Seven Companies, Three Countries, One Race: Who Actually Controls the Humanoid Supply Chain

Open up any humanoid robot and you will find a geography lesson. The batteries are almost certainly Chinese. The AI processor is almost certainly American. The precision gears may be Japanese. The depth sensors might be German, Israeli, or American depending on the model. And the rare earth elements inside the magnets that make the actuators move probably came from a mine in Inner Mongolia regardless of who assembled the final product.

The humanoid robot race is usually framed as a competition between companies. Unitree versus Tesla. AgiBot versus Figure AI. But that framing misses the deeper contest happening underneath. The real race is between supply chains, and no single country controls the entire stack. This creates dependencies that could reshape the industry overnight if any link breaks.

This is the story of those links.

Anatomy of a humanoid robot: where the money goes

Before tracing the supply chain, it helps to understand what a humanoid robot actually is in terms of components. Strip one down and you get five major subsystems, each with its own supplier ecosystem, cost structure, and geopolitical exposure.

The cost breakdown of a typical humanoid robot looks roughly like this:

These numbers vary by manufacturer and price tier. A $16,000 Unitree G1 allocates proportionally more to actuators and structure because it uses a less expensive compute platform. A $150,000 Figure 02 spends proportionally more on compute and sensors. But the general ranking holds across the industry: batteries first, actuators second, compute and sensors competing for third.

What makes this breakdown geopolitically interesting is that no country dominates all five categories. China leads in batteries and actuators. The United States leads in AI compute. Japan leads in precision gears. Sensors are split across multiple countries. And the raw materials underneath all of it follow their own geography entirely.

Batteries: CATL and the 30-40% cost advantage

The single most expensive component in a humanoid robot is its battery pack. For a robot that needs to operate for two to eight hours while carrying its own weight plus a payload, battery capacity is not optional. It is load-bearing in every sense.

And this is where China’s advantage is most decisive.

CATL (Contemporary Amperex Technology Co., Limited) and BYD together control roughly 77% of the global EV battery market. The same lithium-iron-phosphate (LFP) and nickel-manganese-cobalt (NMC) cells that power electric vehicles also power humanoid robots. When Unitree builds a G1, when AgiBot assembles an A2, when UBTECH produces a Walker S2, they are drawing from a battery supply chain that China spent over a decade and hundreds of billions of dollars building for the electric vehicle industry.

The numbers are stark. According to BloombergNEF, the average price of a lithium-ion battery pack in China fell to roughly $94 per kilowatt-hour in 2025. The global average outside China was closer to $133 per kWh. That gap translates directly into the finished cost of a humanoid robot.

Consider the battery requirements across current humanoid platforms:

A 2,300 Wh battery pack at $94/kWh costs roughly $216. The same pack at $133/kWh costs $306. That is a $90 difference on batteries alone. Scale that across thousands of units and the gap compounds into a massive cost advantage for Chinese manufacturers who can source batteries domestically.

But it is not just price. Chinese humanoid makers benefit from proximity. CATL’s headquarters in Ningde, Fujian Province is a short truck ride from the electronics manufacturing clusters in Shenzhen and the robotics hubs in Shanghai and Hangzhou. American companies face longer lead times, higher shipping costs, and, increasingly, tariff complications when sourcing Chinese battery cells.

What happens if this link breaks

A US-China trade decoupling scenario would hit American humanoid companies hard on batteries. Samsung SDI and LG Energy Solution in South Korea offer alternatives, but at 20-30% higher cost and with less available capacity. The Inflation Reduction Act has spurred domestic US battery production, but those factories are focused on EV cells. Retooling for the smaller, specialized packs that humanoid robots require would take years.

Figure AI appears to recognize this risk. The company has invested in producing batteries in-house at its BotQ manufacturing facility, one of the few Western humanoid companies attempting vertical integration on battery production. Whether this approach can match the cost efficiency of CATL remains to be seen.

Actuators: the muscles nobody talks about

If batteries are the heart of a humanoid robot, actuators are its muscles. Every joint, every finger, every head tilt is driven by an actuator. A typical humanoid has 20 to 82 of them depending on the number of degrees of freedom. The quality, cost, and availability of actuators is arguably the most important supply chain story in humanoid robotics, and it is the least discussed.

An actuator for a humanoid robot joint typically contains three critical subcomponents: a brushless DC motor, a precision gear reducer (often a harmonic drive or strain wave gear), and a control board with position and torque sensors. Each of these has its own supply chain story.

