Original Title: Robots Haven't Started Selling Yet, but Embodied Data Volume is Already Surging
In the embodied AI market, data investment precedes shipments. Even though terminal sales are still limited, embodied AI companies have already begun investing resources in hundreds of thousands of hours, or even larger scales, of training data. Based on this, He Hongling, COO of Datatang, predicts that the most significant increment in the AI data service market in 2026 will be embodied AI. The changes are not only occurring in the scale of demand.
As large models enter the post-training phase, data services are shifting from large-scale operations emphasizing labor costs and scale to special forces teams relying on high-quality talent and rapid response to model feedback. Meanwhile, the data demand for embodied AI has formed a pyramid ranging from teleoperation and UMI to first-person perspective (Ego) data, with training value and scalability presenting different rankings. In He Hongling's view, large-scale data demand in any single AI field is cyclical.
What data service providers truly need to accumulate is the ability to migrate project implementation across cycles, flexible data production platforms, and copyrighted data assets. Centering around the changing demands of AI data services, the value and cost of embodied data, the commercial boundaries of simulation, and the future convergence of the domestic market, iFenxi communicated with He Hongling.
Core Insights
- Data precedes robot shipments, and embodied AI has become a significant increment in AI data services.
Robots must first be trained and possess sufficient intelligence before they can enter the market. Therefore, embodied data demand cannot be judged solely by terminal shipment volumes. He Hongling predicts that the main growth driver for players in the data industry in 2026 will come from embodied AI.
- Large model data services have entered the post-training and expert data phase, and the organizational model is shifting from large-scale operations to special forces teams.
Tasks require higher-quality, more creative talent with professional judgment. Team sizes are shrinking, but the value and cost of individual talent have significantly increased. The service process also requires more dynamic adjustments in response to model feedback.
- Data demand in a single track is cyclical, and data service providers must continuously shift to new data mines.
After a model reaches a high level in a specific field, large-scale data investment will gradually shift to testing and optimization for a small number of issues. For service providers to operate across cycles, the key lies in accumulating talent and project operation mechanisms, technology platforms, and copyrighted datasets.
- Embodied data is a pyramid: the closer to the robot's physical body, the higher the training value; the further from the body, the easier it is to scale.
Teleoperation data sits at the top of the pyramid, featuring high value, high cost, and small volume; first-person perspective data has lower costs and stronger universality, making it more suitable for scaled production.
- Simulation is an indispensable training method for embodied AI, but it may not independently become an external data service.
Models need to "struggle and learn" in simulated environments to complete reinforcement learning; however, there is currently no consensus on whether this process should be completed internally by the model company or outsourced to external service providers.
- Embodied AI is still in the stage of parallel attempts on multiple routes, and it may gradually converge over the next three to five years.
He Hongling holds an optimistic view on the long-term value of embodied AI, but also emphasizes that current data modalities and data recipes are not yet determined, and specific players and implementation forms still carry uncertainty.
The following is the transcript of this interview, slightly modified without changing the original meaning.
01. Data Precedes Robot Shipments, Embodied AI Becomes the Largest Increment
iFenxi: What substantive changes have occurred in the AI data service market over the past two or three years?
He Hongling: There have been tremendous changes in the past two years. From the business and market situations we have observed, the businesses of major participants are all growing, and the overall market is expanding. The growth comes from two aspects: one is the upgrade of existing data service demands, and the other is the new demands brought by large models and physical AI. For the upgrade of existing demands, voice is a very typical example. After the capabilities of large models improved, voice naturally became the interface for human interaction with large model applications, creating new demands for speech recognition and speech synthesis. Dialect data, which was once silent in the past, has regained attention. Manufacturers hope to make recognition more granular and want the system to reply in the corresponding dialect when a user asks a question in a specific dialect. The new demands mainly come from large models and physical AI. Large models include various directions such as voice, multimodality, understanding, and generation; embodied AI has also continuously seen new collection demands since last year. AI data services are inherently designed to serve model training, and models are one of the fastest-changing fields. Therefore, market demands will keep changing, and the pace of change will only accelerate.
iFenxi: Which new demands have already formed a considerable scale? Is embodied AI still in the exploration stage?
