Huawei, Tencent, and Baidu have all joined the fray, kicking off an all-out ecological battle for robot brains

Just two days ago, the gong at the Hong Kong Stock Exchange rang for the "first stock of the robot brain". Seer Robotics went public, and its share price soared by more than 38% during the intraday trading on the first day. What's eye - catching are the data from the prospectus stage: the public offering part received nearly 6,000 times over - subscription, and the winning rate for one lot was only 5%. Eight cornerstone investors, including Hillhouse Capital and Yuanbao Family Office, invested a total of HK$462 million.

A company that has been established for only six years, with an annual revenue of 442 million yuan and still in a loss - making state, has received such enthusiastic pursuit in the capital market. Its core selling point is not the robot body, but the controller known as the "robot brain". According to data from CIC Consulting, in terms of sales volume, Seer Robotics' robot controllers have captured 24.8% of the global market and 45.2% of the Chinese market, both ranking first.

Seer Robotics' listing happens to coincide with the industry inflection point.

In the past two years, most of the attention in the embodied intelligence track has been focused on humanoid robot bodies. The focus of public opinion has always been on which company's robots can walk more steadily, move more flexibly, and look more human - like. However, in 2026, the trend has changed.

Capital has begun to bypass the "body" and directly flow towards the "brain". According to statistics from QbitAI, in the first half of 2026, the total domestic financing in the embodied intelligence track was about 43.8 billion yuan. More than half of it flowed to companies focusing on the "brain", while less than 20% went to body manufacturers.

This is a shift in the focus of industrial value during the process of embodied intelligence moving from concept verification to large - scale implementation. As the hardware body gradually matures and the supply chain becomes more stable, the core variable that determines what a robot can do and how well it can do it is shifting from the mechanical structure to the intelligent system.

01

Bypass the body and go straight for the brain

Back in 2024, the most popular companies in the embodied intelligence track were humanoid robot complete machine companies.

UBTECH went public on the Hong Kong Stock Exchange. Companies such as Zhipu AI, Unitree, and Fourier successively launched new products, and each product launch attracted the collective attention of the technology circle. The narrative logic at that time was straightforward: embodied intelligence is the next - generation computing platform, and humanoid robots are the ultimate form. Whoever can produce mass - marketable bodies first will gain an advantage.

However, in just over a year, the focus of capital has shifted.

The financing structure in the first half of 2026 is quite telling. Brain - focused companies received half of the funds, while full - stack, core component, and body - focused companies have moved to the periphery. From another perspective, companies related to the "brain" have received nearly 70% of the total financing, and pure hardware body companies have been squeezed to the edge.

The financing pace is also accelerating. Many brain - focused companies complete a round of financing on average every month. For the two fastest companies, there was only a two - week interval between two rounds of financing. This speed is not common in the hard - tech field. Usually, only the Internet and pure software tracks see such intensive capital injections.

The logic behind this is not complicated. After two years of explosive development, the technical threshold for robot bodies is rapidly decreasing. Unitree Technology announced that the price of its bipedal humanoid robot Unitree R1 has been reduced from 39,900 yuan to 29,900 yuan and is now available for immediate purchase.

Previously, humanoid robots were generally considered products in the hundreds of thousands of yuan range. Behind the price drop is the maturity of the supply chain and the start of large - scale production. The costs of core hardware such as joint modules, motors, and reducers are all rapidly decreasing, and the advantages of Chinese manufacturing are emerging.

When the "body" is no longer scarce, the value of the "brain" becomes prominent.

The reason is simple: for the same robot hardware, different brains can lead to vastly different tasks. A low - end controller can only make a robot repeat the same handling tasks along a pre - set route, while a high - end intelligent controller can enable a robot to sense environmental changes, plan paths autonomously, cooperate in multi - robot operations, and even understand natural language instructions through large models. The former sells hardware, while the latter sells capabilities.

The difference in gross profit margins best reflects the value stratification. Seer Robotics' prospectus shows that the gross profit margin of its controller business is as high as 79.8%, and that of its software business reaches 89.3%, while the gross profit margin of its robot complete machine is only 38.4%, and that of its accessory business is only 15.7%. A nearly 80% gross profit margin is very rare in the hardware - dominated robot industry and is closer to the profit level of pure software companies.

This is also the core reason why capital is chasing brain - focused companies. Compared with the hardware body, which has heavy assets, low gross profit, and a long cycle, the marginal cost of controllers and intelligent systems is lower, and the scale effect is stronger. Once an ecological barrier is established, the profit margin will be very considerable.

Of course, the connotation of the term "brain" in the current context is much richer than it was a few years ago.

Early robot controllers were essentially motion control boards, responsible for commanding motor rotation and coordinating joint movements. The technical barriers mainly lied in real - time performance and stability. However, today's "robot brain" has evolved into an intelligent system that integrates perception, decision - making, and control, integrating SLAM positioning and navigation, visual semantic recognition, reinforcement learning, multi - robot scheduling, and even starting to connect to large language models and world models.

