Produced by | RoboIsland

Recently, DeepSeek was reported to be leading the seed round investment in Winverse AI, a company founded just a few months ago. Its founder, Hu Han, previously led the development of Hunyuan multimodal and embodied foundation models at Tencent.

Prior to this, in August, DeepSeek participated in a strategic placement of Unitree Technology with an investment of 141 million yuan and a lock-up period of 36 months. Both parties signed a bidirectional preferential procurement memorandum.

Sam Altman, CEO of OpenAI, disclosed on a podcast that OpenAI is venturing into humanoid robot R&D, focusing on developing intelligent robot systems. He predicted that OpenAI could unveil an impressive demonstration prototype by 2027.

Other large model companies like Zhipu and Kimi are also venturing into embodied AI. In July, Zhipu invested in the general embodied AI company Yuanli Lingji and jointly open-sourced the 2.6 billion-parameter VLA model ZR-0 along with a matching (supporting) robot dataset. Its fund has also invested in multiple robot ontology (robotics) companies. Kimi is recruiting talent in the field of embodied AI.

The most straightforward interpretation is that 'large model companies are collectively entering the field of embodied AI.' However, DeepSeek has chosen to invest without touching the robotics ontology (hardware) or establishing production lines. In contrast, OpenAI has shifted from investing in others to in-house development, building actuators, establishing its own data collection systems, and managing its supply chain.

Beyond whether they enter the market or not, SenseTime's data provides another perspective, perhaps sparking more thought and imagination.

SenseTime integrated GPT-6 Astra into the RoboDojo testing platform to run a series of robotic operation tasks. The results were interesting: nine tasks that had previously failed all five attempts succeeded in all five tries when the model was allowed more attempts, pushing the success rate to 81.9%. This indicates a step forward in capability.

However, the test also recorded another set of numbers: completing these tasks took approximately 131 minutes and cost $19 for stacking bowls, and about 263 minutes and $55 for placing pens in a pen holder. In real-world scenarios, these tasks would typically be completed by a worker in just a few minutes.

The cost of a single robotic execution is several times higher than manual labor, and the time required is tens of times longer. This means that even if the model achieves an 81.9% success rate in a testing environment, the time and financial costs of completing a task in a real-world scenario remain prohibitively high for most factories or warehouses.

Thus, the bottleneck in physical interaction lies not in whether capabilities have been breakthrough (broken through) but in the price and speed at which they can be achieved.

The significant gap in cost structure explains why DeepSeek and OpenAI, seemingly both entering the field of embodied AI, are moving in opposite directions from the starting point.

RoboIsland attempts to analyze the following questions in this article: Why have DeepSeek and OpenAI, facing the same threshold, moved in opposite directions? What has pushed them to their current positions? In the market, who is bearing the fixed costs of physical interaction, and who is avoiding them? How long can these different positions be maintained when cost curves change?

I. Two Opposite Choices

First, consider DeepSeek's two investments in embodied AI, neither of which targets ownership of robotics hardware or real-world data.

The 141 million yuan strategic placement in Unitree centers on bidirectional preferential procurement: Unitree prioritizes DeepSeek for model training services, while DeepSeek prioritizes Unitree when exploring embodied business opportunities. Winverse AI, which DeepSeek plans to lead in investment, is widely regarded by industry insiders as 'more focused on underlying infrastructure and model acceleration' rather than a complete robotics company.

DeepSeek has secured access to 'bodies' and interfaces for 'brain' acceleration through capital but has not paid fixed costs for any production lines or data collection facilities.

OpenAI has taken a different path. As of October 2026, 27 positions related to 'robotics' can be found on its official website, up from 11 in May of the same year. New roles span actuators, motor design, manufacturing, supply chain, and data collection.

Among them, the actuator positions explicitly require designing custom robotic actuators, covering motors, transmissions, sensing, and thermal management. On the data collection side, they are hiring 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. From the job postings, it appears that OpenAI is building bodies designed for AI control while establishing a factory to produce data for those bodies.

