The "inflection point" of humanoid robots has arrived: the production volume is only the lower limit, and the model determines the upper ceiling.
Capital has confirmed a judgment with real money: in the second half of the AI era, pricing power is shifting from hardware to models. The same logic is accelerating its replication in the embodied intelligence industry — capital is re-pricing the value composition of robots with a similar valuation logic.
Many years later, when reviewing the evolution history of artificial intelligence, the release of DeepSeek R1 may be regarded as a watershed.
In January 2025, DeepSeek, a Chinese artificial intelligence company, released its reasoning model R1. Its performance is comparable to the most cutting-edge closed-source large models at that time, while its training cost is far lower than the industry level, and the model weights were open-sourced simultaneously on the release day.
A week later, NVIDIA's stock price fell by more than 17% in a single day, with its market value evaporating by more than 560 billion US dollars, falling below the 3 trillion US dollar mark. This is the day with the largest market value evaporation in NVIDIA's history, and also the highest record of single-day market value evaporation for a single company in the history of US stocks by then.
Before the release of R1, the mainstream expectation of the market was that the development of AI requires continuous and large-scale computing power investment, and NVIDIA almost monopolizes the supply side. R1 broke this expectation — the improvement of model capability does not necessarily require a proportional increase in computing power procurement. Computing power is no longer the only bottleneck, and the model itself is the core of valuation premium.
This impact was quickly transmitted to the capital market. In the months after the release of R1, the investment logic of the global large model track was re-examined. For the whole year of 2025, financing activities in China's AI sector remained active, the valuation of leading large model companies went through multiple rounds of revaluation within the year, and capital further concentrated on teams with core technical capabilities.
The track remained hot in 2026, and even witnessed a landmark event: DeepSeek received leading investment from the National Large Fund, with its primary market valuation anchored at about 45 billion US dollars; the overall valuation center of Chinese large model companies moved up on the whole.
Capital has confirmed a judgment with real money: in the second half of the AI era, pricing power is shifting from hardware to models.
The same logic is accelerating its replication in the embodied intelligence industry — capital is re-pricing the value composition of robots with a similar valuation logic.
Policy Clearance and Capital Inflow: "Brain" Becomes the New Valuation Anchor for Embodied Intelligence
Capital chases certainty. In the era of large language models, capital is the pathfinder; while for embodied intelligence, policies have taken the lead in making their attitudes clear.
The 2026 government work report clearly proposed to foster future industries including embodied intelligence, which marks the second consecutive year that embodied intelligence has been included in the government work report. The "15th Five-Year Plan" proposal also lists it as a future industry for forward-looking layout. The focus of policies was further clarified in the second half of the year: the Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission jointly launched a special action, listing "improving the large and small brain model algorithms" as a key research task; Beijing, Shanghai, Shenzhen and Hangzhou successively released action plans, all of which put the "large and small brain" models at the top of key technologies without exception.
The signal is clear enough, and the response of capital is more concentrated.
Data from IT Juzi shows that the total financing amount of China's domestic embodied intelligence track in the first half of 2026 reached 935 billion yuan, an increase of more than 5 times compared with the first half of 2025. For enterprises that have entered the listing process, the top fundraising projects in their prospectuses are mostly model and algorithm R&D. On July 24, the Hong Kong Stock Exchange further optimized the Chapter 18C listing system to lower the listing threshold for technology enterprises — under this set of rules, the valuation of enterprises is mainly based on technical capabilities, and the technical capabilities of embodied intelligence enterprises, in the final analysis, lie in the "brain".
Overseas, capital is also gathered in leading companies focusing on the "brain" level.
In July 2026, General Intuition, the U.S. "robot brain" company, completed a 320 million US dollar Series A financing with a valuation of 2.3 billion US dollars, whose core proposition is not to build robots but only to build "brains"; the UK-based Humanoid completed a 152 million US dollar Series A financing, becoming the first humanoid robot unicorn in Europe.
A more landmark case is Figure AI. The post-investment valuation of this company has reached 39 billion US dollars, with cumulative financing of 1.9 billion US dollars, and its investors include Microsoft, NVIDIA, OpenAI and Jeff Bezos. In March 2026, it obtained an additional financing of about 1.1 billion US dollars, targeting the general-purpose "robot brain". When a "brain" company can support a valuation of 39 billion US dollars, the market's pricing logic for "brains" speaks for itself.
The product side is also accelerating the shift to the brain. Tesla directly migrated the FSD autonomous driving neural network to Optimus, and the two share the same world model and visual perception system. At the 2026 Q2 performance meeting, Tesla confirmed that Optimus has entered mass production, with its focus shifted to the intelligent software layer. NVIDIA launched the Halos for Robotics security architecture, positioning itself as an "infrastructure provider for robot brains".
On the one hand, emerging brain startups are receiving excess financing; on the other hand, tech giants are shifting their decades-accumulated technical assets to the "brain". All paths lead to the same goal: the whole world is paying for the same narrative: the value of robots is shifting from the "body" to the "brain".