Seer Robotics' prospectus shows that in the global robot controller market, the supply of controllers by independent controller suppliers has increased from 6,000 units in 2021 to nearly 50,000 units in 2025, and is expected to exceed 300,000 units by 2030. In terms of revenue, the global robot controller market size has grown from 700 million yuan in 2021 to 2.4 billion yuan in 2025, and is expected to reach 8.4 billion yuan by 2030, with a compound annual growth rate of 28.8% from 2026 to 2030.

Some professional industry institution reports also state that in 2025, the market for robot brain controllers reached 2.236 billion yuan, and the market for robot motion control systems reached 6.073 billion yuan. If we consider the new - generation intelligent controllers with AI support and the supporting software, algorithms, and cloud services, the market space is much larger.

A report released by UK - based Future Market Insights shows that the global physical AI market is expected to grow from approximately $383 billion in 2026 to $3.26 trillion in 2040.

Physical AI essentially means equipping various physical entities with intelligent brains. Robots are just one of the most typical forms.

02

Cloud providers enter the game

Compete to be robot brain suppliers

If the financing boom among startups is just a signal within the track, then the collective entry of Internet giants and cloud providers means that the war for the robot brain has escalated to the ecological level.

Since this year, Huawei, Tencent, Baidu, and Alibaba have successively launched their own embodied intelligence platform products, and without exception, they have all chosen the path of "not making the body, but making the brain". They do not directly manufacture robots but provide intelligent systems, development tools, and cloud service bases for robot manufacturers.

Huawei has taken the fastest and most thorough action. At the Huawei Cloud INSPIRE Conference this month, Huawei Cloud officially launched the CloudRobo Embodied Intelligence Development Platform, positioning it as the "world's first one - stop full - process embodied intelligence development platform". According to the official statement, this platform covers the entire chain of data synthesis, model development, simulation verification, and cloud - edge - end deployment. It has built - in millions of data assets and more than 20 Ascend - friendly models, enabling "robots to be connected to the cloud within hours and models to be deployed within minutes".

Huawei's thinking is clear: use cloud capabilities to lower the threshold of robot intelligence. In the past, robot manufacturers had to develop their own algorithms, train models, and build simulation environments, which was costly and time - consuming. Now, they can directly use Huawei Cloud's platform to complete the entire process from data to training and deployment in one stop. Robot manufacturers only need to focus on making good hardware bodies and implementing scenarios.

On the day of the product launch, more than 20 enterprises, including Youibot and Huayan Robotics, announced that they would be the first to join the CloudRobo platform. Huawei also simultaneously launched the ecological cooperation plan of "One Hundred Models, One Thousand Forms, Cloud - based Win - win". Combining Huawei's layout in Ascend chips, industrial software, 5G networks, etc., CloudRobo is essentially building an embodied intelligence infrastructure that coordinates the "end - edge - cloud".

Tencent has chosen a lighter approach. Tencent Robotics X Lab participated in the exhibition with a complete technology matrix for the first time, launched the Tairos Embodied Intelligence Open Platform, open - sourced the HY - Embodied series of the Hunyuan Embodied Large Model, and also demonstrated the robot body interconnection technology RoboFusion.

Zhu Yajuan, the person in charge of the Tairos product ecosystem at Tencent Robotics X, said on - site that Tencent positions itself as an "indispensable titanium screw" in the robot industry. It does not manufacture robot bodies but focuses on software and cloud services. This statement is quite interesting. It not only acknowledges its boundary of not making hardware but also emphasizes its core value at the system level.

Specifically, the Hunyuan Embodied Large Model solves the problem of "understanding and thinking" for robots, enabling robots to understand the environment, understand instructions, and plan tasks autonomously. The Tairos platform solves the problem of development efficiency by providing a standardized toolchain. RoboFusion solves the problem of interconnection between different robots. The combination of these three constitutes Tencent's solution for the robot brain.

Baidu has taken the path of "data + model + infrastructure". At the Create 2026 Baidu AI Developer Conference in May, Baidu Smart Cloud clearly stated that it would increase investment in three aspects: AI Infra infrastructure, scenario connection, and industry standard construction. In April before that, Baidu had already jointly launched the "Embodied Intelligence Data Supermarket" with several robot companies, building a hierarchical data labeling system.

In terms of investment, Baidu also has a deep layout. Baidu was involved in the 1 - billion - yuan Series B financing of Zhipingfang and the 700 - million - yuan Series A financing of the Beijing Humanoid Robot Innovation Center. Baidu's logic is to bind hardware companies with the capabilities of the Wenxin Large Model to complete the closed - loop of AI from the virtual world to the physical world.

