Embodied Intelligence, the game has completely changed.

An embodied intelligence data service firm established around a year and a half ago has received more data procurement orders in the past two weeks than the total volume of the previous two months, which acts like a shot in the arm against the backdrop of tightening financing environment across the entire embodied intelligence industry.

"Orders are not only coming from the embodied intelligence sector, some large foundational model manufacturers that previously only inquired about prices without placing orders have started to place orders now," the founder of this company told *Dingjiao One*.

The change started with the release of GPT-6 Astra. In early September, after OpenAI launched its latest generation model, a large number of tests of using Astra to control robotic arms to perform various tasks appeared on social media. Many embodied intelligence practitioners found that the general-purpose multimodal foundational model also has spatial understanding capabilities, which can directly control robots to execute tasks, with a success rate even exceeding that of some dedicated embodied models.

As the test videos spread more and more widely, voices like "General large models will sooner or later 'eat up' embodied models", "Large language models will swallow embodied models, and embodied intelligence is just a subset of general large models" began to circulate in the industry. Wang Xingxing, founder of Unitree Robotics, even said bluntly that after GPT-7 comes out next year, it may directly break through all the bottlenecks of embodied intelligence. However, there are also different voices in the industry. Some people believe that the current general large models are far from reaching the level of dimensionality reduction strike against embodied intelligence, and there are still obvious shortcomings in many links such as long-horizon task success rate, latency, and cost.

The answer may not be unique, but the impact Astra has brought to the industry is actually happening. Jin Ge, CEO of Lingyu Intelligence, told *Dingjiao One* that the release of Astra itself is exciting news for the embodied intelligence industry. It first verified a new feasible direction, and proved that data from the physical world and non-physical world complement each other. "Under the Scaling Law, training physical data into large language models can also improve their non-physical capabilities."

This endows embodied physical data with new value, and at the same time puts forward higher requirements for embodied "brain" companies. An embodied intelligence investor told *Dingjiao One* that after the release of Astra, the value of some embodied model startups and small and medium-sized enterprises will be questioned. "If a company does not have more than 1,000 computing power cards, it will be difficult for them to survive in the competition with large foundational model manufacturers in the future."

The emergence of Astra forces the industry to answer: Is embodied intelligence an independent business, or a branch on the capability tree of large language models that is destined to be covered?

01. Large manufacturers enter the market, is embodied intelligence "dimensionality reduced"?

"The barbarians at the gate are coming," an embodied intelligence practitioner told *Dingjiao One*, which is the first reaction of many people around him to GPT-6 Astra.

In his view, once AI giants with sufficient capital, computing power cards and talent reserves extend their capabilities to the embodied intelligence field, it is like a group of barbarians standing outside the door who may break in at any time. What is even more unsettling is that Astra makes many practitioners realize that the know-how accumulated in this industry may not be as deep as everyone thought.

The core change brought by Astra is a cross-domain capability improvement. General large language models can handle robot operation and navigation tasks that were previously mainly undertaken by dedicated physical models such as VLA and WAM, and also demonstrate capabilities of zero-shot generalization and in-context learning.

NVIDIA's research team tested Astra in mid-October. In the object grasping and placing test, without special fine-tuning with robot data, Astra can directly control the robot to perform tasks based on visual information and task instructions. The success rate of directly outputting actions relying on the Astra model reached 66.3%, which is far higher than the dedicated model π0.5 of the embodied intelligence company Physical Intelligence.

Domestic embodied intelligence company Galaxy Universal also conducted similar tests. In the selected RoboDojo tasks, the success rate of Astra directly controlling the robot is 26%, while the success rate of π0.5 is only 15.67%. If the two are combined, with Astra making decisions and π0.5 responsible for execution, the success rate will further increase to 48%.

The "brain" of a robot is usually divided into three layers: System 2 is responsible for understanding scenarios, planning tasks and reasoning, System 1 accepts instructions from System 2 and combines environmental feedback to plan and adjust actions, and finally System 0 is responsible for underlying motion control. From the above experimental results, it can be seen that the capabilities of large language models can already better perform System 2 and part of System 1, and the advantage interval of embodied proprietary models is compressed to part of System 1 and System 0.

Whether embodied intelligence can be implemented depends on the brain, which is a long-standing consensus in the industry, and the emergence of Astra has accelerated the development process of the brain. Wang Xingxing's judgment at the recent Unitree partner ecosystem conference is that some capabilities of GPT-6 can already solve part of the robot's planning problems. With the improvement of general large model capabilities, physical AI may usher in a major breakthrough in 1 to 2 years at the fastest, or 2 to 3 years at the slowest.