He Hongling: At least from our current business perspective, embodied AI can no longer be considered purely in the exploration stage. We could say that a few months ago, but currently, its proportion in the company's revenue is already quite considerable, even surpassing the autonomous driving data business. Our current main directions include voice, large models, general images, embodied AI, and autonomous driving. Among them, voice remains the largest field, while embodied AI is growing rapidly and has formed a certain business volume. My judgment is that the main growth driver for players in the AI data service market in 2026 will be embodied AI, not large models. The scale of data demand for large models is still very large, but from last year to this year, there has been no essential change in training methods and service models; embodied AI, on the other hand, was very small or even close to zero last year, and has rapidly reached a considerable scale this year, so the increment it brings is more obvious.
iFenxi: But from the terminal perspective, the shipment volume of embodied AI is still not large. Why has data service scaled up first?
He Hongling: Because we are discussing the AI data service market, which inherently precedes the shipment of embodied AI products. Robots must first complete training; only with data and intelligence can they be sold. We can make a conservative estimate. Suppose a model company needs 100,000 hours of data, and the cost per hour is calculated at CNY 100, the data investment would be CNY 10 million. Many other companies are planning for 1 million hours, or even setting targets of 10 million hours. From this, we can see how massive their investment in data is. The logic for robots and autonomous driving vehicles is not entirely the same. A car can be sold first, with certain intelligent driving features unlocked later; but no one would buy a robot that cannot do anything and just sits there. Robots must have sufficient intelligence before they can enter the market. Therefore, the scale of the data service market cannot be deduced solely based on the current shipment volume of robots.
iFenxi: Why hasn't the data demand for large models continued its explosive growth?
He Hongling: I think the training methods for large models have been basically established. Whether it is code, basic subject knowledge, vertical fields such as finance, law, and healthcare, or the aesthetic data needed for image and video generation and editing, there has been no paradigm-level change this year compared to last year. The investment directions of different companies will be adjusted according to product priorities. For example, some companies will increase investment in office applications, while others will strengthen their code capabilities. Budgets will still grow, and the requirements for personnel levels will also increase, but the overall model has not changed, and participants are further consolidating. Therefore, for large model data services this year, the scale continues to grow, rather than the demand pattern exploding anew.
02. Post-Training Enters the Expert Data Phase, Shifting from Large-Scale Operations to Special Forces Teams
iFenxi: Where are the current data demands for large model training mainly concentrated?
He Hongling: Currently, it is mainly the expert data required for post-training. General pre-training data has been heavily consumed, and the key investment directions roughly include code, basic knowledge of the physical world, vertical fields such as finance, law, and healthcare, as well as aesthetic data related to images and videos. These tasks all require higher-level talent. Often, only those in the top 10% of capability within a field are worth providing knowledge and judgment for the model. The higher the model's capability, the less it wants data filled with mediocre content; instead, it needs data that is creative, attractive, and truly insightful.
iFenxi: How will this change the organizational methods of data service companies?
He Hongling: In the past, when we did data services, we emphasized cost and large-scale delivery more. Taking autonomous driving as an example, a project might organize over ten thousand people, with relatively stable work content that could be done continuously for several months. This is more like large-scale military operations. Post-training for large models is more like special forces operations. A niche field might only need a dozen people; a team of a hundred is already quite large, but these people must be elites. The tasks are not fixed either, but need to be continuously adjusted according to the needs of the model companies. Therefore, the changes occur not only in recruitment but also affect organizational structure, project processes, and even the choice of team locations. Only by changing the organizational methods together can data service providers adapt to this business format. Embodied AI will also face similar issues.
iFenxi: Will personnel outsourcing become a new mainstream cooperation model?