Alibaba's DAMO Academy launched the RynnBrain Embodied Basic Model in February this year, open - sourcing seven models at once, including the industry's first 30B MoE - architecture embodied model. According to official data, this model has enabled robots to have spatio - temporal memory and spatial reasoning capabilities for the first time, setting a new SOTA in 16 embodied evaluation lists and surpassing Google's Gemini Robotics ER 1.5.

Although ByteDance has not separately launched an embodied intelligence platform, it has listed the world model as its top - priority goal for 2026. At the Volcengine FORCE Conference on June 23, ByteDance mentioned that the Seedance video generation model can be applied to the data synthesis stage of embodied intelligence. Considering ByteDance's technical accumulation in multi - modal large models and video generation, as well as Volcengine's cloud service capabilities, it is almost inevitable for ByteDance to enter the embodied intelligence brain track in the future.

The collective choice of large companies to "make the brain rather than the body" is essentially an economic decision.

Although the hardware body is intuitive, it has heavy assets, low gross profit, and complex supply - chain management. Moreover, it is likely to end up in a price war, which has been repeatedly verified in the smartphone industry. The brain and platform layers require large upfront R & D investments, but once an ecosystem is formed, the marginal cost will rapidly decrease, and continuous revenue can be generated through cloud services.

More importantly, cloud providers have natural advantages in making robot brains. The training of embodied intelligence requires massive computing power, large - scale simulation environments, and multi - modal large - model bases, which are the strengths of cloud providers. It is neither economical nor realistic for robot body manufacturers to build an AI infrastructure from scratch.

From this perspective, the future embodied intelligence industry is likely to form a hierarchical structure similar to that of smartphones: the bottom layer is chips and computing power, the middle layer is the operating system and intelligent brain, and the upper layer is various forms of robot hardware and scenario applications. What cloud providers are competing for is the most core system position in the middle layer.

03

High - margin brains

Why can't they support revenue?

Although the story of the brain is appealing, when it comes to the business reality, problems still exist.

Seer Robotics' prospectus is worth reading carefully. On the one hand, the 79.8% gross profit margin of the controller business and the 89.3% gross profit margin of the software business fully prove the commercial value of the robot brain. On the other hand, the revenue of the controller business in 2025 was only 85 million yuan, accounting for 19.3% of the total revenue, while the revenue of the low - margin robot complete machine was 300 million yuan, accounting for 67.9%.

In other words, the most profitable business has the smallest scale, while the business that supports the revenue has low profits. This is almost the common dilemma of all third - party controller manufacturers.

The reason is not complicated. At present, robot manufacturers, especially industrial robot manufacturers, prefer to purchase the whole - machine solution rather than buying controllers separately. On the one hand, the whole - machine delivery can be quickly implemented and directly used, and customers do not need to do their own integration. On the other hand, as a core component, if purchased separately, customers need to have strong secondary development capabilities, which is too high a threshold for many traditional manufacturing customers.

So, in reality, controller manufacturers often need to sell their products through the whole - machine supporting method. They first install the controllers into their own whole machines and then sell them to end - customers. The brain is good, but it has to rely on the body to be sold.

Xu Zhaoyun, a partner at Lihan Investment, once analyzed to the media that domestic humanoid robot complete machine manufacturers are generally in the stage of ramping up mass production. Their core requirement is to control the hardware BOM cost first. Third - party controller manufacturers can only bind customers by offering low - price and high - volume products, and the value of software is discounted due to the hardware carrier. This has led to a somewhat embarrassing situation: controllers have high technical barriers and high gross profit margins but cannot be sold separately; whole machines have relatively low technical barriers and low gross profit margins but are the main source of revenue.

Seer Robotics has not escaped this rule either.

From 2023 to 2025, the company's revenue increased from 249 million yuan to 442 million yuan, with a compound annual growth rate of 33.2% over three years. However, during the same period, the net losses were 47.7 million yuan, 42.3 million yuan, and 47.07 million yuan respectively, with a total loss of about 137 million yuan over three years. Although the adjusted net loss is narrowing, from 20.91 million yuan in 2023 to 2.87 million yuan in 2025, it is still far from making a real profit.

This is not a problem unique to Seer Robotics but a phased characteristic of the entire industry.

Currently, embodied intelligence is still in the early stage of implementation. Customers are more willing to pay for the "tangible" hardware, and their willingness to pay for software, algorithms, and intelligent systems has not been fully established. Just like in the early days of smartphones, consumers initially paid for the hardware. It was not until the mobile Internet ecosystem matured that the value of software and services gradually emerged.

Another challenge comes from the self - research of large companies. As cloud providers such as Huawei, Tencent, and Baidu have successively launched their own embodied intelligence platforms, will the survival space of third - party independent controller manufacturers be squeezed?

At present, their positions are still different. The platforms of cloud providers are more inclined to general large models, training tools, and cloud - based scheduling, while the controllers of companies like Seer Robotics are more