Guoxin Securities: Model iteration nurtures a new blue ocean in the data industry, which is expected to achieve 10 billion market space

The Zhitong Finance App learned that Guoxin Securities released a research report saying that data is a key resource for the development of large physical AI models and is the underlying driving force supporting model training, the emergence of capabilities, and intelligent evolution. Currently, the accumulation of physical data is mainly visual information. Human operation videos provide rich behavioral and semantic information, while tactile data requires simultaneous recording of contact states during interaction. The supply of existing tactile data is still relatively limited. There are various physical data collection routes, and the value of multi-source data training is outstanding. The embedded intelligent data industry chain covers infrastructure, data collection, data processing and application.

Guoxin Securities's main views are as follows:

High-quality physical interaction data is the key to achieving generalization with embodied intelligence. Embodied ScalingLaw was initially revealed

Throughout the development of large digital AI models, breakthroughs in model capability boundaries have always been accompanied by the collaborative evolution of data, computing power, and algorithms. Data is a key resource for the development of large physical AI models, and is the underlying driving force supporting model training, ability emergence, and intelligent evolution. Scaling Law in the field of embodied intelligence has begun to receive initial empirical support. For example, Dyna-2 discovered in an experiment with Ego's first-perspective human video pre-training data that when the data scale expanded from 1,000 hours to 1 million hours, the model's performance showed a marked improvement trend, providing preliminary evidence for Embodied Scaling Law, and it is expected that data demand will continue to expand.

General-purpose model manufacturers are speeding up entry and embodying intelligence, and tactile data is scarce

The growth rate of ARR in the digital AI business of some basic model manufacturers is slowing down, and the extension to the physical world is expected to become an important direction for them to expand their application boundaries, and demand for physical data procurement is growing rapidly. The multi-modal general model represented by GPT-6Astra continues to expand into the physical world, showing strong task understanding, planning, and closed-loop error correction capabilities in specific fields, further expanding the number of data requirements. Currently, the accumulation of physical data is mainly visual information. Human operation videos provide rich behavioral and semantic information, while tactile data requires simultaneous recording of contact states during interaction. The supply of existing tactile data is still relatively limited.

Specific data collection routes are diverse, and the value of multi-source data training is highlighted

The physical data collection technology route includes types such as Internet video data, Ego first-person video data, simulated synthetic data, and real data. Data from different sources have complementary values. 1) The scale of Internet video data is huge, and it lacks power and tactile information; 2) Ego first-person view data is performed naturally by the collector wearing a head-mounted or neck-mounted device, and does not rely on a specific robot body. It is a route with high scaling potential and low deployment threshold. The price of unlabeled Ego data is about tens of yuan/hour. The price of unlabeled Ego data is about tens of yuan/hour. The high-quality data after processing is a few hundred yuan/hour, and the Ego device is at the level of 1,000 yuan; 3) UMI general operating data interface is held by the operator As a collection tool, collect Terminal operation information provides structured motion information. The main advantage is that it is cross-body reuse and does not rely on specific robot body hardware, but there is a gap between the motion space and the actual robot. The data price is tens to hundreds of yuan/hour, and the core equipment is 10,000 yuan; 4) Remote operation data provides real robot execution demonstrations. It is a high-value data type, mainly used for post-model training. The price of tactile gloves is remarkable. The price of a single set of domestic manufacturers is 10,000 yuan; 5) The simulation data supports efficient generation and scene expansion, but faces core-SIM to- Real gap conundrum.

Domestic manufacturers are actively arranging data collection and processing, and physical data is expected to achieve 10 billion market space

The embedded intelligent data industry chain covers infrastructure, data collection, data processing and application, and enterprises with software and hardware collaboration, real scenario resources, and continuous data delivery capabilities are expected to form a differentiated advantage. In terms of market space, looking ahead to next year, industry demand will advance to 100 million yuan per hour, and the data market corresponds to a market space potential of about 10 billion yuan; in the long run, drawing on Tesla's intelligent driving data accumulation process, it is estimated that until 2024, when there is a clear breakthrough in the FSDV13 model capability in 2024, based on the assumption that vehicles will continue to be in use and bicycles will drive for 1 hour per day, Tesla's global fleet driving time is estimated to be about 6 billion hours (estimated value, for reference only). Theoretically, the data complexity of robots is far higher than that of intelligent driving cars. If they form a breakthrough in the ability of a major model similar to FSDV13, the data demand is likely to exceed this time period. Considering future declines in data collection costs, the long-term data market is also expected to exceed 100 billion yuan, and the market space is broad.

Risk warning: The progress of the artificial intelligence model fell short of expectations, the volume of physical robot shipments fell short of expectations, the demand for the procurement of physical data fell short of expectations, changes in data technology routes, falling data prices, and increased industry competition.