A factory has terabytes of sensor logs. A model team wants them. The factory cannot hand them over, because the logs contain customer names and trade secrets.
Everyone in the room agrees the data would be useful. Nobody can safely move it. That is the problem China's trusted data spaces are meant to solve, and the government has just opened a second round of experiments.
What was actually issued
On November 18, the National Data Administration (国家数据局) published a notice on its website: the second batch of trusted data space (可信数据空间) innovation and development pilots, document number Guoshu Zong Ziyuan [2026] No. 53.
State media picked it up the next day. Xinhua reported the launch on November 19, and the People's Daily covered it on November 20.
Here is the stage, stated precisely, because it matters. This is a pilot-application notice, not a draft law and not a regulation that took effect after a legislative vote. It was formally issued and published, so the application window is open now. The requirements inside it bind the organizations that choose to participate; they do not impose new obligations on the general public. Anyone describing this as "new AI rules now in force for everyone" would be reading a stage wrong.
The notice rests on two earlier documents: the Trusted Data Space Development Action Plan (2024–2028), issued in 2024, and the national 15th Five-Year Plan outline, which calls for implementing a trusted data space action plan.
What is different in the second round
Three kinds of pilot remain on the table: enterprise, industry and city-level trusted data spaces. The stated goal is a data market that is open, shared and secure, with rules that allow value to be created and shared rather than just data to be stored.
The new emphasis is the phrase "two-way empowerment between data and artificial intelligence (人工智能)." In practice that means the pilots should not merely build safe warehouses of data. They should produce high-quality data supply that AI training and AI use can actually draw on, while also letting AI tools improve how data is organized and governed.
The notice also points toward a commercial question that has dogged data markets for years: how do you charge? It calls for exploring billing and revenue-sharing methods based on token (词元) calls and on application results. That is a meaningful shift from selling a dataset once to charging for the value delivered each time a model uses it.
Who can apply, and who cannot
The thresholds are specific enough to filter out casual applicants.
- Applicants must generally be independent legal entities registered in China, or operators of national-level innovation platforms such as national laboratories, national science data centers and national AI application pilot bases.
- The applicant must be the operator of the space. Joint applications by multiple organizations are not supported.
- Enterprise spaces must be led by a chain-leading or leading enterprise. Industry spaces should have the ability to integrate that sector's data resources and ideally carry a joint recommendation from the relevant national industry regulator.
- City spaces need the consent of the city government and must show they can integrate the city's data resources.
- Each applicant may submit only one project, and organizations that took part in the first batch in 2025 cannot apply again.
Applicants must also show real technical foundations: an environment for trusted data control, resource interaction and value creation, plus enough operating staff and funding to run a marketplace rather than a demo.
Why AI developers should care
For most of the past decade, the constraint on Chinese AI was not only model architecture. It was access to clean, licensed, domain-specific data. Hospitals, factories and local governments sit on exactly that data, but legal and security concerns keep it behind walls.
A trusted data space tries to change the unit of exchange. Instead of copying a dataset to someone else's server, the data stays where it is, while the system controls who can use it, logs what happened, and measures the result. If the pilot's billing experiments work, a model team could pay for a specific use rather than buy a whole archive.
That matters for small teams. The cost of entering a frontier model market is increasingly the cost of data, not the cost of downloading weights. A mechanism that lets a small lab pay for a narrow, certified slice of a factory's data could be more useful than another open dataset of web text.
Key takeaways
- The National Data Administration published a second batch of trusted data space pilots on November 18, 2026, in a formally issued application notice rather than a draft regulation.
- The three pilot types — enterprise, industry and city — carry over, with a new emphasis on two-way empowerment between data and AI.
- The notice invites experiments with charging by token calls and by application results, pointing toward usage-based rather than one-off data deals.
- Eligibility is narrow: one project per applicant, no joint bids, and no re-entry for 2025 pilot units.
Honest limitations
- This is a pilot notice with an application process, not a nationwide mandate and not proof that any particular billing model works. Selection results, project lists and timelines are not yet published here.
- The "token-based and result-based billing" language describes an avenue for exploration. It does not set a price, a standard or a legal definition, and no specific company or city is named in this article.
- The article reads the notice through the lens of AI training because that is the stated breakthrough point. Data governance, privacy and cross-border questions are larger than that lens and are only touched on here.
What readers can do now
- If your organization holds data that an AI team wants, watch the National Data Administration site for the application guide and the selection list before assuming your sector is included.
- For developers, treat "data you can license narrowly" as a planning assumption rather than a given. Ask vendors what usage rights and logs come with a dataset.
- Follow the first batch's published cases, if any, to see whether usage-based billing produced real revenue or only pilot paperwork.
