2026.10.09 | TB | Robotics | China Robot Chips: Unitree Expert

Client-led • 3 Aug 2026

Companies: NVDA US, 9660 HK, 603893 CN, 300458 CN

Expert Executive, Operations at Unitree Robotics (Ningbo) Co Ltd

Agenda

• Around five years' experience in the robotics industry, focusing on market analysis, sales team building and marketing strategy promotion
• Deep insights into the application scenarios and development paths of large-compute and small-compute chips in the embodied AI robot industry
• Well-placed to discuss the gap between Chinese robot chips and Nvidia chips in compute speed and performance, as well as the development of Chinese chips' penetration rate

Questions

00:00:01] We would like to learn about the chip field of embodied AI and robots. What are the core differences between edge AI chips used in robots and those used in consumer electronics and automobiles?

Expert [00:00:01]: There are two main parts. One is 3588, which can be regarded as an on-device chip. However, it is a small compute and requires low latency and fast communication. There are also chips like Nvidia and Horizon Robotics that require high compute but high bandwidth. Compared with consumer electronics, they don’t have such high requirements for power consumption but have higher requirements for communication and compute capabilities.

Client 1 [00:01:14]: Which company made the first chip you mentioned?

Expert [00:01:14]: Like Rockchip and Allwinner Technology, the most classic one is Rockchip’s RK 3588.

Client 1 [00:01:30]: The robot’s on-device chip is mainly small compute and low latency, right?

Expert [00:01:30]: That’s right.

00:01:42] What are the major robot on-device chip suppliers in the market? What are their characteristics, advantages and disadvantages and technical moat?

Expert [00:01:42]: At present, there are only two chip manufacturers in this segment. One is Rockchip and the other is Allwinner Technology. Rockchip can basically occupy 95% of this market share and a small part is used by Allwinner Technology. The characteristics of each company are that Rockchip currently has a large overall supply and the entire chip development tool chain is relatively complete. Secondly, the price is relatively cheap. For example, 3588 is commonly used now and the price of a single one is USD 40, which is relatively cheap. The price of Allwinner is about 20-25% higher than that of Rockchip, which is about USD 50.

Client 1 [00:03:05]: Do Horizon Robotics or other companies produce chips for robots?

Expert [00:03:05]: This is mainly a large compute on-device chip. Currently, the three major suppliers are Horizon Robotics, Nvidia and Huixi. At present, Nvidia’s advantage lies in its product types. It has a wide range of products from low compute to high compute. Second, it has a complete ecosystem and tool chain. Many SDKs or ecological tools are developed based on Nvidia, which is friendly to developers at present. However, its disadvantage is that it is expensive. The price of a small compute chip ranges from USD 500-4,000 and there are some restrictions on importing some products.

The second type is Horizon Robotics, which mainly promotes RDK S100. It has a compute of 80 TOPS and is priced at about RMB 3,000. Its competitor should be Nvidia’s NX chip. Under the same compute level, it is about 35% cheaper than Nvidia. Its disadvantage is that it lacks an ecological chain at present, especially for some developer tools. There is also a company called Huixi, based in Beijing. It mainly produces large compute 500 TOPS chips. The disadvantage of Huixi Guangzhi R1 is that its ecosystem chain is incomplete. This chip was originally intended for autonomous driving. Since last year, it has been deployed on robots such as AgiBot and Leju Robot.

00:05:55] What kind of robots or products need to use a large compute chip? What kind of robot can use a small compute chip?

Expert [00:05:55]: Generally speaking, in most scenarios, small compute chips such as RK 3588 with 6 TOPS can basically be used for some simple motion control development and voice interaction. This is enough and similar to a small on-device compute box. Basically, humanoid robots or embodied AI within RMB 100,000 are equipped with a 3588. For secondary development users, basically all those with secondary development functions worth more than RMB 100,000 need to be equipped with compute chips of at least 100 TOPS. Nvidia’s products are widely used, mainly including NX and AGX. Large compute chips are popular this year, so Thor’s T5000 chip will be used more frequently this year. To put it simply, a small compute chip can be used for simple speech interaction and operation control; while a larger compute chip may be used for running some large AI models or multi-modal interaction.

Client 2 [00:08:06]: Are there any specific application cases of small compute chips in robots? What products on the market use small compute chips? In addition to the Vbot consumer-grade robot that uses a large compute chip, what else is there?