The reason lies in DeepSeek's advantage in language models, which lies in algorithmic efficiency and inference cost control. However, this advantage cannot be directly translated into the physical world because the bottleneck in physical interaction costs is not model parameters but the fixed investment in real-world data collection and training facilities.

According to industry estimates, the cost of collecting one hour of effective real-world data ranges from 500 to 1,000 yuan. A training facility spanning thousands of square meters requires an investment of tens of millions to over 100 million yuan. Fixed costs such as equipment, labor, scenario construction, and maintenance do not automatically decrease with improved model efficiency.

Source: Unitree Technology

In past public reports, DeepSeek has relatively limited experience in hardware engineering, supply chain management, and real-world data collection. It has not directly demonstrated training results in physical interaction. While developing its own complete robotics is not impossible, the cost of crossing this fixed cost threshold likely exceeds the reasonable range of its current investment structure.

Therefore, DeepSeek's investments in Unitree and Winverse represent flexible expenditures like investments and procurement agreements, substituting for the fixed investments in production lines and data factories. Unitree provides bodies and real-world scenarios, Winverse provides interfaces for model acceleration, and DeepSeek only needs to contribute algorithms and computing power.

The reason for investing in Winverse is not complicated. During his time at Microsoft Research Asia, Hu Han co-authored papers such as SimMIM and Swin Transformer V2 with Xie Zhenda, who later joined DeepSeek and was one of the key contributors to DeepSeek-Coder, V2, V3, R1, and other core works, as well as the first author of the mHC paper.

At the same time, Winverse's focus on underlying infrastructure and model acceleration precisely fills the gap where DeepSeek lacks the most in the physical world. DeepSeek has invested in a team whose technical judgment has already been validated by its core researchers.

OpenAI's heavy asset investment is also driven by cost structure, albeit in a different direction. Earlier, OpenAI followed a light asset approach, investing in Figure and signing an agreement to develop next-generation AI models for Figure's humanoid robots. The collaboration ended in February 2025, with Figure shifting to in-house development of Helix. Adcock publicly stated that 'robotic AI must be vertically integrated and cannot be outsourced.'

However, the more fundamental lesson came from data: if the body belongs to someone else, the model cannot access the perception and operation data generated by the robot in real-world environments. When OpenAI disbanded its robotics team years ago, the reason given was a lack of training data.

After Figure closed the door, OpenAI's position became narrow. Continuing as a supplier of just the 'brain' meant that data access remained in the hands of the body's owner. To gain ownership of physical data, OpenAI needed to take matters into its own hands.

Therefore, understanding the strategic differences between the two companies requires looking at the fixed cost threshold of physical interaction. DeepSeek has chosen not to bear the fixed costs of building its own bodies, opting for relational control instead. OpenAI, cut off from data access through relational control, has chosen to bear fixed costs and shift toward asset control.

While their specific implementation directions are opposite, the constraints they face are of the same origin: the fixed cost threshold of physical interaction and the extent to which each can bear it determine their respective positions.

II. Who Bears the Costs of Physical Interaction?

Across the market, financing in the domestic embodied AI sector reached approximately 43.8 billion yuan in the first half of 2026. Among this, companies specializing in brain-layer algorithms and control software absorbed 22.253 billion yuan, accounting for 50.8%, while complete robotics hardware companies accounted for only 12.8%.

Although capital is flowing toward the 'brain,' the fixed costs of physical interaction have not disappeared due to this capital flow. Players with different capability structures have chosen their positions based on the cost threshold.

Based on who bears the fixed costs of physical interaction and to what extent, four types of market positions exist:

Type 1: Minimal Bearing of Fixed Costs for Physical Interaction, Shallow Positioning in Models and Interfaces

Alibaba released the Qwen-Robot series in June, including VLA operation models, VLN mobility models, and world models, attempting to create a universal foundation for model interfaces akin to 'hands, feet, and brains.'

