China's robotics push creates software bottleneck despite hardware dominance
China produced 12,800 humanoid robots in 2025-about 90 percent of the global total-yet most remain tethered to preprogrammed tasks or remote control. The country's manufacturing strength masks a critical weakness: dependence on Nvidia for the AI software that would make robots truly autonomous.
This gap matters for developers. China's robotics sector is where the EV industry was a decade ago-a massive manufacturing ecosystem racing to build intelligent machines. But the software layer that enables real autonomy still comes from the US.
How China got here
China installed more industrial robots over the last five years than all other countries combined. Chinese companies supplied over 57 percent of the domestic market in 2024, the first time they overtook foreign competitors.
The advantage stems from overlap with electric vehicles. EV makers already control batteries, sensors, and supply chains. Companies like XPeng repurpose autonomous-driving sensors for robot navigation. GAC and Spirit AI adapted EV batteries for their robots. This vertical integration gives Chinese manufacturers a 63 percent share of key humanoid-robot component suppliers globally.
Cost reflects this advantage. Unitree's G1 humanoid costs around €11,650. Boston Dynamics' Atlas costs over €120,000. Tesla's Optimus runs about €25,800.
The precision and autonomy problem
Lower cost doesn't mean ready for factories. Chinese humanoids remain "only half as efficient as humans," according to UBTech's founder. Most deployed robots handle single, repetitive tasks-quality inspections, material sorting, simple assembly.
The real challenge is autonomy. Corporate videos show humanoids dancing or kickboxing, but these demonstrations are usually preprogrammed or teleoperated. They don't perceive their environment and respond in real time.
This is where software matters. Vision-language-action (VLA) models-AI systems that take visual input, process language, and generate robotic actions-represent the frontier. Google DeepMind leads with RT-2 and Gemini Robotics. Nvidia offers GR00T. Chinese teams including AgiBot, Galbot, and AI² Robotics are accelerating VLA development, but they lack foundational generalist models that Western competitors have built.
AgiBot announced its first general model GO-1 in 2025, which uses a visual-language model but isn't a full VLA. Galbot's three models-GraspVLA, TrackVLA, and GroceryVLA-each handle narrow tasks, illustrating how difficult building fully autonomous robots remains.
Where Nvidia fits in
Most major Chinese robotics players depend on Nvidia's stack. The company's Jetson modules embed its Blackwell GPUs and Isaac development platform directly into robots. UBTech, Galbot, Unitree, EngineAI, and AgiBot were among the first to receive Jetson Thor chips.
Nvidia provides simulation tools for training. Galbot builds synthetic datasets using Nvidia's Isaac Sim. Unitree trains robots in complex actions using Nvidia's Isaac Lab reinforcement-learning framework.
This dependency creates risk. If US export controls tighten around edge-computing chips-currently not restricted-Chinese robotics could stall. For now, these chips aren't the focus of US restrictions, but that could change as they become more powerful.
What's actually deployable
Industrial robots in Chinese factories today are fixed, single-purpose machines. They excel at assembly, inspection, material handling, picking, and sorting. These 556,000 units produced annually dwarf the 12,800 humanoids.
Kepler Robotics' K2 handles logistics-carrying loads up to 30 kg. UBTech's Walker series performs quality inspections and simple assembly in EV factories. Midea's six-armed MIRO U supports washing-machine assembly.
None operate fully autonomously across varied environments. They work in controlled, site-specific trials.
The cost barrier to scale
Commercial viability requires prices to fall significantly. Chinese humanoids average 300,000-500,000 yuan (€36,000-€62,500). Guotai Securities calculates the threshold for profitability at 160,000 yuan, based on annual labor costs of 80,000 yuan and a two-year return on investment.
Even domestically made components remain expensive. High-end precision parts still come from Europe and Japan. Germany's Schaeffler and Japan's THK and NSK supply 90 percent of special ball screws used for precise positioning.
A ZY&YR Innovation Hub ranking of overall component competitiveness rates Japan "A", China and Europe "B", and the US "C". Chinese firms lead in low- and mid-range segments. Foreign makers like ABB and FANUC still dominate high-end markets.
Policy backing the sector
Beijing allocated 1 trillion yuan (€120 billion) over 20 years to robotics through the National Venture Capital Guidance Fund and regional funds. The Ministry of Industry and Information Technology set up a Standardization Committee for Humanoid Robots.
Shanghai's "Action Plan for the Development of the Embodied Intelligence Industry 2025-2028" targets algorithmic breakthroughs. Provinces and cities subsidize up to 30 percent of automation project costs or offer discounts on humanoid purchases.
The government included embodied AI in the 2025 annual work report and the 15th Five-Year Plan through 2030, signaling long-term commitment.
The employment question Beijing isn't solving
Public anxiety about job displacement is growing. Estimates suggest robots could replace at least 70 percent of China's manufacturing jobs. China's nearly 300 million migrant workers-many with precarious employment and no social safety net-face particular vulnerability.
The talent shortage compounds the problem. China faces a 5 million-person AI talent gap even as youth unemployment sits around 17 percent.
Official rhetoric emphasizes "empowerment" over substitution. But studies conclude AI will displace jobs faster than it creates them, straining employment, wages, and the social security system. China's weak welfare infrastructure makes this transition difficult.
Policy documents increasingly acknowledge the risk. The 15th Five-Year Plan and "AI+" Action Plan both call for assessing AI's employment impact. Yet the government has no intention of building a welfare state, leaving solutions unclear.
What this means for IT development
For developers, the takeaway is straightforward: hardware production is concentrating in China, but the software that makes robots autonomous remains a Western advantage. Understanding generative AI and LLM concepts-particularly vision-language-action models-is essential for anyone working on robotics integration.
The dependency on Nvidia's ecosystem also matters. Developers should understand both the capabilities and constraints of edge-computing chips and how reinforcement learning frameworks apply to robot training.
China's robotics sector will likely follow the EV playbook: many domestic competitors, state support, cost reduction, and gradual technological catch-up. For European and US firms, the competition is real and accelerating. For developers, the skills gap between hardware production and autonomous software remains the defining challenge.
Your membership also unlocks:
