The year 2025 has marked a definitive turning point in the history of artificial intelligence and robotics. While the last decade saw the United States leading in fundamental AI research and generative models, the physical manifestation of this intelligence—embodied AI—has found a new epicenter. According to recent industry reports, China has secured a staggering lead in the humanoid robotics sector, deploying approximately 13,000 of the 16,000 humanoid robots installed globally in 2025. This 80% market share signals not just a manufacturing victory, but a strategic leap in the race to gather the real-world data necessary to build the next generation of AI "world models."
For observers at Creati.ai, this development represents more than just industrial statistics; it highlights a divergence in strategy. While Western companies have focused on perfecting the "brain" of the robot through simulation and controlled demos, Chinese firms have prioritized putting "bodies" into the real world. This approach is creating a data flywheel effect that could accelerate the maturity of general-purpose robots far faster than previously anticipated.
The scale of China's dominance in 2025 is illustrated by the shipment volumes of its top domestic manufacturers compared to their Western counterparts. Data reveals that Shanghai-based Agibot and Hangzhou-based Unitree Robotics have effectively cornered the market, leaving high-profile American competitors like Tesla and Figure AI trailing significantly in terms of actual units deployed.
Agibot alone shipped over 5,000 units, capturing nearly one-third of the global market. These robots are not merely prototypes confined to R&D labs; they are being deployed in manufacturing logistics, automotive assembly lines, and educational institutions. In contrast, despite the immense media attention surrounding Tesla’s Optimus, the company secured less than a 5% share of the global market in 2025, with deployments largely restricted to internal testing or pilot programs.
The following table breaks down the shipment estimates for the leading players in the humanoid robotics space for 2025:
Global Humanoid Robot Shipments by Manufacturer (2025)
Manufacturer|Headquarters|Est. Units Shipped|Global Market Share
---|---|---
Agibot (Zhiyuan)|Shanghai, CN|~5,100|32%
Unitree Robotics|Hangzhou, CN|~4,200|26%
UBTECH|Shenzhen, CN|~1,000|6%
Tesla|Texas, USA|< 800|< 5%
Figure AI|California, USA|~300|< 2%
Others (Leju, Fourier, etc.)|Global|~4,600|29%
The implications of these deployment numbers extend far beyond hardware sales revenue. In the AI industry, data is the primary currency. Current Large Language Models (LLMs) have been trained on the entirety of the text-based internet, but they lack "physical common sense"—an understanding of gravity, friction, weight, and spatial manipulation. To bridge this gap, AI models need training data from the physical world, often referred to as "world models."
By deploying 13,000 active units, China is effectively building a massive distributed data collection network. Every step an Agibot takes in a factory and every object a Unitree robot manipulates contributes to a proprietary dataset that refines the robot's VLA (Vision-Language-Action) models. This creates a feedback loop: more robots generate more data, which leads to smarter models, which in turn justifies further deployment.
While US companies rely heavily on "Sim2Real" (simulation-to-reality) training—teaching robots in virtual video-game-like environments—Chinese firms are validating their algorithms in the messy, unpredictable reality of factory floors. This "brute force" application of hardware could allow Chinese AI models to leapfrog Western counterparts in robotic dexterity and adaptability, much like how the vast user base of WeChat allowed for rapid iteration in mobile software.
The primary driver of this rapid adoption is cost. Chinese manufacturers have successfully applied the supply chain efficiencies from the smartphone and EV sectors to robotics. The bill of materials (BOM) for Chinese humanoids has been aggressively reduced through the domestic production of core components such as harmonic drives, frameless motors, and lightweight sensors.
For instance, Unitree’s G1 humanoid robot was introduced at a price point ranging between $6,000 and $16,000, significantly lower than the projected $20,000–$30,000 price tag for Western equivalents like Tesla's Optimus (which has yet to reach mass commercial availability). This aggressive pricing strategy makes humanoid robots accessible not just to elite tech giants, but to mid-sized manufacturing firms and research universities.
Price and Application Comparison: China vs. West
| Model | Manufacturer | Est. Price (USD) | Primary Application |
|---|---|---|---|
| Unitree G1 | Unitree Robotics | $6,000 - $16,000 | Education, Research, Light Service |
| Agibot X2/G2 | Agibot | $15,000 - $25,000 | Industrial Manufacturing, Logistics |
| Optimus (Gen 2) | Tesla | Target: $25,000+ | Internal Automotive Manufacturing (Pilot) |
| Figure 02 | Figure AI | RaaS Model* | Warehouse Logistics (BMW partnership) |
| Digit | Agility Robotics | RaaS Model* | Logistics, Material Handling |
(Note: Many Western firms currently prefer "Robots-as-a-Service" (RaaS) leasing models over direct sales to manage high upfront hardware costs.)
The explosive growth of the Chinese humanoid sector is not accidental; it is the result of coordinated policy support. The Chinese government's "AI+" initiative treats artificial intelligence and robotics as essential infrastructure, similar to electricity or high-speed rail. Subsidies for factory automation, the establishment of national innovation centers, and tax incentives for robotics startups have lowered the barrier to entry.
In 2025, the Ministry of Industry and Information Technology (MIIT) emphasized the goal of mass-producing humanoid robots by 2025 and reaching advanced world-class levels by 2027. This top-down mandate has encouraged local governments in Shanghai, Shenzhen, and Beijing to compete for the title of "Robotics Capital," creating clusters of innovation where hardware supply chains are tightly integrated.
Conversely, Western markets face a more fragmented regulatory landscape and a venture capital environment that has recently become more cautious regarding hardware-intensive startups. While US investment in AI software remains robust, the capital-heavy nature of building factories for robots has slowed the scaling process for American hardware firms.
Looking ahead, analysts project that the gap in installation numbers may widen before it narrows. Counterpoint Research estimates that the global installed base of humanoid robots will exceed 100,000 units by 2027. If current trends persist, Chinese manufacturers are poised to capture the lion's share of this expansion, particularly in the industrial and automotive sectors which currently account for 72% of deployments.
However, the race is far from over. Western companies still hold a perceived edge in the sophistication of the underlying "brains"—the generalizable AI models that power these machines. The challenge for the US and Europe will be to solve the manufacturing bottleneck. Unless Western firms can ramp up production capacity and lower costs, they risk a future where the world's physical AI infrastructure runs on Chinese hardware, regardless of whose software is running in the cloud.
For the AI industry, the lesson of 2025 is clear: innovation without scale is merely a prototype. As we move into the era of embodied intelligence, the ability to manufacture, deploy, and maintain fleets of robots is becoming just as critical as the neural networks that guide them. China’s 80% market share is a wake-up call that the next frontier of AI will be fought not just in silicon, but in steel and servos.
In a significant escalation of the tech trade war, Chinese authorities have reportedly blocked imports of Nvidia's H200 AI chips, a move that comes despite the US government having cleared them for export.
TechRadar covers DeepSeek's new Engram technique that separates static memory from computation, reducing expensive HBM requirements and addressing the global DRAM shortage.
