The Zhitong Finance App learned that Guojin Securities released a research report saying that the digital collection service is the link with the highest order visibility and the most stable cash flow in the entire data collection industry, and leading companies will continue to benefit from the widening supply and demand gap and the concentration of patterns brought about by route convergence. Focus on digital acquisition service providers with order, production capacity, and delivery capabilities in the short term. Medium- to long-term companies that can platformize data processing and extend it back to model evaluation, post-training, and deployment. Focus on three types of companies: (1) traditional leaders with traditional AI data service experience; (2) simulation leaders that use simulation assets as a base to build an integrated software and hardware platform; (3) a full-process platform that lays out a closed loop of “data - simulation - training - migration - scenario”.
Guojin Securities's main views are as follows:
The data bottleneck of physical AI is not only a quantitative gap, but also a gap between quality and closed loop
(1) Quantity: The industry expects at least 10 million hours of data required for mature and full-sized models, and it may take 100 million hours to further advance to general intelligence. Currently, the industry's high-quality training data is only hundreds of thousands to 1 million hours, and the gap between supply and demand is about 90%. (2) Quality: Physical AI needs to establish timing relationships between task instructions, environmental observation, robot state, control actions, and execution results. The data must include not only vision, language, but also information such as joint position, force, and touch, and meet the requirements of “accurate, complete, broad, and true”. (3) Interaction: Physical data collection requires collaboration between edge side collection and cloud processing to form a closed loop of “data acquisition - cleaning, alignment, labeling, processing - quality evaluation - delivery feedback”.
Judging from the picture of the industry, the digital acquisition industry is in the early stages of fighting over the technology path and the business model is still being explored
The divergence of routes between real machines, no body, and simulation is essentially a balance between precision and scale. Real machines have the highest accuracy, no body has the fastest scaling speed, and simulation has the broadest potential for development. The line anticipates that mixed data types will become the mainstream training method for physical models. Supply gaps and unsettled technology have spurred many players to enter the market. Judging from the commercial paradigm, the business model for digital acquisition equipment and services is the clearest, while data assets and SaaS platforms have more long-term potential.
The core competitive elements of digital acquisition services revolve around border adjustments and asset upgrades
(1) Increased base model capabilities impact and redefine the boundaries of physical data companies. One of the biggest differences in the market about embodied data companies is whether independent embodied data companies will be attacked by downsizing if the general model continues to improve. The bank believes that the general model ability circle is “understanding instructions, disassembling tasks, and calling skills,” and that precision task capabilities are still insufficient. The success rate of GPT in the precision operation group was only 4.0%, while the DM0.5 in the control group was 16.75%. Embodied data companies have rich physical information assets, and their core barriers are reflected in additional marginal gains in the company's data and services after switching to a stronger base model.
(2) Data companies' products are not videos but training assets. Rich physical information, complete data processing capabilities, and continuous optimization of training feedback are the core of building data assets. There is a general power law relationship between the diversity of training data and the generalization performance of strategies in an unseen environment. Taking the DROID project as an example, after adding multi-scene external data, the model's performance within the division and in unseen scenarios increased by an average of 22PP and 35PP, and the generalization ability was significantly improved.
(3) In the long run, the long-term moat for digital acquisition service companies comes from entering the scene and evaluating closed loop and reuse capabilities. The most valuable companies in the medium to long term are often those that master the deployment network or deeply embed the deployment process. The data source continues to be generated as the business runs, making it difficult for competitors to replicate through one-time capital expenses. Having a scene doesn't necessarily mean having data value. It is also necessary to establish a complete closed loop of “failure diagnosis — data formulation — collection and processing — training and recharging — actual machine review” to actually achieve model capability gain. At the same time, reusability after data migration to different entities is an important guarantee for the profitability of data companies after competitive homogenization.
Risk Alerts
The development of robots fell short of expectations, data collection technology routes changed drastically, and the main business development of various companies fell short of expectations.