He Hongling: Indeed, many projects now adopt the personnel outsourcing approach, but I do not think this is a structural change. In the past, settlement was based on data volume, with the premise that tasks could be clearly defined in advance and remain stable during execution. Now, many advanced tasks change with model performance, making it difficult to determine a unified workload and unit price at the very beginning. Therefore, it is easier for both supply and demand sides to reach an agreement by cooperating based on headcount and time first. If a task forms a stable process after a period of exploration, it may also revert to volume-based settlement. For example, providing services based on headcount for the first month or the first three months, and calculating based on output volume in the subsequent months. This is more of a business negotiation method, not a fundamental change in the business model of data services.
03. Single Tracks Have Cycles, Data Service Providers Must Continuously Shift to New Data Mines
iFenxi: If large model clients gradually consolidate and model capabilities continue to improve, where are the long-term opportunities for data services?
He Hongling: We have always been talking about AI data services, not just large model data services. I believe that large models will not always require external data services on the same scale. If an AI always needs people to repeatedly teach it, it also indicates that it hasn't truly learned. For example, when a model enters fields like finance, law, or healthcare, it might take half a year to a year to reach a level surpassing 90% of practitioners in that field. By then, large-scale investment will end, requiring only a small number of personnel to test and optimize for problematic cases. Next, the model might enter accounting, finance, or other fields, kicking off a new round of data investment. This is a bit like mining; you cannot expect a single mine to produce ore forever. Data service providers must continuously explore new fields, while ensuring that something can be accumulated during the process of shifting mines. Data demand in a single field will inevitably face cyclical changes, and embodied AI might only require large-scale data for a few years.
iFenxi: Crossing demand cycles, which capabilities can be continuously accumulated?
He Hongling: The capabilities we have accumulated over the long term mainly fall into three categories. The first category is standardized project implementation and delivery capabilities. This includes talent recruitment, management, and motivation, project manager experience, standardized project management processes, and the management of quality, schedule, cost, and client communication. These professional service capabilities can be transferred to new data tasks. The second category is the data engineering platform. In fields such as autonomous driving, embodied AI, and voice, clients tend to use our data platform. Based on the goals of improving efficiency and endlessly pursuing safety and controllability, the platform will be frequently and meticulously refined to continuously improve data production efficiency. The third category is copyrighted datasets. We can further accumulate project experience into data assets. The most leading model companies may have already invested a lot of money and time and temporarily no longer need our existing general datasets, but subsequent followers can shorten the gap with the advanced by purchasing copyrighted datasets.
iFenxi: Are copyrighted datasets more suitable for general training or vertical industry training?
He Hongling: Not necessarily only for vertical fields. Taking dialects as an example, top tech companies can invest massive resources to build data themselves, which is hard for other enterprises to accomplish at the same speed. If we have already accumulated copyrighted dialect datasets, latecomers can acquire capabilities faster by purchasing the data. Data demands in vertical industries will more often adopt customized approaches. After a general model enters specific businesses like power maintenance or power generation, it still needs industry data for post-training to understand the characteristics of that field. Building these datasets costs both money and time, but it is hard to bypass. Our own copyrighted datasets will focus more on general and pre-training-related data, while vertical fields will mainly rely on customized services.
04. Embodied Data is a Pyramid: Teleoperation at the Top, Ego Easier to Scale
iFenxi: How is the value of embodied data such as teleoperation, UMI, first-person perspective, and simulation ranked?
He Hongling: First, it needs to be clarified whether we are looking at the value from the perspective of the model trainer or the data service provider. For the model trainer, the closer the data is to the robot's physical body, the higher the value. The general order is teleoperation, UMI, first-person perspective (Ego), and then simulation methods. The further from the physical body, the more complex the mapping the model needs to learn, and the more information attenuation there will be. However, from the perspective of scaling for data service providers, the order is almost the opposite. First-person perspective data is easier to generalize and scale, followed by UMI, and then teleoperation. Costs also follow a similar order: first-person perspective is the lowest, UMI is in the middle, and teleoperation is the highest. Simulation is a very important method for data production and training, but whether it can independently become an external data service still has different views in the industry. Model training definitely requires a simulation environment, but whether this part of the work needs to be handed over to external service providers has not yet formed a consensus.
iFenxi: How is Datatang deploying these types of embodied data?