The rare earth problem

The permanent magnets inside brushless DC motors require neodymium-iron-boron (NdFeB) alloys. China controls roughly 60% of global rare earth mining and an even larger share of processing and refining. In 2025, China tightened export controls on several rare earth elements, sending prices for neodymium and dysprosium up 15-25%.

This matters because every humanoid robot contains dozens of these magnets. The Unitree G1, with 23 degrees of freedom, has at least 23 motors with NdFeB magnets. The XPENG Iron, with 82 degrees of freedom, has even more. Chinese manufacturers buy these magnets at domestic prices. Everyone else pays export prices plus shipping and tariffs.

Precision gears: Japan’s quiet leverage

The gear reducers inside humanoid actuators are among the most precision-engineered components in the entire robot. Harmonic Drive Systems, a Japanese company, has dominated this market for decades. Their strain wave gears can achieve backlash of less than one arc-minute, which is essential for precise joint control.

But Harmonic Drive cannot supply the entire humanoid industry. Their production capacity has historically served industrial robots, which need perhaps six gears per arm. A humanoid needs 20 to 50 precision gears. As humanoid production scales, the gear bottleneck becomes acute.

Chinese companies are aggressively building alternatives. Leaderdrive in Suzhou, Green Harmonic in Shenzhen, and Laifual Drive in Beijing are all producing harmonic-style gears that have reached 85-90% of Japanese quality at 40-60% of the cost. For the cost-focused Chinese humanoid industry, “good enough” gears at half the price are a perfectly rational choice. Unitree and AgiBot both source primarily from domestic suppliers.

Fourier’s actuator innovation

Fourier Intelligence has developed its own FSA (Fourier Smart Actuator) 2.0 series, delivering 380+ Nm of peak torque. By designing actuators in-house, Fourier controls a critical piece of its own supply chain. The partnership with BASF for advanced materials gives them another edge: lighter, more durable actuator housings that reduce overall robot weight.

Figure AI has taken a similar approach, producing actuators in-house at its BotQ facility. This vertical integration costs more upfront but reduces dependency on external suppliers. For well-funded American startups, building your own actuators is a strategic hedge against supply chain disruption. For smaller companies, it is simply unaffordable.

AI compute: NVIDIA’s near-monopoly

Turn from the physical to the digital and the geography flips. Where China dominates batteries and actuators, the United States holds an equally commanding position in AI compute, primarily through one company: NVIDIA.

Across the humanoid robots tracked on this site, the processor landscape looks like this:

NVIDIA’s Jetson platform has become the de facto standard for humanoid robot compute. The Jetson Orin, with up to 275 TOPS (Trillion Operations Per Second) of AI performance, runs the brains of the Unitree G1, AgiBot A2, Apptronik Apollo, 1X NEO, NEURA 4NE-1, and Leju Kuavo. Boston Dynamics uses NVIDIA Isaac GR00T for Atlas. Fourier Intelligence trains its robots using NVIDIA Isaac Gym and deploys with TensorRT.

This concentration is not accidental. NVIDIA’s CUDA ecosystem, built over two decades, creates a software moat that is extraordinarily difficult to replicate. When a robotics researcher trains a neural network for locomotion or manipulation, they almost certainly use PyTorch on NVIDIA GPUs. When that model needs to run on a robot in real time, the NVIDIA Jetson platform offers the most optimized inference pipeline. The entire software stack, from training to deployment, assumes NVIDIA hardware.

The GR00T foundation model

In 2024, NVIDIA launched GR00T (Generalist Robot 00 Technology), a foundation model specifically designed for humanoid robots. GR00T provides pre-trained capabilities for locomotion, manipulation, and human interaction that robot companies can fine-tune for their specific hardware. Boston Dynamics, Apptronik, Fourier, and NEURA Robotics have all announced GR00T integrations.

This is a powerful lock-in strategy. If your robot’s AI foundation is trained, optimized, and deployed on NVIDIA’s stack, switching to an alternative compute platform means retraining models, rewriting inference code, and accepting performance regressions. For most companies, the switching cost makes NVIDIA’s dominance self-reinforcing.

The exceptions

Two notable exceptions stand out. Tesla uses its own custom FSD (Full Self-Driving) chip for Optimus, leveraging the same silicon it designed for autonomous vehicles. This gives Tesla independence from NVIDIA for inference, though Tesla still uses NVIDIA GPUs for training on its Cortex supercomputer cluster.