Expert [00:08:06]: For example, AgiBot’s Lingxi-AR X1 and Youth X2, Noetix Robotics’ Bumi chip and Booster Robotics’ K1 all use small compute chips. At present, there are many large compute chips, such as AgiBot’s Expeditionary A3 and Vbot. Most models priced above RMB 200,000 will be equipped with large compute chips, including EngineAI’s T800.

Client 2 [00:09:34]: Previously, there were some first-generation robots priced below RMB 100,000. They are not widely used or haven’t achieved the desired effect in terms of applicability. What do you think of the future or current trends? Will the large compute chip dominate future development?

Expert [00:09:34]: I think big compute will be a trend in the future. Starting this year, we can see that both industrial and consumer-grade embodied AI need to run larger large AI models with multimodal interaction as the main focus. Therefore, the demand for compute is expected to double this year. Take Vbot as an example. It is a typical consumer product that needs to be equipped with large compute chips because there is potential demand for consumer-grade embodied AI to carry such large compute chips, which will also be a great growth point in the future. For example, when embodied AI enters industrial scenarios, the demand for AI models and multimodal recognition, especially data processing of visual recognition, requires large compute chips at an exponential growth rate.

00:11:19] Industrial-grade robots won’t be widely used for large compute chips because the operations in factories or wharves are more complicated. However, industrial-grade robots have relatively fixed operation paths and movements in factories and warehouses. Is there a great demand for large compute chips? Which players are more engaged in industrial robots now?

Expert [00:11:19]: There are two parts of embodiment. Simple path planning is enough for a small compute. However, when we talk about embodied AI, it must be equipped with some large AI models. Otherwise, it is no different from mobile robots such as AGV and AMR. Therefore, there is a potential demand for the installation of large compute chips. For example, Horizon Robotics, Huixi and Nvidia are moving towards industrial-grade embodied compute with a TOPS of more than 500 this year. On the contrary, except for extremely few consumer-grade embodied products, there is still a demand for small computes within 20 TOPS.

Client 2 [00:12:44]: Consumer-grade robots include UAV and quadruped robots. How about the market for some simpler consumer scenarios before embodied robots? Is it saturated?

Expert [00:12:44]: At present, except for UAVs, quadruped robots are a bit saturated. According to our estimate, the sales volume of consumer-grade quadruped robots this year may decrease by 10% compared with last year. In terms of overall growth trend, it will slightly decline or remain at the same level as last year. At present, the demand is still small and has not yet reached such a stage of explosive growth.

00:13:49] Some reports say that the technical roadmap of embodied robots is not very clear yet. How to understand this?

Expert [00:13:49]: The path for industrial scenarios is indeed unclear. I think it will be difficult to realise this scenario in the next three years for three main reasons. First, the hardware is immature at present, especially the endurance mileage, payload, zero sales and some gears, which have not met the requirements of industrial hardware; Second, large AI models are indeed lacking in capabilities. At present, most industrial scenario demos are still running through RL. Relatively speaking, it is impossible to change clients or scenarios and run a large amount of data, which is unrealistic for long-term industrial deployment. Therefore, we have to wait for the development of the whole large AI model; The third part is that the industrial requirements for robots are relatively high, especially focusing on MTBF and trouble-free running time. At present, it is far from meeting this requirement. For an industrial product, it may run for tens of thousands or even 100,000 hours before one error is allowed. However, for embodied products, they may be down after running for hundreds of hours, which is absolutely unrealistic for industrial scenarios. In addition, it is difficult to calculate the ROI of embodied products at present. I think it will take several years to reduce costs.

Client 2 [00:15:50]: What about consumer-oriented ones? What are the main product focuses or R&D product types of leading technology companies?

Expert [00:15:50]: There will be more, such as humanoid robots and quadrupeds. Humanoid robots will be a hot topic this year. In addition to the earliest Booster Robotics, Vbot and Primebot will also enter this year. They have begun to plan for the consumer-grade embodied humanoid robot segment, which may be priced at RMB 10,000-30,000. There are fewer players in the consumer-grade quadruped market, such as Vbot and Azure. In this price range, players may compete more fiercely and focus on about 10,000. Of course, the functions of consumer-grade embodied AI are still limited. It is only an interaction with a large AI model based on voice interaction. The functions of consumer-grade embodied AI are indeed limited, so the current market demand is still limited. We may focus on less than 20,000 units per year.

00:17:28] Among embodied robots, humanoid robots and quadruped robots are mostly used in consumer scenarios. What is the approximate upstream chip supply of these robots?

