Why large model vendors cannot decide their own position in embodied intelligence

Recently, DeepSeek was reported to be planning to lead Winverse AI's seed round. The company was founded only a few months ago, and its founder, Hu Han, previously oversaw Hunyuan multimodal and embodied foundation models at Tencent.

Prior to that, in August DeepSeek had already participated in Unitree Robotics' strategic placement with 141 million yuan, with a 36-month lock-up period, and the two sides signed a two-way preferential procurement memorandum.

OpenAI CEO Sam Altman disclosed on a podcast that OpenAI is laying out a humanoid robot R&D track, and the related project will make robot intelligence system R&D a core focus. He predicted that OpenAI is expected to launch an impressive demo prototype in 2027.

Major model vendors such as Zhipu and Kimi are also laying out embodied intelligence. In July this year, Zhipu took a stake in general embodied intelligence company Yuanli Lingji and jointly open-sourced the 2.6-billion-parameter VLA model ZR-0 and a supporting robot dataset; its affiliated fund had previously invested in several robot body companies as well. Kimi, meanwhile, is recruiting talent in the embodied intelligence direction.

In this regard, the most convenient interpretation is that "major model vendors are collectively entering embodied intelligence." However, the approach DeepSeek chose is investment: it does not touch the body and does not build production lines; OpenAI has shifted from investing in others to in-house development, building its own actuators, establishing its own data collection system, and managing its own supply chain.

Beyond whether to enter the field, SenseTime's data offers another observational perspective that may bring more thinking and room for imagination.

SenseTime connected GPT-6 Astra to the RoboDojo testing platform and ran a set of robot manipulation tasks. The results are quite interesting: 9 tasks that had previously failed all 5 consecutive times became 5 out of 5 successful after the model was allowed more steps to try, pushing the success rate to 81.9%. Judging by the improvement in success rate, capability has indeed taken a step forward.

But the test also recorded another set of numbers: to complete these tasks, stacking bowls took about 131 minutes and 19 dollars, and putting pens into a pen holder took about 263 minutes and 55 dollars. In a real scenario, what it wants to replace is work that a worker could originally complete within a few minutes.

The cost of a single robot execution is several times that of manual labor, and the time required is dozens of times that of manual labor. This means that at the current stage, even if the model achieves an 81.9% success rate in the test environment, every time it completes a task in a real scenario, the time and money cost is still so high that almost no factory or warehouse is willing to pay.

So the bottleneck of physical interaction is not whether capability has broken through, but at what price and at what speed it can be done.

Thus, the order-of-magnitude gap in cost structure can also explain why DeepSeek and OpenAI appear to both be entering embodied intelligence but are moving in opposite directions at the starting point.

Zhixie Island attempts in this article to analyze the following questions: Why have DeepSeek and OpenAI moved in opposite directions when facing the same threshold? What has pushed each of them to where they are today? Across the entire market, who is paying for the fixed costs of physical interaction, and who is avoiding them? At these different positions, how long can each hold out when the cost curve changes?

1. Two opposite choices

First, look at DeepSeek. Neither of its two investments in embodied intelligence points to ownership of robot bodies or real-machine data.

The 141 million yuan strategic placement in Unitree has as its core term two-way preferential procurement: when Unitree purchases model training services, it gives priority to DeepSeek, and when DeepSeek explores embodied business, it gives priority to Unitree. As for Winverse AI, which DeepSeek plans to lead the investment in, multiple industry figures have made clear that it is "more oriented toward underlying infrastructure and model acceleration" and is not an embodied complete-machine company.

DeepSeek uses capital to lock in access rights to the "body" and an interface for accelerating the "brain," but it has not paid fixed costs for any production line or any data collection factory.

OpenAI has taken another path. As of October 2026, 27 positions can be found on its official website using "robotics" as the keyword; in May this year the number was still 11. The new positions span actuators, motor design, manufacturing, supply chain, and data collection.