Tencent launched the Tairos open platform, jointly developed by the Futian Laboratory and Tencent Robotics X. It does not produce hardware but focuses on the 'brain' and has been deployed in scenarios such as Dunhuang tour guiding with Unitree G1. Leading hardware companies like Unitree, Zhiyuan, Zhongqing, and UFactory have all integrated with it.

ByteDance entered the field by investing in Zibianliang Robot, the only embodied AI company to receive strategic investments from ByteDance, Meituan, Alibaba, and Xiaomi.

The common feature of these players is that they do not develop complete hardware, do not establish real-world data collection facilities, or other heavy assets. Instead, they use model capabilities and cloud platforms to secure their position as the 'brain' for robots. They are farthest from the fixed costs of physical interaction, with the lightest investments, and their substitutability is highly conditional, largely depending on whether hardware companies continue to need external models.

Source: Zibianliang

Type 2: No Complete Closed Loop but Sharing Partial Fixed Costs Through Data and Scenarios

JD.com does not develop hardware or general-purpose brains but focuses its resources on data collection. According to public reports, the EgoLive dataset plans to collect 10 million hours of human real-world video data over two years while establishing over 80 RoboBase industrial bases and robot maintenance networks. It has already built eight domestic maintenance centers and plans to cover over 100 countries within five years.

Meituan has invested in at least 16 embodied AI companies, with 10 becoming unicorns. It also develops its own delivery robot matrix, using real-world scenarios like food delivery and warehousing sorting to provide hardware validation for its invested companies.

The logic for these players is that since the bottleneck in physical interaction ultimately lies in data and scenarios, they should occupy the nodes of data production and scenario delivery without choosing between the model layer and hardware layer.

Type 3: Bearing All Fixed Costs of Physical Interaction and Accumulating Barriers Within a Closed Loop

Companies like Unitree, Zhiyuan, and UBTECH, which have the capability for mass production and delivery of hardware, fall into this category.

Unitree shipped 5,500 units in 2025 and aims to ship 10,000 to 20,000 units in 2026. Zhiyuan surpassed 10,000 units in cumulative production by March 2026 and has built the AGIBOT WORLD real-world dataset.

Their barriers lie in the production conditions for real-world data and the engineering accumulation of physical interaction, which general-purpose models cannot directly purchase with computing power.

Figure's decision to terminate its collaboration with OpenAI and shift to in-house development of Helix essentially reflects a choice to internalize the fixed costs of physical interaction. By doing so, they aim to achieve irrevocability once the closed loop scales, seeking high barriers.

Type 4: No Ownership of Any Closed Loop but Controlling the Underlying Infrastructure Through Which All Loops Must Pass

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

They do not directly bear the execution costs of physical interaction but make it difficult for physical closed loops to bypass their interfaces.

Large companies tend to focus on models, platforms, and data rather than directly developing robotics hardware, not because hardware is unimportant but because the fixed costs of physical interaction are not their strength.

Their approach itself acknowledges this cost threshold. The current influx of capital into Type 1 reflects a certain expectation: once interfaces are adopted by a sufficient number of hardware companies, the cost of replacement will rise. When switching costs increase, substitutability is no longer an issue.

Of course, this The premise (prerequisite) is that the cost curve of physical interaction declines fast enough, making model interfaces a more critical competitive dimension.

III. Advantages, Pressures, and Windows

When the cost curve of physical interaction changes, how long can the advantages of the four types of positions be maintained, and can they be converted into irreplaceable structural positions before the window closes? The criterion does not lie in how close they are to a physical closed loop.

The current advantage of Type 1, 'shallow positioning,' comes from the continued need of hardware companies for external model capabilities.

Alibaba's Qwen-Robot and Tencent's Tairos have been adopted by multiple leading hardware companies, indicating that using external models for thinking is still a common choice. However, this advantage is being weakened by a trend that has already emerged: leading hardware companies are internalizing 'brain' capabilities.

