technical glossary
Source note: this section now uses only public web sources listed below. Private/local clippings are not used as evidence for the public atlas.
An American brain on a Chinese body
Tesla designs the AI4/AI5 chip, the Optimus Cortex VLA model, and the actuator system specs in California. Public sources support a China-heavy component stack for Optimus, while also making clear that actuator counts vary by body generation and whether hands/fingers are included.
| Component | Specification | Qty | Primary supplier | Country |
|---|---|---|---|---|
| Actuation — joints & muscles | ||||
| Rotary actuator (small) | Frameless torque motor + harmonic reducer · 20 Nm | ~6 | Suzhou Green Harmonic | CN |
| Rotary actuator (mid) | Torque motor + harmonic reducer · 110 Nm | ~4 | Tuopu Group | CN |
| Rotary actuator (large) | Torque motor + harmonic reducer · 180 Nm | ~4 | Tuopu / Sanhua | CN |
| Linear actuator (small) | Planetary roller screw / 行星滚柱丝杠 · 500 N | ~4 | Sanhua Intelligent | CN |
| Linear actuator (mid) | Planetary roller screw / 行星滚柱丝杠 · 3,900 N | ~6 | Sanhua ($685M order) | CN |
| Linear actuator (large) | Roller screw / 滚柱丝杠 · 8,000 N · knees | ~4 | Sanhua / Tuopu | CN |
| Hand micro-actuators | Coreless motor / 空心杯电机 + tendon drive / 腱绳传动 | 11/hand | Zhaowei | CN |
| Mechanical — bearings, screws, structure | ||||
| Planetary roller screws / 行星滚柱丝杠 | Linear motion — ~19% of BOM in some estimates | 14 | Beite, XCC, Best | CN |
| Crossed-roller bearings / 交叉滚子轴承 | Joint stiffness and precision | ~28 | Hengli, Luoyang LYC | CN |
| Chassis & frame | Lightweight aluminum structure | 1 | Tuopu Group (10% lighter) | CN |
| Magnets & energy | ||||
| Rare-earth magnets | NdFeB inside every motor | ~50 | JL Mag, Zhongke Sanhuan | CN |
| Battery pack | 2.3 kWh torso · 500 Wh/kg cell | 1 | CATL (custom cells) | CN |
| Thermal & cooling | Heat exchangers, fluid loops | — | Sanhua Intelligent | CN |
| Sensing | ||||
| Cameras | 8× Tesla-designed (auto-grade) | 8 | Tesla design + Sony image sensors | US/JP |
| IMU | 9-axis inertial measurement | 1–2 | Bosch / TDK / Murata | DE/JP |
| Joint torque sensors | Strain-gauge based, in each rotary joint | 14 | Custom (China-fabricated) | CN |
| Position encoders | Magnetic absolute encoders per joint | 28 | AMS Osram / domestic CN | EU/CN |
| Tactile / fingertip | Pressure sensing on dexterous fingers | 11+ | Custom + emerging suppliers | CN/US |
| Compute & AI | ||||
| Main inference chip | Tesla AI4 (current) → AI5 (next) | 1 | Tesla design · Samsung / TSMC fab | US |
| Vision-Language-Action model | Optimus Cortex (FSD-derived) | — | Tesla AI Cortex (50K-GPU cluster) | US |
| Networking / connectivity | Wi-Fi 6 / 5G / BT | 1 | Qualcomm | US |
Estimated BOM value by country
What Tesla makes vs. what they buy
Tesla designs four things: chassis architecture, the AI4/AI5 inference chip, the Optimus Cortex VLA model, and actuator system specs. Everything else is sourced — and at the BOM level, Sanhua, Tuopu, Suzhou Green Harmonic, Zhaowei, JL Mag, and CATL together supply ~70% of an Optimus body's dollar value.
Musk's October 2025 order with Sanhua alone is reportedly worth $685M, with delivery starting from Sanhua's Mexican plant in Q1 2026 — a tariff-routing move, not a sourcing one.
