Rise of the robots

This article was first published in New Electronics in May 2025 and was written by Rosalie Chen, Senior Analyst at Third Bridge.

Earlier this year the seemingly sudden emergence of Chinese AI pioneer DeepSeek shook many previously safe assumptions about the AI market. In particular DeepSeek challenged the assumption that the US is dominant in AI and sent the shares of companies like Nvidia – heavily exposed to that assumption – tumbling. 

While the reality of DeepSeek hasn’t so far lived up to the hype, it nevertheless revealed that there’s little room for complacency in the rapidly evolving world of tech and away from the digital world of AI, another tech market could be poised to go mainstream. While the world’s attention has been focused on AI, global capital has been quietly rushing into humanoid robots and breakthroughs in this market are now on the horizon. Just as AI has shaken the world of white-collar work, robots too could usher in a new industrial revolution with the potential to lower manufacturing production costs and increase output.

The disruptive emergence of DeepSeek and the subsequent and ongoing trade war between the US and China signals an era of intensifying competition between these two powers for global tech hegemony. 

So, in the race to develop and deploy the armies of humanoid robots needed to build the future of manufacturing, could Chinese technology be poised to disrupt markets once more? 

Rise of the robots

The emergence of DeepSeek lifted the lid on the race for tech supremacy between China and the US in the world of AI and the same dynamic can be observed in the emerging humanoid robot market. In the US, Tesla’s Optimus has proven capable of walking autonomously and performing assembly tasks and is speeding toward commercialization with a target price of USD20,000. In China meanwhile, Unitree’s H1 robot has made it to the national stage, performing alongside dancers at the Spring Festival Gala. Indeed, Chinese companies like CATL and Fuyao have already begun integrating humanoid robots into manufacturing, proving their practical feasibility.

Behind the headlines however, success in this market is dependent upon a wide range of factors ranging from production costs to supply chain management with both China and the US having advantages and disadvantages at different stages of the production process. 

Hardware

While the race to dominate AI is largely about software, delivering commercially viable and deployable humanoid robots also brings with it a number of hard-ware challenges. Humanoid robots’ multimodal systems divide into upper‑limb and lower‑limb mechanics. Today, lower‑limb technology is mature enough to handle parkour, dancing, and other complex motions, and core algorithmic hurdles have largely been cleared. These systems typically run in‑house reinforcement‑learning models (on the order of 10 million parameters, three‑layer neural nets), which multiple Chinese and American firms have already closed the loop on.

By contrast, upper‑limb dexterity remains the principal bottleneck, plagued by insufficient manipulation precision and poor generalization. Two camps are pursuing solutions: most companies adapt open‑source large models for specific use‑case optimization; a few go straight for artificial general intelligence (AGI), building universal large models from the ground up.

Software

The predominant approach is still task‑specific training, which demands massive data sets and top‑tier AI model design. But industrial, commercial, and home environments are far more complex and variable, so robots will need much stronger generalization to flexibly handle real‑world conditions. Current demos excel only in single‑scenario tasks, and the future goal is AGI‑adjacent autonomous learning and iteration to minimize dependence on labeled scenario data.

Notably, very few companies worldwide can independently train full Vision–Language–Action (VLA) models. The heavy hitters here are China’s Zhiyuan Robotics and U.S. players like Figure and Tesla.

A sub‑branch, Vision–Language–Manipulation (VLM) (focused solely on upper‑limb control, i.e. robotic arms without leg mobility) has already seen more adoption in closed, SOP‑driven settings such as manufacturing, senior care, and food service.

Costs

Today’s humanoid robots cost USD100,000 to USD200,000 each, far too high for economical labor replacement. Reducing those costs to USD20,000 to USD30,000 is critical, which explains why these firms need massive rounds of funding to underwrite R&D, manufacturing, and supply‑chain subsidies. 

Only once production scales up can unit costs meaningfully decline. While Tesla’s Optimus is targeting a cost in this range, China nevertheless has an advantage here. Just as China’s planned economy allowed for the seemingly sudden emergence of DeepSeek, the country’s ability to direct industrial policy and capacity towards a centrally defined goal also confers advantages when it comes to industrial production and scale. So, while Western companies like Boston Dynamics, Figure AI, and Apptronik currently lead in motion control and design, they struggle with commercialisation and high costs. In contrast, China’s integrated supply chain allows for rapid production at lower costs, with key components like harmonic reducers and servo motors priced at least 50% lower than international alternatives. Indeed, it seems likely that Tesla may eventually adopt Chinese suppliers to cut costs, in line with the company’s EV policy. This approach could however now find itself heavily disrupted by US tariffs on Chinese goods, making life more expensive for western companies but also eroding potential markets for Chinese made robots. 