The Battle for Robot "Brain": Who is Leading?
To understand why the "brain" is important, we first need to understand how the models of embodied intelligence are layered.
The so-called robot "brain" we often talk about is more accurately called the "whole brain" — it has two key parts: the cerebellum is responsible for movement, such as balance, walking and jumping, to solve the problem of "how to move"; the cerebrum is responsible for planning and decision-making, such as understanding the environment and deciding what to do next. What is truly powerful is the "whole brain" (collectively referred to as "brain" in the following text), which connects the two parts, enabling perception, thinking and action to cooperate and complete together in one model.
At present, the global research on the embodied intelligence industry is mainly concentrated in two extremes.
One extreme is the hardware body. After the explosive development in the past two years, hardware bodies are no longer scarce, the technical threshold of robot bodies has dropped rapidly, and capital flows to the more upstream along the industrial logic. The representative enterprise is Unitree Robotics, which controls robots to complete specific motion tasks through specific programs, and has reached a very high level in the field of motion control.
The other extreme is the general brain. Most of the embodied large models intensively released by many companies at home and abroad in the past two years fall into this category, which solve the problems of task planning and scene understanding. However, at the level of general cerebellum and whole brain, few companies are carrying out systematic layout and thinking.
Overseas, Tesla's path is the most direct. On July 23, 2026, Tesla confirmed at its Q2 earnings call that Optimus is about to enter the mass production stage, and put forward the long-term goal of an annual output of 1 million units for the V3 version and 10 million units for the V4 version. The discussion about Optimus at the performance meeting rarely involved the hardware side, and focused more on how to achieve large-scale mass production, the hardware support required for the software layer, and the training of the VLA end-to-end model — which indicates that the hardware carrier has taken initial shape, and the focus is shifting to the model layer of intelligent capabilities.
Elon Musk made it clear that Optimus will adopt the same end-to-end strategy as FSD autonomous driving — that is, "input pixels, output control instructions". Tesla is building a single, huge and unified neural network, which receives sensory input through multiple cameras, receives proprioceptive signals from joints, receives force and tactile feedback from hands, and generates continuous control output. Different from the layered modular architecture adopted by many competitors — which separates perception, planning and prediction into independent code bases — Tesla abandoned these systems and turned to a single neural network.
In China, there are many enterprises making robot bodies, but few making "brains".
The R&D of the "brain" for embodied intelligence requires capabilities in computer vision, robotics, large-scale model training and real-scene engineering at the same time. These directions belong to different departments and disciplines in universities, and researchers who can integrate all of them are very few across the globe. In China, there are even fewer academic labs and established R&D teams focusing on this interdisciplinary field, which is the fundamental reason why there are far fewer startups in the "brain" direction than in the "body" direction.
The deeper bottleneck lies in data.
Wang He, founder of Galaxy General Robotics, shared a set of data in a public speech this year: a leading car manufacturer has about 100 million pieces of driving data flowing back every day, while the world's largest embodied intelligence data collection is only at the order of millions of pieces, with a gap of two orders of magnitude. The data of autonomous driving comes from real driving vehicles, with almost zero marginal cost; while the data of embodied intelligence comes from laboratories, each frame requires manual annotation or teleoperation collection, with high cost, low speed and limited scenarios. This means that embodied intelligence cannot copy the "real scene data flywheel" of autonomous driving, and must find a completely different solution on the data side.
The industry consensus is: to make effective use of multi-source heterogeneous data resources. Just as large language models achieved breakthroughs relying on massive Internet text data, embodied intelligence also needs to build a data system covering all physical interaction scenarios — Internet data, human motion data, simulation data, teleoperation data, and real machine backflow data. The five types of data sources must be integrated into a unified framework, complementing and verifying each other.
In China's exploration in the "brain" direction, the most representative enterprise is Galaxy General.
If Tesla represents the "top-down migration" — migrating from the end-to-end driving model accumulated by FSD to robots, then Galaxy General's path is closer to "bottom-up exploration" — taking the lead in building "whole brain" capabilities, and finally forming a full-stack closed loop of "data-model-body". This system integrates multi-source heterogeneous data on the data side, covers three levels of general cerebrum, general cerebellum and dexterous hand operation on the model side, and has completed verification in extremely dynamic scenarios such as playing tennis — all actions are independently decided by the model, without pre-programmed trajectories, and the cerebrum and cerebellum work cooperatively in the same framework. This achievement was reported by CCTV News, and Elon Musk also responded with attention on overseas social platforms.
The two paths have their own advantages. Tesla's end-to-end strategy has been verified in the field of autonomous driving, and its migration to robots is logically reasonable; while Galaxy General's "bottom-up" path relies on China's world's most complete humanoid robot supply chain and rich manufacturing application scenarios. This means that Chinese enterprises do not need to deploy the body until the "whole brain" is fully mature, but can train while iterating the body, forming a two-line parallel evolution rhythm.