Where public Tesla/industry sources vary
Actuator count: this atlas keeps 28 as the body-actuator count for Optimus Gen 2. Public sources often discuss hands separately, so hand DoF, finger drives, and Gen 3 hand claims should not be mixed into the same count without labeling the scope.
Hand value: public reporting and public patents/commentary put more emphasis on dexterous hand / 灵巧手 value, especially coreless motors / 空心杯电机, tendon transmission / 腱绳传动, and fingertip sensing / 指尖传感. That makes the hand less of an accessory and more of a separate subsystem market.
Top 4 humanoids — component-by-component, country by country
Tesla and Figure are designed in California; Unitree and UBTech are end-to-end Chinese. Public sources strengthen the hardware industrialization thesis, but they do not validate every Figure-specific supplier claim. Treat Figure's supplier mix as undisclosed unless Figure or a named partner publishes it.
US designCN build
USA
China
China
| Spec | Optimus Gen 2 | Figure 02 | Unitree H1 | Walker S2 |
|---|---|---|---|---|
| Physical | ||||
| Height | 173 cm | 168 cm | 180 cm | 176 cm |
| Weight | ~57 kg | 70 kg | 47 kg | 70 kg |
| Total DOF / actuators | 28 body actuators; hand counts tracked separately | 28 | 19 | 52 |
| Hand DOF | 11 (Gen 2) → 22 (Gen 3) | 16 | 0 (no hands base) | 7 per hand |
| Payload | ~9 kg | 25 kg | ~30 kg arm capacity | 15 kg |
| Walking speed | ~5 mph | ~4.3 mph | 3.3 m/s record | ~3.6 mph |
| Actuators | ||||
| Joint torque (peak) | 180 Nm rotary, 8,000 N linear | 450 Nm | 360 Nm (knee) | ~250 Nm (waist) |
| Rotary actuator | Tuopu, Sanhua CN | Custom in-house US | M107 motor (Unitree) CN | Custom UBTech CN |
| Linear actuator | Sanhua ($685M order), PRS / 行星滚柱丝杠 CN | Custom US | None (rotary-only) CN | Custom CN |
| Reducers | Suzhou Green Harmonic CN | Undisclosed; custom/in-house module design plus external precision vendors US/? | In-house planetary CN | Mixed CN + JP CN/JP |
| Planetary roller screws / 行星滚柱丝杠 | Beite, XCC, Best CN | Undisclosed; do not treat THK/NSK as confirmed ? | N/A CN | Hiwin / domestic mix TW/CN |
| Precision bearings / 交叉滚子轴承 | Hengli, LYC assumed CN | Undisclosed; likely external piece-part vendors ? | Domestic CN CN | Domestic CN CN |
| Magnets (NdFeB) | JL Mag, Sanhuan CN | Undisclosed ? | Domestic CN CN | Domestic CN CN |
| Sensors | ||||
| Cameras | 8 Tesla-designed US | 6 (custom) US | Intel RealSense US | Dual RGB stereo CN |
| LiDAR / depth | None (vision-only) US | None (vision-only) US | MID-360 (Unitree) CN | Depth LiDAR CN |
| Force / torque | 14 joint sensors US/CN | Integrated US | Joint torque feedback CN | Arms + wrists CN |
| Tactile (fingers) | 11+ pressure points CN/US | Integrated finger units US | None on base hand CN | Fingertip sensors CN |
| IMU | 9-axis (Bosch/TDK) DE/JP | Custom US | 9-axis CN | 6-axis CN |
| Compute & AI | ||||
| Main inference chip | Tesla AI4 → AI5 US | Nvidia Jetson Thor US | Dual-computer (x86+ARM) US/CN | Intel i7 + Nvidia Orin US |
| AI / VLA model | Optimus Cortex US | Helix (in-house) US | Open-source + Unitree CN | BrainNet 2.0 + Co-Agent CN |
| Power | ||||
| Battery capacity | 2.3 kWh (CATL custom) CN | 2.25 kWh torso US | 864 Wh swappable CN | Dual-pack hot-swap CN |
| Runtime | ~5 hr (target) | 5 hr · 1.5 hr charge | ~2 hr | 24/7 (3-min swap) |
| Manufacturing & market | ||||
| 2025 production target | 5,000 → 100K (2026) | ~thousands (BMW) | ~hundreds (commercial) | Mass production launched |
| Assembly site | Fremont CA + Shanghai US/CN | Sunnyvale CA US | Hangzhou CN | Shenzhen CN |
| Manufacturing-chain layer / 制造链主 | Tesla + component partners | Figure BotQ; in-house assembly of core modules plus external piece-part vendors | In-house China supply chain | In-house China supply chain |
| % BOM from China (est.) | ~70% | Undisclosed; prior estimate removed | ~95% | ~95% |
Three takeaways from the side-by-side
Capability spread is wider than geography spread. Walker S2 has the most DOF (52) and is built for 24/7 industrial shifts via 3-min battery swaps. Figure 02 wins on raw torque (450 Nm) and hand dexterity (16 DOF). Optimus targets the lowest cost ($20K) and largest scale (100K in 2026). Unitree H1 is the lightest and fastest research platform.