XPENG’s Iron humanoid uses three custom “Turing” AI chips delivering a combined 3,000 TOPS, the highest compute density of any humanoid robot. XPENG, like Tesla, came from the automotive industry and has the scale and revenue to design custom silicon. For everyone else, NVIDIA is the only practical choice.

Sensors: the most fragmented layer

If batteries are dominated by China and compute by the United States, the sensor layer is where things get truly tangled. Humanoid robots use an array of sensor types, and each type has its own dominant suppliers scattered across different countries.

LiDAR: China’s quiet takeover

LiDAR sensors, which provide 3D spatial mapping for navigation, tell a supply chain story that mirrors the broader industry dynamics. Five years ago, Velodyne (American) dominated the LiDAR market. Today, Chinese companies Hesai and RoboSense have captured the majority of automotive LiDAR shipments and are increasingly supplying humanoid robot makers.

The UBTECH Walker S2 uses LiDAR for navigation. The AgiBot A2 carries 3D LiDAR among its sensor suite. The Engine AI SE01 combines LiDAR with depth and infrared cameras. Boston Dynamics Atlas uses LiDAR alongside stereo cameras. In nearly every case, the most cost-effective LiDAR option is Chinese.

Hesai’s manufacturing scale, driven by the massive Chinese autonomous vehicle market, has driven LiDAR prices down by 70% since 2022. An automotive-grade LiDAR unit that cost $1,000 in 2023 now costs under $300. For humanoid robots, which need smaller, lighter LiDAR units, costs are even lower.

Intel RealSense: the depth camera standard

For close-range 3D vision, Intel’s RealSense depth cameras have become something of a standard in robotics research and commercial applications. The Leju Kuavo uses Intel RealSense D435i sensors on both its head and chest. The technology provides structured-light and stereo depth perception at relatively low cost.

But Intel’s commitment to RealSense has wavered. The company announced in 2022 that it would wind down its RealSense business, only to reverse course as demand from the robotics industry surged. This uncertainty underscores a broader vulnerability: key sensor technologies sometimes sit inside divisions of large companies that view them as non-core.

Chinese alternatives like Orbbec (based in Shenzhen) are rapidly closing the gap, offering depth cameras with comparable specifications at lower prices. Once again, the pattern repeats: initial Western technology leadership followed by Chinese volume production at lower cost.

The IMU supply chain

Inertial Measurement Units, the sensors that tell a robot which way is up and how fast it is moving, are sourced primarily from Bosch (Germany) and TDK InvenSense (Japan/US). These components are relatively inexpensive (often under $10 per unit) and widely available, making them one of the least geopolitically risky parts of the supply chain.

However, high-precision IMUs for demanding applications, such as maintaining balance during dynamic walking, come from a smaller set of suppliers and cost significantly more. The difference between a consumer-grade IMU and one suitable for bipedal locomotion can be a factor of fifty in price.

The cost advantage, quantified

All of these supply chain dynamics combine to give Chinese humanoid makers a substantial and well-documented cost advantage. The gap is not subtle.

The Engine AI SE01 can sell for approximately $12,000 because it sources virtually every physical component from within the Pearl River Delta and greater Yangtze River Delta manufacturing clusters. Its dual NVIDIA+Intel processors are the most expensive individual components, and even those benefit from NVIDIA’s China pricing and bulk supply agreements.

By contrast, Figure AI’s robots carry the cost of American labor, American real estate, and the decision to produce batteries and actuators in-house rather than buying from Chinese suppliers. Those choices buy independence, but they cost money.

The seven companies that matter

The supply chain story becomes concrete when you look at how specific companies have positioned themselves. Seven companies, representing three countries, illustrate the different strategies for managing supply chain risk.

Unitree Robotics (China) - The supply chain optimizer

Unitree’s entire business model is built on supply chain efficiency. The G1 runs on an NVIDIA Jetson Orin (275 TOPS), uses Chinese-sourced LFP battery cells, and leverages Hangzhou’s dense network of actuator and sensor suppliers. By keeping everything within a few hundred kilometers of its 50,000 square meter factory, Unitree achieves the lowest cost structure of any humanoid maker.

Supply chain risk: Low domestically, but fully dependent on NVIDIA for AI compute. A Jetson export restriction would force a costly pivot to domestic alternatives.