He represents the more radical view in the industry, and there is also a more conservative faction. Wang Qian, founder and CEO of Zili Robotics, believes that it is unlikely for foundational models to generalize from general intelligence in other fields to embodied intelligence. General large models cannot perform reasoning and deployment on robots at a reasonable speed, and a pure embodied model is still needed to do this.

From the perspective of embodied intelligence companies, Astra has compressed the development window and living space of some companies, and at the same time provided new development paths and ideas for other companies. From the perspective of large model companies, the opportunities brought by Astra far outweigh the threats.

A recent industry research report on the embodied intelligence sector from Nomura Securities shows that from September 2025 to August and September 2026, OpenAI procured about 10 million hours of embodied data, accounting for about 70% of the supply in the US market.

These data are the source of Astra's embodied "brain" capabilities. In addition to large-scale data procurement, OpenAI also plans to enter the embodied intelligence field in person. In June this year, a recruitment notice released by OpenAI CEO Sam Altman on social platforms officially announced its entry into the physical robot track. He said that the company is forming a new team called "OpenAI Robotics" and is publicly recruiting full-stack hardware, operation, system and machine learning engineers.

DeepSeek is also preparing for embodied intelligence-related businesses and investments. According to previous media reports, Liang Wenfeng has paid close attention to the embodied intelligence sector intensively, and communicated with many entrepreneurs in this field to discuss directions such as robots, interaction between large models and the real world. DeepSeek's solution to enter embodied intelligence may include setting up a team to conduct independent R&D, or cutting into the track by investing in related enterprises.

DeepSeek previously participated in Unitree's IPO placement. Last month, it was exposed that it agreed to lead the seed round financing of Winverse AI, an embodied model startup founded by Hu Han, a former scientist of Tencent, with an amount of tens of millions of US dollars.

In addition, cutting-edge general model companies such as Anthropic and Kimi are planning to enter the embodied intelligence and physical AI field, and have started relevant personnel recruitment. Embodied intelligence has evolved from the exclusive track of robot companies to the final destination on the roadmap of all AI companies.

02. Data collection "blowout orders", ontology rises sharply, model companies are affected

Facing the high-profile entry of AI foundational model giants with strong strength, companies in all links of the embodied intelligence industry chain are in different situations.

The first to benefit from the "Astra Effect" are the embodied data collectors at the most upstream of the industry chain.

The founder of the aforementioned embodied data service company told *Dingjiao One* that for a period of time before Astra was released and triggered discussions related to embodied models, many data companies were not having an easy time. Due to the narrowing of financing environment, data procurement orders also decreased, and some model companies could not get money from investors, so they correspondingly reduced data procurement.

"The Ego Centric data originally priced at 300 yuan per hour would even be pressed down to dozens of yuan," he said.

In fact, this does not mean that the industry's demand for data is decreasing. The general consensus in the embodied intelligence industry is that for embodied intelligence models to reach the capabilities of GPT-3.5 in large language models, about 100 million hours of data are required, while the amount of cleaned valid data in the industry is still at the level of millions of hours.

Low data efficiency, serious homogenization and involution competition among data collection companies have led to lower and lower data prices. However, the emergence of Astra has changed the structure of data demand to a certain extent. General large language model companies have seen the value of embodied data. Whether they do not plan to enter the embodied intelligence industry and only improve the overall intelligence level of the model through physical data training, or lay a data foundation for entering the embodied intelligence industry in the future, new demands have been generated.

Compared with data collection companies, model companies in the middle reaches of the industry chain need to face changes in the competitive environment.

An embodied intelligence investor told *Dingjiao One* that the launch of Astra by OpenAI has the most obvious impact on primary market companies that focus on embodied "brain" or scenario implementation. As the physical understanding capability of large language models gradually matures, the value of companies engaged in pure general embodied brain R&D will shrink.

Previously, one of the core logics that embodied "brain" companies could obtain high valuations was that investors believed that general and physical AI were two independent and parallel development paths, and a new OpenAI might emerge in the embodied model field. But if OpenAI itself can have certain embodied model capabilities, this narrative will be broken.

However, in today's embodied intelligence industry, there are almost no companies that only do a single link. Robot ontology companies will deploy data collection and model training related businesses, and data companies will also do model training. Therefore, the changes brought by Astra to company valuations and competition patterns have not emerged in a short period of time, and it is more about promoting the overall development of the industry at the technical level.