The U.S. lead is narrow and concentrated. AI compute (Tesla AI5, Nvidia Jetson Thor) and VLA models (Optimus Cortex, Helix) are the only layers where U.S. firms own end-to-end IP. Below those layers — actuators / 执行器, reducers / 减速器, screws / 丝杠, magnets / 磁材, batteries / 电池 — Chinese suppliers either dominate or compete at every Tesla price point.
Figure is the geographic outlier, but its supplier map is mostly undisclosed. Figure publicly says it vertically integrates core module assembly and uses external vendors for piece-part manufacturing. That supports a different manufacturing strategy from Tesla's disclosed China-heavy stack, but not a named THK/NSK/Harmonic supplier map.
China's advantage is industrialization, not just cheap parts. Public market and company sources repeatedly point to manufacturing-chain capability: module design, supplier management, tooling, testing, and after-sales feedback. That makes ODM/OEM and integration platforms a real strategic layer.
PI and Skild are not hardware suppliers — they are competing to own the robot brain
Physical Intelligence and Skild AI sit above the component BOM. Their bet is that a general robot foundation model / 机器人基础模型 becomes the reusable intelligence layer across humanoids, arms, mobile manipulators, and industrial robots.
Physical Intelligence / PI / π
- Positioning: general-purpose physical AI / 通用具身智能 that can control many robot bodies and tasks.
- Model line: π0, π0.5, and π0.7. The April 2026 π0.7 release is framed as a steerable VLA / 视觉-语言-动作模型 with improved compositional generalization.
- Technical wedge: diverse multimodal prompts, visual subgoals / 视觉子目标, cross-embodiment transfer / 跨本体迁移, online RL / 在线强化学习, and memory for longer tasks.
- Atlas implication: PI is a threat to robot-specific autonomy stacks if its model transfers across arms, mobile manipulators, and home tasks without full retraining.
Skild AI
- Positioning: Skild Brain is an omni-bodied robot brain / 多本体机器人脑: any robot, any task, one shared model.
- Deployment wedge: industrial partnerships with ABB Robotics, Universal Robots, MiR, NVIDIA, and Foxconn-linked Blackwell production work.
- Technical wedge: internet-scale human video / 人类视频, simulation / 仿真, synthetic data / 合成数据, sim-to-real transfer / 仿真到现实迁移, and task data post-training.
