China’s National Data Administration has unveiled a long-term plan to build data infrastructure for artificial intelligence development, calling for the construction of sector-specific databases across 19 industries and five innovation areas. The initiative, analyzed in detail by the Jamestown Foundation’s Samantha Hoffman, establishes a “physically distributed but logically centralized” data management architecture designed to aggregate information collected across government agencies and make it available for AI model training, while simultaneously strengthening the social credit system and the public security apparatus.
The plan represents the most comprehensive articulation yet of how Beijing intends to use its unparalleled access to population-scale data as a strategic asset in the global AI competition. While Western AI developers debate the ethics of data collection and grapple with privacy regulations that limit what data can be used for training, China is building a state-supervised infrastructure that systematically converts government data across every major sector of society into high-quality AI training datasets.
The Architecture: Physically Distributed, Logically Centralized
The plan’s central architectural concept, “physically distributed but logically centralized” (物理分散、逻辑集中), describes a system in which data remains stored in its original location across multiple government agencies and institutions, but is made accessible and interoperable through a unified logical layer. The new National Dataset Management Service Platform is designed to serve as this logical layer, providing a single interface through which authorized parties can discover, access, and combine datasets from across the government.
This architecture addresses a persistent challenge in large-scale data governance: the tension between the efficiency of centralized data management and the security and resilience benefits of distributed storage. By keeping data physically distributed while providing logical centralization, the NDA aims to capture the benefits of both approaches, enabling large-scale data fusion for AI training while avoiding single points of failure and security vulnerabilities associated with fully centralized data warehouses.
The plan also establishes data exchanges as a key mechanism for making bulk data transactable across departmental lines. These state-supervised marketplaces are designed so that transactions are logged, participants verified, and audit trails maintained, with cooperation obligations under the 2021 Data Security Law. The effect is to create a government-controlled market for data that is legible to regulators and auditable by security agencies, a structure that is fundamentally different from the commercial data markets that have developed in Western economies.
The 19 Sectors and Five Innovation Areas
The NDA has directed the construction of sector-specific datasets for AI model training across 19 sectors, including healthcare, e-commerce, public security, urban governance, social credit, agriculture, transportation, and education. Five innovation areas have been designated for special attention: the low-altitude economy (drones and urban air mobility), smart manufacturing, digital finance, intelligent transportation, and AI-assisted governance.
The inclusion of public security and social credit in the list of priority sectors is significant. It confirms that the AI training databases being built under this plan are not purely commercial in orientation, they are explicitly designed to improve the effectiveness of China’s surveillance and social control infrastructure. The datasets will be used to train AI models that can more accurately identify individuals, predict behavior, and assess social credit scores, integrating AI capabilities into the existing apparatus of state control.
Defining “High-Quality Datasets” — Technical and Political Standards
The term “high-quality dataset” has a specific official definition under the NDA’s framework, encompassing standards issued by the National Technical Committee 609 on Data of Standardization Administration of China (SAC/TC609). A high-quality dataset is a collection of data that has been collected, processed, and otherwise handled, can be used directly to develop and train AI models, and can effectively improve model performance.
The standards establish a three-tier classification system for high-quality data, arranged by sensitivity. “General knowledge datasets” contain broadly applicable information with limited security sensitivity. “Industry general knowledge datasets” contain sector-specific information that requires more careful access controls. “Industry professional knowledge datasets” contain the most sensitive and specialized information, subject to the strictest access restrictions and oversight.
The NDA’s High-Quality Dataset Construction Guidelines identify three main technical approaches to data collection: multi-source heterogeneous data fusion collection, which combines data from multiple sources with different formats and structures; edge-side collection, which captures data at the point of generation rather than after the fact; and synthetic data generation, which uses AI to create artificial training data that supplements real-world datasets.
Implications for China’s AI Competitive Position
The data infrastructure plan has direct implications for China’s ability to train and improve AI models. Access to large, high-quality, diverse datasets is one of the most important inputs to frontier AI development, and China’s ability to systematically construct such datasets across 19 sectors gives its AI developers a structural advantage that is difficult for Western competitors to replicate.
China’s token economy has exploded, with daily AI calls jumping from 100 billion to 140 trillion in just 14 months, generating vast amounts of inference data that can feed back into training. The NDA’s plan adds a systematic government data layer on top of this commercial data foundation, creating a training data advantage that compounds over time.
The plan also raises serious concerns about the dual-use nature of China’s AI data infrastructure. The same datasets that train commercial AI models will improve the social credit system and the public security apparatus. The same data exchanges that enable commercial AI development will make government data flows more legible to security agencies. China’s approach to AI governance, as illustrated by Premier Li Qiang’s Summer Davos speech, emphasizes the importance of state oversight and control. The NDA’s data infrastructure plan is the operational expression of that philosophy: AI development as a state-directed enterprise, built on data that serves both economic and security objectives simultaneously.