AgiBot (China) - Vertical integration, Chinese style

AgiBot took Unitree’s supply chain advantage and added scale. Its Shanghai factory in the Lingang Special Area can draw from both the local supply chain cluster and AgiBot’s corporate connections to BYD (batteries) and SAIC Motor (automotive-grade actuators). The A2 model uses an NVIDIA Jetson Orin 64G and combines LiDAR, RGB-D cameras, and fisheye cameras in a sensor suite that mixes domestic and imported components.

Supply chain risk: Similar to Unitree. Strong domestic positioning, NVIDIA dependency on compute.

Tesla (USA) - The only true vertical integrator

Tesla is unique in the humanoid industry because it designs its own AI chip (the FSD SoC) and manufactures much of its own supply chain. The Optimus uses Tesla Vision cameras (the same hardware from Tesla vehicles), a custom FSD chip for inference, and battery cells from Tesla’s own cell production lines. Tesla also trains on its own Cortex supercomputer cluster rather than relying on external cloud compute.

Supply chain risk: Lower than most American competitors due to vertical integration. However, Tesla still depends on Chinese battery materials (lithium, cobalt, nickel) and rare earth magnets for actuator motors. The company’s long-term target price of $20,000-$30,000 per unit will require cost reductions that may necessitate Chinese component sourcing.

Figure AI (USA) - Building independence from scratch

Figure AI has made the most aggressive bet on supply chain independence among Western startups. The company produces robots, batteries, actuators, and control systems in-house at its BotQ facility. The $1.85 billion in total funding helps finance this approach. Investors include NVIDIA, Intel, Microsoft, and Amazon, essentially a who’s-who of the American tech supply chain.

Supply chain risk: Moderate. In-house production reduces dependency but increases cost. Figure must achieve manufacturing efficiency that typically requires years of iteration. The company’s 12,000-unit annual target for BotQ will test whether vertical integration can scale.

Boston Dynamics (USA/South Korea) - The Hyundai advantage

Boston Dynamics benefits from a unique position: American engineering with South Korean manufacturing backing. Hyundai Motor Group, which acquired Boston Dynamics for $1.1 billion in 2021, brings automotive supply chain expertise, battery technology from Hyundai’s EV division, and manufacturing scale. The Electric Atlas uses hot-swappable battery packs and integrates Google DeepMind’s Gemini Robotics alongside NVIDIA Isaac GR00T.

Supply chain risk: Diversified. Hyundai provides access to South Korean battery suppliers (Samsung SDI, LG Energy Solution) as an alternative to Chinese cells. The Google DeepMind partnership reduces NVIDIA dependency on the AI model side, though Atlas still uses NVIDIA compute hardware. The planned 30,000-unit/year factory benefits from Hyundai’s automotive factory expertise.

UBTECH Robotics (China) - The government-backed champion

UBTECH, listed on the Hong Kong Stock Exchange, benefits from the full weight of China’s industrial policy. The Walker S2 is deployed in NIO and BYD factories. With 800 million yuan in orders, UBTECH has the volume to negotiate favorable terms with every domestic supplier. The Walker S2 uses a combination of Intel Core i7 and NVIDIA Jetson AGX Orin for compute, plus a cloud-based BrainNet 2.0 system for cognitive functions.

Supply chain risk: Low for domestic components. Dual Intel/NVIDIA compute dependency is notable. UBTECH’s expansion into the Middle East (NEOM project) will test whether the Chinese supply chain can support global deployments.

1X Technologies (Norway/USA) - The European outlier

1X is the only non-Chinese, non-American company in the top tier of humanoid shipments. Manufacturing from Moss, Norway, the company uses NVIDIA Jetson AGX Orin compute and builds its distinctive tendon-driven actuators in-house. The NEO’s 842 Wh battery and 30 kg weight reflect a design philosophy that prioritizes energy efficiency over brute force.

Supply chain risk: High. Norway lacks a domestic robotics supply chain. 1X must import nearly every component: NVIDIA chips from the US, battery cells from Asia, sensors from multiple countries. The OpenAI partnership provides AI capability but adds another American dependency. The company’s plan to scale to tens of thousands of units will stress-test this distributed supply chain.

The geopolitical dependencies map

The dependency map reveals a striking asymmetry. Chinese humanoid companies depend on the United States primarily for one thing: AI compute chips. American humanoid companies depend on China for multiple things: batteries, actuator components, rare earth magnets, and increasingly LiDAR sensors.