- Atlas implication: Skild is the most enterprise-deployment-oriented of the robot-foundation-model startups in this set; it is trying to attach to existing robot OEM installed bases rather than sell a humanoid body.
| Company | What it owns | Evidence found | What remains unclear |
|---|---|---|---|
| Physical Intelligence / PI | Generalist VLA models / 通用VLA模型 for physical tasks | π0 open-source release; π0.5 open-world generalization; π0.7 steerable model with cross-embodiment and compositional generalization claims | Commercial deployment depth, pricing, OEM integration model, and production economics |
| Skild AI | Skild Brain / 多本体机器人脑 for many robot types | ABB, Universal Robots, MiR, NVIDIA collaborations; Foxconn/NVIDIA Blackwell assembly use case; reported $14B valuation in Jan. 2026 | Production revenue quality, customer concentration, model performance outside partner demos, and hardware-control boundaries |
| Figure Helix | Robot-specific VLA stack inside Figure's own humanoids | Public Figure 02/03 and Helix materials show in-house integration with Figure robots | Whether Helix becomes a platform outside Figure hardware or stays vertically integrated |
| Tesla Optimus Cortex | Tesla-owned autonomy stack tied to Optimus and FSD data/compute | Clear vertical integration around Tesla AI compute and Optimus hardware strategy | How broadly the model generalizes outside Tesla-controlled environments and tasks |
Why this changes the atlas
The first atlas treated AI mainly as an internal capability of Tesla, Figure, Unitree, and UBTech. PI and Skild add a separate software-platform thesis: the winner might not build the body at all, but instead provide the robot brain / 机器人脑 that turns existing hardware into adaptive labor. This does not weaken the Chinese hardware thesis; it adds a U.S.-led model layer above it.
Where the data came from
Primary sources are public: IFR (World Robotics 2025), the Shenzhen Robotics Industry White Paper 2025, IEA, public company posts, public market-research abstracts, and public robotics coverage. Private/local clippings are excluded from the public atlas.
Market & installations
- IFR — World Robotics 2025: global robot demand doubles over 10 years
- IFR — World Robotics 2025 Industrial Robots executive summary
- IFR — China Tops World Record of 2 Million Factory Robots
- IFR — World Robotics 2025: Americas press release
- China Daily — China surges ahead as global robotics powerhouse
- 深圳市机器人产业发展白皮书 (2025) — Shenzhen Robotics Industry White Paper
- Statista — China's robotics boom leaves all other nations behind
- Straits Research — United States Industrial Robots Market Size
- IMARC — China Industrial Robotics Market Size, Share and Report
- TechNode — China Installed Nearly 300,000 Industrial Robots in 2024
Supply chain, components & humanoids
- Morgan Stanley — Mapping the Humanoid Robot Value Chain
- McKinsey — Humanoid robots: Crossing the chasm from concept to commercial reality
- Bank of America Institute — Humanoid Robots 101
- Interact Analysis — Joint actuators: power and dexterity
- IEA — China's share in rare earth magnet production, 2024
- Adamas Intelligence — Record rare earth magnet exports from China in 2024
- Yicaí Global — Three core supply chains of humanoid robots
- 36Kr — Elon Musk Places $685M Order with Sanhua for Optimus
- OptimusK Blog — Tesla Optimus Supply Chain: Who Makes the Parts
- HumanoidSpecs — Tesla Optimus Gen 2
- The Robot Report — Figure 02 humanoid robot is ready to get to work
- Figure — BotQ: A High-Volume Manufacturing Facility for Humanoid Robots
- Figure — Introducing Figure 03
- Unitree — H1 product page
- UBTech — Walker S2 product page
- CNBC — China ships more humanoid robots than the U.S.
- SCMP — UBTech surges as humanoid robot sales jump 23-fold
- TrendForce — Optimus bottlenecks and dexterous hand challenges
- QYResearch / Valuates — Planetary Roller Screw for Humanoid Robot market
- China Energy News — Donghai Securities humanoid components report summary
- Eastmoney — Donghai Securities report page
- Sina Finance — Yongxing Securities 2026 machinery strategy
- HumanoidSpecs — Learn about humanoid robot component stack
Robot AI / foundation models
- Physical Intelligence — company and research index
- Physical Intelligence — π0.7: a steerable model with emergent capabilities
- Physical Intelligence — Open Sourcing π0
- Skild AI — partnering with ABB Robotics, Universal Robots, and NVIDIA
- Skild AI — expanded collaborations announcement
- TechCrunch — Skild AI hits $14B valuation