Among them, actuator positions explicitly require designing custom robotic actuators, covering motors, transmission, sensing, and thermal management. On the data collection side, it is recruiting technical project managers who can combine robots, sensors, operator interfaces, and data pipelines into "production-ready" systems, tracking utilization, cycle time, throughput, and downtime.

This is almost a set of factory production metrics. Judging from the recruitment information, OpenAI is building a body designed for AI control, and at the same time wants to build a factory that produces data for the body.

The reason is that DeepSeek's advantage in language models is algorithmic efficiency and inference cost control, but this advantage cannot be directly converted in the physical world, because the bottleneck of physical interaction cost is not model parameters, but the collection of real-machine data and the fixed investment in training grounds.

According to public industry estimates, the collection cost of one effective hour of real-machine data is 500-1,000 yuan, and a training ground of several thousand square meters requires an investment of tens of millions to hundreds of millions of yuan. Fixed costs such as equipment, manpower, scenario construction, and maintenance will not be automatically diluted just because model efficiency improves.

Image source: Unitree Robotics

In past public reports, DeepSeek's accumulated experience in hardware engineering, supply chains, and real-machine data collection has been relatively limited, and it has never directly demonstrated training results in physical interaction. Building its own complete robot is not something it cannot do; rather, the cost of crossing this fixed-cost threshold would likely exceed what is reasonable within its current investment structure.

So DeepSeek's investments in Unitree and Winverse use reversible expenditures such as investment and procurement agreements to replace fixed investment in production lines and data factories. Unitree provides the body and real-machine scenarios, Winverse provides the interface for model acceleration, and DeepSeek itself only needs to supply algorithms and computing power.

As for why the investment was in Winverse, the reason is not complicated. During Hu Han's time at Microsoft Research Asia, he co-authored papers such as SimMIM and Swin Transformer V2 with Xie Zhenda. Xie Zhenda later joined DeepSeek and is one of the authors of a series of core works including DeepSeek-Coder, V2, V3, and R1, as well as the first author of the mHC paper.

At the same time, Winverse's positioning, which leans toward low-level infrastructure and model acceleration, happens to sit exactly at the interface DeepSeek most lacks in the physical world. DeepSeek invested in a team whose technical judgment has already been validated by its own core researchers.

OpenAI's heavy-asset investment was likewise driven out by its cost structure, only in a different direction. Earlier, OpenAI also took a light-asset route, investing in Figure and signing an agreement to develop the next generation of AI models for Figure's humanoid robots. The cooperation ended in February 2025, and Figure shifted to developing Helix in-house. The reason Adcock publicly gave was that "we must vertically integrate robot AI and cannot outsource AI."

But the more fundamental lesson came from data: if the body belongs to someone else, the model cannot obtain the perception and manipulation data generated by the robot in real environments. When OpenAI disbanded its robotics team years ago, the reason it gave was precisely the lack of training data.

After Figure closed the door, OpenAI's position became very narrow. If it continued as a supplier that only provides the "brain," access to data would always remain in the hands of the body owner; to obtain ownership of physical data, it would need to do it itself.

Therefore, to understand the strategic difference between the two companies, one must look at the threshold of the fixed cost of physical interaction. DeepSeek chose not to bear the fixed cost of building its own body and instead pursued relational control; OpenAI, having been cut off from data access within relational control, chose to bear the fixed cost and shift to asset control.

The specific directions the two are implementing are opposite, but the constraints they face share the same source: the threshold of the fixed cost of physical interaction, and the extent to which each can bear it, determines where each stops.

II. Who is bearing the cost of physical interaction?

Across the entire market, in the first half of 2026, financing in China's embodied intelligence sector totaled about 43.8 billion yuan, of which brain-layer algorithm and control software companies absorbed 22.253 billion yuan, accounting for 50.8%; complete-robot body companies accounted for only 12.8%.