He Hongling: We are working on all three categories. For first-person perspective data, we jointly define high-quality, high-productivity collection kits with core partners. On the one hand, we use them to produce our own copyrighted datasets; on the other hand, we can also provide collection services around client needs. UMI is relatively close to specific equipment, and currently mainly focuses on customized client services. Teleoperation is also mainly customized services. We have set up dedicated teleoperation collection sites and simulated scenarios in Baoding, and have also accumulated a group of personnel with practical collection experience. From being unskilled in the early stages to now being able to smoothly solve problems during collection, this experience itself is an accumulation of capabilities.
iFenxi: Why is it said that teleoperation data sits at the top of the pyramid?
He Hongling: Teleoperation is "golden data." Its price is very high; a few hundred CNY for an hour of valid data is already considered cheap, and it could reach one or two thousand CNY. But its proportion in all embodied data might only be one-thousandth, or even one-ten-thousandth, so it must be at the top of the pyramid. UMI will have a larger data volume, but the production efficiency is still not high, requiring professional collection personnel to operate equipment like grippers. Further down is first-person perspective data, which allows people who are actually doing the task in real scenarios to collect it directly. For example, in a cleaning scenario, the collector should be someone who has been doing cleaning work for over ten years, not a fresh graduate who is only good at operating equipment. The experiences recorded by the two are different. I personally estimate that there might be an order-of-magnitude difference in data volume among first-person perspective, UMI, and teleoperation. Only by forming such a pyramid can the overall economics of embodied data be established.
iFenxi: Since teleoperation is hard to scale, why continue to invest in it?
He Hongling: Because many clients need complete solutions. The direct contribution of teleoperation to revenue may not be the most significant, but without this capability, it is impossible to cover the complete data chain. For a scaled data service provider like us, we will not simply give up any important data format. What really needs to be judged is which type of business bears the scale growth, and which type bears high-value capabilities and complete delivery. First-person perspective and UMI are closer to the focus of scaling, while teleoperation is an indispensable part of complete capabilities.
05. Simulation is Indispensable, but May Not Become an Independent Business
iFenxi: What role does simulation data play in autonomous driving and embodied AI?
He Hongling: Simulation is definitely very useful; it requires simulation data assets and simulation scenarios. In embodied AI, data can first be used to teach robots some basic capabilities, and then they are placed in a simulated environment to "struggle and learn," continuously correcting strategies through reinforcement learning. It is like a master first teaching a person how to do things, after which they must still experience the real society to truly grow. Autonomous driving follows a similar logic. The system needs to face scenarios such as "ghost probing" (sudden appearances from blind spots), children crossing the street, and falling objects from the vehicle ahead, to judge whether it can respond correctly. Without a scenario library, test case library, and simulation environment, it is difficult to systematically train and verify these capabilities. Therefore, simulation is an indispensable part of model training.
iFenxi: Can simulation data become an independent business?
He Hongling: Not necessarily. Necessary technical means do not equal that it will definitely form an independent external business model. Model companies inherently need reinforcement learning during the training process; they can build simulation environments themselves, or they might purchase externally provided environments or synthetic data. What I observe now is more of the former: simulation exists as part of the model training process. Whether it will be largely spun off in the future and provided by external service providers is currently uncertain. Therefore, while a clear judgment can be made about the technical value of simulation, boundaries need to be reserved regarding whether it can become an independent business.
06. Flexible Data Production Lines Must "Change Machine Tools" According to Demand
iFenxi: What is the core value brought to data services by the data engineering platform?
He Hongling: We call our platform a "flexible data manufacturing production line." It is first and foremost a production line, where data flow and business closed loops are completed within the system. How much data has been submitted, what problems have occurred, what the quality is, and how each individual and team performs can all be seen, thereby quickly identifying production issues. After the processes are moved online, the efficiency of data flow will also improve. More importantly, the platform not only manages data but also supports the closed loops of cash flow, information flow, and data flow for the entire enterprise, including settlement, bonuses, external payments, quality acceptance, and client delivery. Why each payment occurs, which production task it corresponds to, and whether the pricing and quality are reasonable should all be traceable. For a data service enterprise operating continuously, this system is similar to the ERP of a manufacturing enterprise. Small teams can temporarily rely on manual work and spreadsheets for management, but after the business scale expands, it is difficult to balance security, efficiency, quality, and compliance without a system.
iFenxi: Where is the flexibility reflected?