This means the leverage is not balanced. The US could theoretically restrict NVIDIA chip exports and hobble Chinese humanoid companies in the short term. But China could restrict rare earth exports, battery material exports, and processed magnet exports simultaneously, hitting American companies across multiple subsystems at once.

The Taiwan variable

The supply chain analysis is incomplete without mentioning Taiwan. TSMC (Taiwan Semiconductor Manufacturing Company) fabricates the silicon for both NVIDIA’s Jetson processors and many of the sensor chips used in humanoid robots. A disruption in Taiwan would affect every humanoid company regardless of nationality, making it the single point of failure for the entire global industry.

NVIDIA’s move toward domestic US fabrication with Samsung and Intel Foundry provides some hedge, but leading-edge chip production remains concentrated in Taiwan for now.

The three scenarios nobody is planning for

Scenario 1: US restricts NVIDIA Jetson exports to China

If the US tightened export controls to include the Jetson Orin module (currently permitted), it would force Unitree, AgiBot, Leju, and other Chinese humanoid companies to switch to domestic alternatives. Huawei’s Ascend 310/910 processors are the most likely fallback, but they lack the CUDA software ecosystem. Chinese companies would face 12-18 months of disruption while rewriting software stacks and retraining models.

The ironic outcome: this would accelerate Chinese development of competitive AI chip alternatives, reducing American leverage in the long term. The same dynamic played out with Huawei smartphones, where US sanctions ultimately pushed China to develop the Kirin 9000s chip.

Scenario 2: China restricts rare earth magnet exports

If China imposed export quotas on NdFeB magnets or the rare earth elements that go into them, it would hit actuator production for every non-Chinese humanoid maker. Alternative magnet technologies exist (ferrite magnets, for example) but deliver significantly less torque density, which means heavier, larger, and less capable actuators. The only near-term mitigation would be stockpiling, which several companies have reportedly begun doing.

Scenario 3: Harmonic Drive hits capacity limits

Harmonic Drive Systems already operates near capacity. If humanoid production scales to the tens of thousands of units per year that multiple companies are targeting for 2027, precision gear supply could become a binding constraint for Japanese and Western manufacturers. Chinese gear makers would benefit as humanoid companies accept the quality trade-off in exchange for availability.

Who wins the supply chain race

The supply chain analysis points to an uncomfortable conclusion. In the near term, China has the structural advantage. Chinese humanoid companies can source 80% of their bill of materials domestically, with only AI compute as a significant foreign dependency. American companies, by contrast, face foreign dependencies across multiple critical subsystems.

But the long-term picture is less clear. Three forces could shift the balance:

Vertical integration by funded American companies. Figure AI, Tesla, and to some extent Boston Dynamics (via Hyundai) are all investing in producing critical components in-house. If they succeed, American companies could reduce their Chinese supply chain exposure to raw materials only.

Chinese AI chip development. Huawei, Horizon Robotics, and several other Chinese chipmakers are building AI inference processors. They are two to three years behind NVIDIA on performance, but the gap is narrowing. If Chinese chips reach “good enough” performance for humanoid applications, NVIDIA’s leverage evaporates.

The raw materials wildcard. Lithium, cobalt, nickel, and rare earths are the foundational inputs for the entire supply chain. China dominates processing of most of these materials today, but mining investments in Australia, Canada, the Democratic Republic of Congo, and Chile are diversifying the upstream supply. This diversification takes a decade to mature, but it is happening.

The supply chain timeline

What this means for the industry

The humanoid robot supply chain is not a neat story with a single winner. It is a tangled web of mutual dependencies where every country holds some leverage and no country holds enough.

China can build cheaper humanoid robots today because its supply chain was built for electric vehicles and consumer electronics. The United States can build smarter humanoid robots today because its AI chip and software ecosystem is unmatched. Japan provides precision components that both sides need. South Korea offers diversification. Europe contributes materials science and sensor technology.

The companies that navigate this web most skillfully will win. That means Unitree and AgiBot cannot ignore their NVIDIA dependency, and Tesla and Figure cannot ignore their materials dependency. Vertical integration helps at the margin, but no company can realistically produce everything in-house.

For the industry as a whole, the lesson is structural. The humanoid robot is the most geopolitically complex consumer product ever conceived. Every joint is a trade policy risk. Every sensor is a diplomatic relationship. Every battery cell is a mining concession on another continent.

The race to build humanoid robots is not just about engineering. It is about supply chains. And in 2026, the supply chain is the strategy.