Expert [00:17:28]: At present, the chip supply is sufficient. For example, RK 3588 is stable. However, for large compute chips, there is still a shortage of chips or price increases. Let’s take Nvidia as an example. This year, the price of on-device compute chips has increased by 25%. Therefore, the impact of prices is critical this year. We also hope to see Chinese transformation or import substitution and our goal is mainly around Horizon Robotics.

00:18:50] In terms of performance or practical application, how big is the gap between Chinese chips and Nvidia’s chips? What is the approximate penetration rate of Chinese chips in domestic robots?

Expert [00:18:50]: I think this can be divided into two parts. One is the closed ecosystem. Let’s take Vitasoy as an example. As we know, there are more than a dozen companies equipped with Horizon Robotics’ embodied AI. These applications launched by original manufacturers are very good. In terms of potential, this non-development version of embodied AI robots is expected to have a penetration rate of more than 60-70%. As for the developer version, whether it is Horizon Robotics or Huixi, its entire developer ecosystem or tool chain ecosystem penetration rate may continue to remain at a low level of 10-20%.

Client 2 [00:20:28]: How do we understand this development version and non-developmental version? For example, the non-development version has some functions set after leaving the factory and is directly installed on the robot. Therefore, some actions and instructions are fixed and cannot be changed again, right?

Expert [00:20:28]: Yes. It depends on whether the embodied AI product has some development interfaces. Generally speaking, the definition of non-development version and development version in this industry is based on this. Users are allowed to carry out secondary development of applications.

Client 2 [00:21:09]: Which form is more commonly used in China now?

Expert [00:21:09]: In terms of total volume, the development version accounts for about 75% of the entire embodied AI ecosystem.

Client 2 [00:21:27]: So the development version accounts for about 75%, so people will still be inclined to use this development version in the future, right?

Expert [00:21:27]: It depends on whether the functions are perfect from the perspective of manufacturers, especially in terms of consumer-grade embodiment. Consumer-grade embodied products are basically non-developmental versions because consumers don’t have the level of development. Therefore, the consumer-grade embodied market is still small at present, far smaller than that of scientific research, education and industry.

Client 2 [00:22:18]: In terms of consumer-grade applications, non-development versions will account for a relatively high proportion, right?

Expert [00:22:18]: Yes.

00:22:35] If AI is embodied and its functions are improved after a certain period of development, people will have higher demand for it at the application level in the future. Therefore, there will be continuous growth in the demand for non-developer versions. The developer version is in high demand now mainly because it is still in the R&D stage. Universities, research institutions and scientific research teams may need the developer version, right?

Expert [00:22:35]: That’s right. At present, I think embodied AI is still in the development stage and hasn’t been applied yet. I think this is a matter of phase transition.

Client 1 [00:23:15]: How long do you think it will take for such robots to be widely used? Is it 3-5 years or 10 years?

Expert [00:23:15]: I think it will take at least 5 years for the application to start. Depending on the scenario, I think it will take 3-5 years in the short term and about 10 years in the long run.

Client 1 [00:23:46]: What is the approximate proportion of chip cost in the cost of a complete vehicle product?

Expert [00:23:46]: The cost of a high-compute chip is about 10-15%, while that of a low-compute chip only accounts for about 1-2% of the total cost of a complete vehicle product.

00:24:11] Horizon Robotics has both high-compute and low-compute products. Which one has a larger market share? I just heard you say that Horizon Robotics takes the lead in high-compute Chinese chip manufacturing. What is its share in low-compute chip manufacturing?

Expert [00:24:11]: Rockchip focuses on the fusion of large and small brains. Its RDK S100 chip may be low in compute, but its embodied chip is medium in compute. About 80-100 TOPS, which can cover most application scenarios. Its high compute products are expected to rise to 500-600 TOPS.

00:25:21] Do these body manufacturers tend to develop by themselves or will they still choose to purchase?

Expert [00:25:21]: At present, some leading manufacturers had indeed shown signs of import substitution. However, I believed that in the short term, they would still rely on Nvidia within 2-3 years. However, some low-compute products gradually switched to Chinese compute systems, mainly from Horizon Robotics and Huixi Automation, which began to be applied in 2025. However, Nvidia is still the mainstream and its procurement volume will certainly increase further in the future.

Client 1 [00:26:20]: That is to say, these auto manufacturers will not choose to develop their own chips. They will either continue to use Nvidia or other suppliers’ chips.

Expert [00:26:20]: The difficulty is very great.

Client 1 [00:26:28]: In addition to compute, what indicators do they pay attention to when purchasing chips?

Expert [00:26:28]: At present, one is compute, the second is developers’ toolchains or ease of chip development and the third is price.