What is more noteworthy is that China is building unique institutional advantages in the data dimension. Liu Liehong, Director of the National Data Administration, stated during his investigation at Galaxy General Robotics that providing high-quality data sets for the embodied intelligence industry will be taken as a new breakthrough to realize the marketization and value realization of data elements.
This is the first time that China's top-level data management institution has clearly taken embodied intelligence data as a policy focus — when data is defined as the fuel of the "brain", whoever masters the data system will master the iteration speed of the "brain". Under the background that there is no unified global data standard, China is using institutional strength to add momentum to the "brain" competition.
Following closely, organizational support has also been implemented simultaneously. Recently, Beijing Municipal Science and Technology Commission and Zhongguancun Administrative Committee officially recognized the "Beijing Key Laboratory of Embodied Intelligence Large Model", with Wang He as the director. This is a key measure for Beijing to lay out the embodied intelligence track. From the policy setting of the National Data Administration to the platform implementation of the Beijing Key Laboratory, the dual efforts of system and institution are building a systematic infrastructure from data to R&D for the "brain" competition.
Why Capital is Willing to Pay a Premium for the Embodied Intelligence "Brain"
Back to the issue of valuation.
DeepSeek R1 has proved a fact: in a technological competition, when hardware is no longer scarce, the party that masters model capabilities takes the largest premium. The embodied intelligence industry is replicating the same scenario — a growing industry consensus is taking shape: the "brain" is not an unnecessary luxury, but the only way for embodied intelligence to move from laboratories to real scenarios.
Lu Kunpeng, Audit Managing Partner of KPMG China's Technology Industry, pointed out the key at the technical level — the embodied large model represented by the world model can extract physical laws and achieve low-latency reasoning, which is the core base for solving the four major problems of data, task, control and deployment.
Zheng Nanning, an academician of the Chinese Academy of Engineering, gave the judgment criteria from the value dimension: to measure the real value of an embodied intelligence system, "the most important thing is not the demonstration demo on the screen, but whether it can enter the real physical environment and solve real problems in the real world".
The judgment from the academic side has been quickly verified on the capital side. Many investors pointed out in their reviews that the "embodied brain" is recognized as the most valuable investment target in the track. Gao Luning, Investment Director of Huagai Capital Technology Fund, said straightforwardly: "The 'brain' and data of robots are the real moat. A strong brain can even make up for mediocre hardware." Zhao Bowen, Chairman of Huangpu River Capital, observed the fundamental transformation of the primary market pricing logic — the valuation basis is shifting from "the background of the founding team and paper achievements" to "how much industry pricing power the core technical barriers can bring".
An intuitive comparison is: an embodied brain team with academic background only takes 8 to 10 months to go from a valuation of several hundred million yuan at the angel round to becoming a unicorn with a valuation of 1 billion US dollars; while "many robot companies that have completed commercial operations have been developing for ten years, and may not be able to reach a valuation of 100 billion yuan". This kind of time axis compression precisely reveals the capital market's scarcity pricing for the "brain" route.
The value of this judgment can be seen from the financing structure of leading enterprises. As of June 2026, 8 domestic embodied intelligence enterprises have been included in the "200 Billion Yuan Valuation Club", and the core narratives supporting these valuations are highly converged — all pointing to their respective general intelligent models. Galaxy General Robotics is one of the most recognized cases on this path. Its investors include "national team" funds such as the National Artificial Intelligence Industry Investment Fund, Sinopec, and CITIC Group — the National Large Fund's first layout in the embodied intelligence track selected exactly a representative enterprise of the "brain" route. From national industrial funds to local governments, the recognition of the "brain" route is expanding from commercial capital to strategic capital.
The output of machines determines the lower limit, while the capability of the "brain" will open the ceiling. After the explosive growth of hardware bodies in the past two years, the technical threshold has dropped rapidly. Whether the general intelligent model can continue to generalize in real scenarios is the key to determining the upper limit of enterprise value. This means that to judge the value of an embodied intelligence enterprise, the primary question is no longer "how many robots can be delivered this year", but "what stage its model capability has reached, and whether the data flywheel has started to rotate".
Once this set of pricing criteria is established, the subsequent actions of the capital market will be traceable. In the primary market, the valuation gap between "brain" companies and "body" companies will likely continue to widen. The window of the secondary market has also opened — with the further relaxation of the threshold of Chapter 18C of the Hong Kong Stock Exchange, a number of embodied intelligence enterprises with the "brain" as their core narrative are queuing up for listing. Referring to the valuation revaluation path that DeepSeek brought to the AI sector, after the leading "brain" companies complete listing pricing and form a public valuation anchor, the value coordinate system of the entire track will be recalibrated. The first to be revalued will inevitably be those scarce targets that master the model and data system.
History will not repeat itself simply, but the pricing logic of capital will. In the era of computing power, NVIDIA took the hardware premium; in the era of large models, DeepSeek proved the premium of models; in the era of embodied intelligence, the premium will belong to the "brain" companies that take the lead in running through the "data-model-body" closed loop.
The final judgment criteria for this competition is clear: it is not about who builds more robots, but about who builds a smarter "brain" — and who can make the capital market believe this first.