Unitree has open-sourced UnifoLM-WLA-1.0, while Zhiyuan has released AGILE 2.0 and GE-Act 2.0. Both are integrating model capabilities into their technology stacks.

The sustainability of Type 1's position depends on the race between the speed at which leading companies build their own brains and the pace of improvement in external model capabilities.

Image source: Zhiyuan

The current advantage of the second type, 'Data and Scenario Entry Points,' stems from the irreplaceability of real-world machine data.

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

However, this advantage is being partially alleviated by declining hardware costs. The labeling cost of synthetic data has dropped from over ten yuan per frame to just a few jiao or even a few fen. The price of a single tactile sensor has fallen from tens of thousands to 199 yuan, and the cost of dexterous hands continues to decline.

The fixed costs of data collection are decreasing, thereby partially diluting the scarcity of data entry points. It is important to distinguish that real-world machine data in high-precision industrial scenarios still requires on-site collection, and this scarcity will not disappear due to sensor price reductions.

The position of the second type is weakening in general-purpose scenario data but remains valid in high-precision industrial data, albeit with a narrowing scope of advantage that has not disappeared.

The temporary advantage of the third type, 'Inside the Physical Closed Loop,' comes from Unitree and Zhiyuan's shipment volumes and accumulation of real-world machine data.

Unitree's real-world machine data, Ontology iteration (Note: ' Ontology iteration ' is a specialized term that may not have a direct English equivalent; it refers to the iterative improvement of the robot's physical body or platform. For the sake of maintaining technical accuracy, it is left untranslated here, but could be paraphrased as 'platform iteration experience' if context requires.) experience, and delivery capabilities are assets that external model companies cannot reclaim, making them the most irrevocable.

However, it also carries the highest risk. According to some reports, the annual comprehensive operation and maintenance cost for a single wheeled-arm or bipedal humanoid robot is approximately 500,000 to 700,000 yuan, higher than the labor costs of most positions. In terms of order structure, educational and research institutions account for about 55.5%, while industrial and technology-oriented enterprises account for only about 21.1%.

Whether the closed loop can scale before fixed investments are exhausted depends on the speed of cost reductions and the speed of order structure transformation.

The extent of the temporary advantage of the fourth type, 'Underlying Infrastructure,' is built on the premise of whether the existing ecological gravity is strong enough.

The history of CUDA has proven that once an interface becomes the default path, the cost of removal is high. NVIDIA's GR00T/Jetson system is already aligned with the CUDA ecosystem, and Huawei's CloudRobo is a one-stop development platform, both possessing strong 'inevitability' attributes.

Alibaba Cloud and Tencent Cloud are closer to optional infrastructure. Their 'inevitability' depends on the penetration depth of cloud services in training and deployment, making them slightly less 'inevitable' than NVIDIA and Huawei.

If the cost curve of physical interaction declines rapidly and robot body manufacturers prefer to build their own 'brains,' the 'inevitable' status of underlying infrastructure will also be compressed.

IV. Conclusion

When considering the current industrial stage, whether it's powerful entities like DeepSeek and OpenAI or rapidly developing large model vendors, all face a common limiting condition: the cost of physical interaction remains too high to support large-scale commercialization.

Before the final outcome, the positions of large model vendors in the embodied AI track (Note: ' track ' is translated as 'arena' to convey the competitive aspect; alternatively, 'field' or 'domain' could also be used depending on context.) are even narrower than they themselves believe and more constrained by the pace of the physical world than any 'control' narrative can describe.

Cover image source: 'One Flew Over the Cuckoo's Nest'

Featured image source: Unitree Technology

Disclaimer: The content of this article is for reference only and does not constitute investment advice. Investors act at their own risk.

Copyright Zhixie Island. All rights reserved. Unauthorized reproduction is prohibited.