The downstream ontology companies benefit the most from technological progress, especially those robot ontology companies that already have mass production and cost advantages. They can quickly make up for their "brain" capabilities by accessing general large language models.

According to Wang Xingxing's disclosure, after the release of Astra, Unitree internally tried to use this model to carry out several robot experiments. For example, letting the robot draw only through prompts, or letting the robot play Go without training, Astra successfully completed the tasks.

In the last two weeks of October, Unitree's stock price rose by more than 70% cumulatively, and its market value once approached 300 billion yuan. An important reason behind this is that the emergence of Astra raised the market's expectation for the progress of embodied large models. When large language models begin to try to solve the problems that embodied large models need to solve, the value of original equipment manufacturers is re-anchored in motion control, ontology reliability, cost control and other aspects, and the "brain" has instead become a commodity that can be plugged and unplugged at any time.

Although the impact brought by Astra has affected all links of the industrial chain, it is still too early to say that Astra will completely subvert the embodied intelligence industry at this stage.

03. Under large models, what kind of embodied companies can survive well?

Zero-shot generalization, spatial understanding, deep reasoning, these obvious advantages are not enough to support Astra to swallow and replace embodied models, and its shortcomings are equally obvious.

The first is the latency problem. The popular videos of Astra operating robotic arms on social platforms are usually played at double speed. At normal speed, the "slowness" of general large models will be exposed.

When controlling robots, large language models often need to stop and think about what to do next. In NVIDIA's test report, when Astra directly outputs the end target on RoboLab-120, the average reasoning time per time reaches 18.56 seconds, and it still needs 13-18 seconds of thinking time in Agent mode. In contrast, the dedicated embodied model π0.5 only takes about tens of milliseconds.

If robots enter production scenarios or home service scenarios, the impact of such latency and pauses will be magnified exponentially.

In addition, large language models run in the cloud, which requires consideration of network latency and stability issues. However, there are different voices in the industry on this point.

One view is that robot operations have high requirements for real-time response. If the "brain" runs in the cloud and the execution layer runs locally, it will lead to unstable robot operation, especially when performing long-horizon tasks. If a sub-task fails to execute due to signal fluctuation, the whole process may get stuck.

But another view holds that the robot's brain should run in the cloud. Jin Ge from Lingyu Intelligence told *Dingjiao One* that the ceiling of end-side computing power is not high. Taking Thor, the most advanced end-side chip in the industry at present, as an example, it takes about 100+ milliseconds to calculate one frame of image, while NVIDIA's A100 chip from several years ago only takes more than 40 milliseconds to calculate one frame.

In addition, if the large language model is only responsible for the decision-making links of System 2 and part of System 1, it does not actually require particularly high immediacy.

In addition to the shortcomings in latency, cost is also an important factor that has to be taken into account. Also in NVIDIA's experiment, a successful direct output of actions consumes an average of 160,000 tokens, which is equivalent to about 1.96 US dollars; if Astra is responsible for planning and the embodied model executes the actions, it consumes an average of 11 million tokens per time, about 17.72 US dollars; and further, if Astra generates a successful action by adjusting the action according to the interaction history, it needs to consume 30.9 million tokens, about 36.40 US dollars.

This is only the cost of one action. From a commercial perspective, the cost performance is very low.

At present, the most ideal situation for the application of large language models in the embodied intelligence field is to "work" in layers with embodied models. In the planning layer and semantic layer, general models perform better and cheaper in capabilities such as task decomposition, scenario understanding, and instruction following. The execution layer of millisecond-level motion control, force and tactile feedback, ontology adaptation, and the most critical real machine data closed loop is handed over to the embodied model.

Under this new division of labor system, the value chain of the embodied intelligence industry will also be redistributed. Wang Qian from Zili Robotics once put forward a view that he believes that today's intelligent ontology exists in data, and the model is more of a distiller and a container. It is a distiller during training and a container during deployment.

This means that high-quality embodied data is one of the key factors that determine the unique competitive advantage of future large embodied models. In addition, data such as tactile sense, force control, and force feedback that large language models cannot obtain through calculation and reasoning will be more valuable.

More than a month after the release of Astra, the industry has voted with actions: data orders are back, the value of ontology companies is re-anchored, and companies that make "brains" have to face head-to-head competition with large foundational model manufacturers.

Some windows are being closed, while some new windows are being opened at the same time, and the pace of development of the entire industry is getting faster.

* All illustrations in the article are from Unsplash.

This article is from the WeChat official account <