On 28 July 2026 the National Data Administration held a press conference in Beijing for the 2026 China International Big Data Industry Expo, disclosing a set of progress figures on how data elements are empowering artificial intelligence.
Two sets of numbers
The first is datasets. Seventy-two industry chain-leader units had built more than 45 PB of high-quality datasets covering over 759 high-value scenarios. This means that around key fields such as scientific research and industrial manufacturing, high-quality data is turning from “scattered raw data” into “usable datasets”.
The second is computing power. 1.4 million PFLOPS of intelligent computing power had been connected to the national integrated computing network trial platform. The point of the network is to link dispersed computing resources into a schedulable grid, so that smaller teams can obtain training and inference capability at a lower threshold.
Why high-quality datasets are the key variable
The gap in model capability increasingly depends not on algorithms but on data. The dividend of public internet data is peaking, and industry high-quality datasets are becoming a new scarce resource.
The National Data Administration’s approach is “chain-leader-led plus pilot-first”: industry leaders drive the building of datasets around specific scenarios, which are then opened to companies in the ecosystem. For AI companies, this means that in entering the industry large-model race, whether you can access these datasets may matter more than the model itself. See the related direction in the ‘AI + Human Resources and Social Security’ Opinions.
The computing network: from “buying cards” to “using the grid”
For small and medium AI teams, computing cost has always been a hard constraint. The advance of the integrated computing network, together with local “computing vouchers” and other policies, is changing how computing is obtained — from building your own data centre to on-demand scheduling.
This aligns with the deployment in the Action Plan for High-Quality Development of Cyberspace Enterprises to support companies in connecting to national computing network hubs.
Data compliance remains the precondition
The more centralised and cross-entity datasets become, the more prominent compliance issues are. Companies participating in building and using industry datasets must attend to at least three points: the traceability of data sources and authorisation, the boundaries of personal information and important data processing, and purpose limitation in data-sharing agreements.
For the relevant boundaries, see our Data Compliance service page and the Network Data Security Risk Assessment Measures.
Three points for AI companies
First, proactively connect with chain-leader and pilot-first units to enter the dataset ecosystem. Second, factor computing cost into the product pricing model and use computing vouchers and the computing network to lower marginal cost. Third, when connecting to industry datasets, clarify the authorisation chain first to avoid “usable but unclear provenance”. For support on data authorisation, sharing agreements and compliance arrangements, see our Data Compliance and Commercial Contracts services.
This article is based on public information from the press conference of the National Data Administration on 28 July 2026. It is for reference only and does not constitute legal advice.