Although capital is moving toward the "brain," the fixed cost of physical interaction has not disappeared because of capital flows. Players with different capability structures, when facing the cost threshold, have each chosen their own position.

According to who bears the fixed cost of physical interaction and to what extent, there are four types of positions in the market:

Type 1: Barely bearing the fixed cost of physical interaction, taking shallow positions in models and interfaces.

Alibaba released the Qwen-Robot series in June, including VLA manipulation models, VLN navigation models, and world models, attempting to make the model interface a general foundation for "hands, feet, and brain."

Tencent launched the Tairos open platform, jointly built by the Futian Laboratory and Tencent Robotics X. It does not produce bodies and focuses on the "brain." It has already connected Unitree's G1 for scenarios such as Dunhuang guided tours, and leading body manufacturers such as Unitree, AgiBot, EngineAI, and Dobot have all connected.

ByteDance entered through investment in X Square Robot, which is the only embodied intelligence company to simultaneously receive strategic investment from ByteDance, Meituan, Alibaba, and Xiaomi.

The common feature of such players is that they do not make complete robots or build heavy assets such as real-machine data collection. Instead, they use model capabilities and cloud platforms to occupy the position of "what robots use to think." They are farthest from the fixed cost of physical interaction and invest the least, so their substitutability is also the most conditional, depending largely on whether body manufacturers continue to need external models.

Image source: X Square Robot

Type 2: Not owning a complete closed loop, but sharing part of the fixed cost through data and scenarios as entry points.

JD.com does not make bodies or general brains, and instead directs resources toward data collection. According to public reports, the EgoLive dataset plans to collect 10 million hours of real-world human scenario video data over two years, while also laying out more than 80 RoboBase industrial bases and a robot repair network. It has built 8 major repair centers in China and plans to cover more than 100 countries within 5 years.

Meituan has cumulatively invested in at least 16 embodied intelligence companies, 10 of which have become unicorns, while also developing its own delivery robot matrix and using real scenarios such as food delivery fulfillment and warehouse sorting to provide real-machine validation conditions for invested companies.

The logic of such players is that the bottleneck of physical interaction ultimately lies in data and scenarios, so they should occupy the nodes of data production and scenario delivery, rather than choosing between the model layer and the body layer.

Type 3: Bearing the full fixed cost of physical interaction and accumulating barriers within the closed loop.

Companies with the capability to mass-produce and deliver robot bodies, such as Unitree, AgiBot, and UBTech, belong to this category.

Unitree shipped 5,500 units in 2025 and targets shipments of 10,000 to 20,000 units in 2026; AgiBot's cumulative output surpassed 10,000 units by March 2026, and it has built the AGIBOT WORLD real-world scenario dataset.

Their moat lies in the production conditions for real-machine data and the engineering accumulation of physical interaction—things that general-purpose models often cannot simply buy with computing power.

Figure's termination of its partnership with OpenAI and pivot to developing Helix in-house is essentially a choice to internalize the fixed costs of physical interaction in exchange for the irreversibility that comes once closed-loop scale takes off, in pursuit of a high barrier.

Type 4: Owning no closed loop, but controlling the underlying infrastructure that closed loops must pass through.

NVIDIA's GR00T/Jetson system, Huawei's CloudRobo platform, and the layouts of Alibaba Cloud and Tencent Cloud at the computing power, simulation, and training framework layers belong to this category.

They do not directly bear the execution costs of physical interaction, but they make it impossible for as many physical closed loops as possible to bypass their interfaces.

In direct layout, big tech companies prefer to build models, platforms, and data rather than personally entering the robot body business—not because the body is unimportant, but because the fixed costs of physical interaction are not a strength in their capability structure.

Their approach to layout is itself an acknowledgment of this cost threshold. Today, capital is pouring heavily into the first category, which also reflects a certain expectation: once an interface is adopted by enough body manufacturers, the cost of replacing it will rise, and once switching costs increase, substitutability is no longer a problem.