Supply chain fragmentation

In response to these challenges and to keep pace with rapidly evolving technology, the supply chains feeding the humanoid robot market are growing more complex.

Across a broad range of supply chain businesses ranging from startups to legacy automakers and internet giants,  a range of strategies are emerging to meet the challenges of the humanoid robot market. 

Tier‑1 “body” firms like Zhiyuan and Unitree Robotics are developing their own motors and integrated joint modules while partnering with upstream suppliers on reducers, encoders, and other key technologies under strategic agreements, cementing their leadership. By contrast, many Tier‑2/3 startups (e.g. Accel Evo, Zhujing Power) prefer turnkey joint modules or dexterous hands, customizing only micro‑servo motors and embracing an asset‑light model.

Traditional manufacturers such as BYD and Seres are exploring flexible approaches at the nexus of autonomous driving and humanoids. With embodied‑intelligence models evolving, experts foresee a convergence between autonomous vehicles and humanoid robots sharing the same algorithmic framework, as their architectures gradually merge.

Viewed by value tier, the highest segment is compute platforms and algorithm models. This upstream layer remains early‑stage but boasts the greatest growth potential and technical barriers. Model architectures continue to evolve: Vision–Language–Action models add action‑sequence encoding to vision and language, enabling robots to translate “recognize a cup” into “grasp and pour” commands. Interpretable frameworks like DeepSeek trace decision paths to boost reliability.

Pioneers such as Figure AI and Tesla are showing clear commercialization paths: Figure by integrating large models for positive cash flow, Tesla by leveraging EV mass production expertise.

Most robot firms today use automotive‑grade industrial PCs (e.g. from Dawon, Chengzhi), which lack optimization in structure, cooling, and power for humanoid needs. As the market matures, purpose‑built control architectures will replace generic PCs—or some companies may skip PCs altogether to design their own chips and break dependency on NVIDIA GPUs.

On the edge‑chip front, firms like Runxin Micro and Allwinner focus on edge‑AI chips to reduce costs and boost real‑time responsiveness.

The mid‑value tier centers on modular structural parts, perception systems, and technically challenging drive modules. Although less exalted than compute, modular trends create structural opportunities. For example, hollow‑cup motors plus screw and tendon‑cable drives now enable lighter, more compact dexterous hands. Companies like Cishi Robotics, Lingxin Smart Hand, and Shanghai Aoyi have already commercialized these, shipping nearly 2,000 units in a year.

Electronic skin, extending tactile sensing from hands and feet to torso and arms, is also ramping up. Core applications include touch perception and safety, with piezoresistive, capacitive‑resistive, magnetoelectric, and visuotactile sensors leading the way. Experts estimate 2025 Chinese humanoid shipments of 5,000–10,000 units, driving demand for 100,000–200,000 tactile sensors.

In drivetrain systems, roller screws, especially reverse‑planetary roller screws, show promise for high‑load ankle and upper‑limb joints, offering self‑locking, heavy‑load, and long‑life traits. Although complexity and cost limit near‑term volume delivery, they promise vast long‑term upside. Companies like Xinjian Transmission and Wuzhou Xinchun are already localizing high‑end screws and could become key partners in robot drivetrains.

Integrated joint modules have become a hotbed for R&D and investment—e.g. unlisted Inkes Technology shipped over 10,000 units in 2024, demonstrating mature tech and supply chains. Frameless torque motors and hollow‑cup motors also have clear niches: the former powers lower‑limb and main‑drive joints (led by Wolong Electric Drive and Boke Co.), the latter dominates dexterous hands (e.g. Mingzhi Electric).

In perception, companies like Orbbec are rolling out 3D‑vision solutions, especially camera‑plus‑low‑end‑chip integrations, standardizing “scene understanding” for humanoids.

At the lowest value tier, standard hardware components, commodity competition reigns, yet upgrade opportunities lurk. On the materials front, high‑performance plastics such as PEEK are gaining traction. Experts note that PEEK producers like Zhongyan Co. and core‑raw‑material supplier Newhan New Materials are laying groundwork to challenge metal‑dominated lightweighting in select applications.

Conclusion

As we can see then, the infrastructure needed to deliver humanoid robots at scale is beginning to emerge. While high costs and supply chain challenges mean the widespread commercial deployment of humanoid robots is still likely 3 to 5 years away, momentum is increasing as the US and China vie for dominance.  

While Western firms lead in AI and embodied intelligence, China’s rapid iteration and cost efficiency may give it a long-term edge. Tesla’s 2025 plan to produce 10,000 humanoid robots—initially for internal use—may signal a shift towards more cost-effective Chinese suppliers. However, the industry’s success will depend on continued AI advancements, component cost reductions, and evolving market demand.