The National Data Administration symposium requires clarifying "data usage needs"; the physical AI data collection segment may lead the way

Article body

On the morning of October 28, the National Data Administration organized a symposium on the work of chain-leading (consortium lead) units in building high-quality industry datasets. The meeting emphasized that the construction of high-quality industry datasets should adhere to the principles of "determining supply by demand, promoting construction through use, and improving quality through evaluation," among which clarifying the data needs for artificial intelligence applications was mentioned as a key task.

The meeting stated that chain-leading units should base themselves on their own needs, proactively expand external demand such as upstream and downstream of the industrial chain, adhere to a demand-oriented approach, and precisely advance dataset construction.

Previously, on October 20, China National Data Group Co., Ltd. was officially inaugurated.

Data collection provides the training "raw material" for embodied large models such as VLA, and is the foundation for building a data flywheel and supporting a commercial closed loop. Unlike text large models, which can scrape corpora from the internet, physical AI cannot rely on web crawling to complete data accumulation. It is necessary to establish temporal correlations among task instructions, environmental observations, robot states, control actions, and execution results. The data must include not only vision and language but also information such as joint poses, force perception, and touch, and meet the requirements of being "accurate, complete, broad, and authentic."

Sinolink Securities stated that the industry estimates that mature embodied large models require at least tens of millions of hours of data, and further progress toward general intelligence may require data on the order of hundreds of millions of hours, while current high-quality training data in the industry is only in the range of hundreds of thousands to one million hours, leaving a supply-demand gap of about 90%.

According to a research report by Zheshang Securities, by early 2026, the globally compliant and usable real physical interaction data will still have a 20-fold gap relative to what is needed to train a single general embodied large model. The industry is currently benefiting from four favorable factors: accelerated construction of standards systems, marketization of data elements, industrialization of data scale, and rapid influx of capital, which are driving the rapid development of embodied intelligence data collection.

In the physical AI field, high-quality data is leading the way, and leading companies in the industry are accelerating commercialization. Scale AI is a global leader in AI data infrastructure, positioned as an AI training data service provider, with core business covering the three major areas of data, evaluation, and applied AI. In 2026, the company's valuation reached $29 billion, and it has order-based cooperation with AI companies, hardware manufacturers, medical institutions, governments, and others.

From the perspective of the industrial chain, Zheshang Securities stated that under a neutral expectation for upstream hardware equipment, the market size is expected to reach 15.3 billion yuan by 2030. Most of it is reuse of embodied intelligence bodies, and the industrial chain is already relatively mature. While barriers and concentration are relatively higher, as upstream "shovel sellers," leading companies are expected to have greater certainty in performance. It is recommended to prioritize attention to cameras and IMUs, which have rigid-demand attributes, as well as some segments with better competitive landscapes and higher barriers; under a neutral expectation for midstream data collection, the market size is expected to reach 174.1 billion yuan by 2030, with a CAGR of 203% from 2026 to 2030. The industry is in a rapid development period from 0 to 1, with clear long-term growth space and greater elasticity. It is recommended to pay attention to dataset suppliers, data collection providers with integrated software and hardware, data post-processing vendors, and simulation platform vendors.

Sinolink Securities stated that the supply gap and technologies that have not yet converged have prompted many players to enter. From the perspective of business paradigms, the business model of data collection equipment and services is the clearest, and it is the segment in the entire data collection industry with the highest order visibility and the most stable cash flow. Leading companies will continue to benefit from the widening supply-demand gap and the industry consolidation brought about by route convergence. In the short term, focus on data collection service providers with orders, production capacity, and delivery capabilities; in the medium and long term, prefer companies that can platformize data processing and extend toward model evaluation, post-training, and deployment feedback.

The institution focuses on three types of companies: (1) traditional leaders with experience in traditional AI data services; (2) simulation leaders that use simulation assets as a foundation to build integrated software and hardware platforms; and (3) full-process platforms that layout a closed loop of "data—simulation—training—transfer—scenario."

(Article source: Science and Technology Innovation Board Daily) Middle of the main text Embedded advertisement