Of course, the premise is that the cost curve of physical interaction declines fast enough to make model interfaces a more important competitive dimension.

III. Advantages, Pressure, and the Window

When the cost curve of physical interaction changes, how long the advantages of each of the four positions can last, and whether they can be converted into irreplaceable structural positions before the window closes, is not judged by who is closer to the physical closed loop.

The current advantage of the first category, "shallow positioning," comes from the fact that body manufacturers still need external model capabilities.

Alibaba's Qwen-Robot and Tencent's Tairos have been adopted by multiple leading body manufacturers, showing that using external models for thinking remains a common choice today. But this advantage is being eroded by a trend that has already occurred: leading body manufacturers are internalizing "brain" capabilities.

Unitree has open-sourced UnifoLM-WLA-1.0, and AgiBot has released AGILE 2.0 and GE-Act 2.0 in succession; both are writing model capabilities into their own tech stacks.

Whether the first category's position can be maintained depends on the race between the speed at which leading manufacturers build their own brains and the speed at which external model capabilities advance.

Image source: AgiBot

The current advantage of the second category, "data and scenario entry points," comes from the fact that real-machine data remains irreplaceable.

JD.com's 10 million hours of first-person human perspective data and Meituan's delivery and warehousing scenarios are indeed harder to bypass than the pure model layer.

But this advantage is being partially alleviated by falling hardware costs: the annotation cost of synthetic data has been compressed from over ten yuan per frame to a few mao or even a few fen, the price of a single tactile sensor has dropped from tens of thousands to hundreds of thousands of yuan down to 199 yuan, and the cost of dexterous hands continues to fall.

The fixed cost of data collection is declining, so the scarcity of data entry points will be partially diluted. What needs to be distinguished is that real-machine data for high-precision industrial scenarios still requires on-site collection, and this scarcity will not disappear because of sensor price cuts.

The second category's position is thinning in general-scenario data but still holds in high-precision industrial data; the scope of its advantage is narrowing, but it has not disappeared.

The temporary advantage of the third category, "inside the physical closed loop," comes from Unitree's and AgiBot's shipment volumes and real-machine data accumulation.

Unitree's real-machine data, body iteration experience, and delivery capability are assets that external model companies cannot reclaim, and its irreversibility is the strongest.

But its risk is also the highest: according to some reports, the annual comprehensive operating and maintenance cost of a single wheeled-arm or bipedal humanoid robot is about 500,000 to 700,000 yuan, higher than the labor cost of most positions; in terms of order structure, educational and research institutions account for about 55.5%, while industrial and technology enterprises account for only about 21.1%.

Whether closed-loop scale can take off before fixed investment is exhausted depends on the speed of cost decline and the speed of order structure transformation.

How large the temporary advantage of the fourth category, "underlying infrastructure," is depends on the premise of whether the gravitational pull of the existing ecosystem is strong enough.

CUDA's history has already proven that once an interface becomes the default path, the cost of removal is high. NVIDIA's GR00T/Jetson system has existing alignment with the CUDA ecosystem, and Huawei's CloudRobo is a one-stop development platform; both have strong "must-pass" attributes.

Alibaba Cloud and Tencent Cloud are closer to optional infrastructure; their degree of "must-pass" depends on the depth of cloud services' penetration in training and deployment, and their strength is slightly weaker than that of NVIDIA and Huawei.

If the cost curve of physical interaction declines rapidly and body manufacturers are more inclined to build their own brains, the "must-pass" status of underlying infrastructure will likewise be compressed.

IV. Conclusion

Placing the current industry stage in context, whether strong players like DeepSeek and OpenAI or fast-growing large-model manufacturers, all face a common constraint: the cost of physical interaction remains too high to support large-scale commercialization.

Before the endgame, the position of large-model manufacturers in the embodied intelligence race is even narrower than they themselves believe, and more constrained by the rhythms of the physical world itself than any "control" narrative can describe.

English text is being prepared.