He Hongling: I often use the phrase "iron camp, flowing soldiers" to describe it. The production line has both automated and manual links; whether they can be replaced like plugins determines whether the platform can adapt to demand changes. For example, after the demand for autonomous driving data declines, the original platform cannot be directly abandoned. We hope to retain the experience from autonomous driving projects and, by replacing or adding some modules, shift the platform to data collection and annotation for embodied AI. Just like a manufacturing production line can process new products after changing machine tools, without needing to tear down and rebuild the entire production line. This requires the platform to be designed at a higher level from the very beginning. Today's R&D investment not only serves current projects but should also become an asset for entering new fields tomorrow. Therefore, the platform is one of the three core capabilities we have accumulated across demand cycles.
07. Embodied AI is Still in the Cambrian Period, and the Market Will Gradually Converge in Three to Five Years
iFenxi: Where are the differences between the domestic and international AI data service markets mainly reflected?
He Hongling: The capabilities and organizational methods of domestic manufacturers are not weak, and everyone has been researching and learning from overseas practices. The real differences come more from the market structure. There are many participants in the domestic AI data service market. After the same type of demand appears, a large number of companies often enter quickly, making the market relatively crowded. It is not easy for a single company to scale up its size and share. The overseas market relatively more often integrates resources, combining technical capabilities, talent scale, and client resources. There are also differences in clients' willingness to pay and their recognition of professional services. This is also the reason why domestic data service companies generally value the overseas market. However, under the current international environment, there are practical difficulties in entering some overseas markets. Therefore, we must both expand into external markets and wait for the domestic market to gradually move from redundant construction to rationality and consolidation.
iFenxi: In the next one to two years, what directions is Datatang most focused on?
He Hongling: Currently, it is undoubtedly embodied AI, or Physical AI. Embodied AI is still in its infancy, and specific technical methods and data forms are not clear, but precisely because of this, it has great room for imagination. Our most important task at present is to consider how to apply existing capabilities to this direction and continue to form new accumulations. The development of large models towards vertical fields is also a definite direction; it's just that the path is relatively clear, and the next step is more about entering different industries piece by piece. Embodied AI, on the other hand, still has many unanswered questions, and how far it will eventually evolve is even more worth looking forward to.
iFenxi: How to judge the duration of data demand for embodied AI?
He Hongling: I personally hold an optimistic view that embodied AI can change the human world. From transportation and mechanical power to computers and informatization, humans have been expanding their capabilities through technology, and embodied AI is logically sound in business. Of course, it is also in a boom now; some participants may become pioneers, and some may become martyrs, but this does not mean that infrastructure like computing power and data has no value. The fiber optics laid during the internet bubble and the routes left by the massive railway construction later supported new prosperity. Embodied AI may also not land at the time and in the form people expect, but the infrastructure will ultimately still play a role. As for data, a clear data recipe has not yet been formed: which data modalities are needed and how different data should be combined. Everyone is still in a trial stage similar to the "Cambrian Explosion." I think the next three years might be an important judgment point, and within three to five years, the industry will gradually find a clearer direction.
iFenxi: Will the domestic embodied data service market simultaneously move towards consolidation?
He Hongling: I think it will. Currently, many investments rely on capital promotion, with redundant construction and redundant investment, which is not entirely economically reasonable. When the market returns to rationality, enterprises still need to operate according to business laws, and industry concentration will also increase. It took about eight years for autonomous driving to develop from its early stages to relative convergence. The development speed of embodied AI might be faster, so my personal judgment is that in a minimum of three years and a maximum of five years, the market will show relatively obvious convergence. But uncertainty still needs to be reserved here: the direction is certain, but which players will eventually emerge, and with what technical routes and business models, still needs time to verify.