On November 10, the Ministry of Natural Resources and the National Development and Reform Commission jointly issued the Implementation Opinions on “AI Plus Natural Resources,” designated Natural Resources Development [2026] No. 150. The document sets two deadlines. By 2028, China aims to establish high-quality datasets, large models and agents across five areas: spatiotemporal geographic information, territorial spatial planning and governance, oceans, geology and mineral resources, and forestry and grasslands, while comprehensively upgrading the intelligence of its “One Map” system. By 2030, it plans to build an integrated, intelligent foundational platform for territorial spatial governance, with secure and controllable core technologies and intelligent equipment.

Those are standard targets found in documents of this kind. The more revealing details are elsewhere: how the ministry plans to obtain computing power, and which legally binding processes will incorporate AI.

Two Sentences on Computing Power

Section 4 is unusually direct: it calls for coordinating the migration of AI data center capacity for natural resources toward the national hubs under the East Data West Computing initiative, and encourages leasing capacity from the national integrated computing power network instead of building new infrastructure.

It is unusual for a ministerial document to explicitly call for “leasing instead of building.” The standard language is to accelerate deployment, coordinate construction and strengthen the foundation. This document takes the opposite approach, effectively acknowledging that building its own infrastructure may not be economical and that leasing capacity from the national network makes more sense.

The rationale is straightforward. Computing workloads in the natural resources system are largely batch-based, offline and non-real-time, including remote-sensing image interpretation, 3D terrain modeling and comparisons of land survey data. These jobs can run for hours or days and do not require millisecond response times. Such workloads are well suited to western China, where electricity is cheaper and data centers have spare capacity, and can readily be scheduled overnight.

But that immediately runs into an unavoidable obstacle: most of the data consists of surveying and mapping products and high-precision geographic information subject to confidentiality requirements. Whether the data can be transferred externally, how far it may travel and who guarantees compliance are not technical questions.

The document’s response is scattered across several sections. Section 4 calls for a trusted data space for natural resources. The section on oceans specifically calls for “high-quality marine datasets and a secure, trusted data circulation network.” The safeguards section also calls for “ensuring compliant data circulation and sharing.” All three point to the same issue.

The premise of a trusted data space is that data remains within its domain and can be used without being exposed: the holder retains control, while the user receives the computing results but not the raw data. Here, the aim is to let classified imagery and map layers be used in training without a computing power scheduling decision triggering a compliance violation.

Whether that model works will depend on the implementation rules. The document goes no further than saying such a system should be developed.

AI Enters the Approval Process, but Accountability Remains Blank

Section 3 lists eight categories of use cases, seven of which cover government functions: monitoring and oversight, planning and governance, asset management, disaster management, marine development, geological exploration, and forestry and grassland conservation. Only the eighth, “AI Plus Earth Science,” concerns basic research. It calls for an “AI superbrain for Earth science” to study interactions among the lithosphere, hydrosphere, atmosphere and biosphere.

Several of the seven government use cases are especially concrete.

“Build an intelligent mining-rights approval system”: mining-rights approval is an administrative licensing decision with legal force. “Develop an agent for cultivated-land protection”: verification under the system for balancing occupied and replenished farmland directly affects local government quotas. “Enable intelligent review of plans”: the system would assess whether planning proposals pass review.

The annex goes further. Column 2 calls for “an AI simulation system for project impacts to assess the comprehensive effects of project implementation on land-sea patterns and ecological security, and explore the automatic generation of assessment reports and optimization recommendations.” Column 4 says geological disaster prevention should include “targeted risk notifications, intelligent alert-and-response evacuation, and emergency-response agents.” Alert-and-response evacuation can mean knocking on doors in the middle of the night and telling people to leave.

These processes have one thing in common: when the system gets a decision wrong, the consequence is not an inaccurate recommendation. It could mean issuing the wrong permit, miscalculating a quota, evacuating people unnecessarily or failing to evacuate them when needed.

How does the document address accountability? Section 5 offers only one sentence: “Strengthen full-lifecycle oversight of AI in the natural resources sector, as well as forward-looking assessments and dynamic monitoring and early warning.” Forward-looking assessments and dynamic monitoring are technical measures concerned with whether models operate reliably. The document does not say who signs off when a model leads an administrative licensing decision astray, or who is responsible when an evacuation alert is wrong.

This is not an omission unique to this document. It remains standard practice for policies of this kind to specify technologies, use cases and platforms while leaving accountability to subsequent regulations and operating procedures. But once AI enters the approval chain, the longer that gap remains, the less willing frontline officials will be to use it.

Five Pilot-Scale Platforms, With No Locations Specified

Section 2 is the shortest, but it carries considerable weight. It calls for the coordinated deployment of five AI pilot-scale platforms covering spatiotemporal geographic information, territorial spatial planning and governance, oceans, geology and mineral resources, and forestry and grasslands. Each would combine five functions: computing power supply, model training, algorithm iteration, use-case validation and ecosystem development.

“Pilot scale” is a manufacturing term for the validation stage between a laboratory result and mass production. Applying it to AI amounts to acknowledging a basic fact: training a model does not make it deployable. It still needs to be validated in real-world operations.

The proposal addresses a specific problem. The natural resources system does not lack models. Its problem is that a model may perform well on one type of data in one province, then fail when moved to another province or a different terrain. These pilot-scale platforms are intended to close that generalization gap.

Yet the section runs to only about 300 Chinese characters. It provides no locations, implementing entities, investment figures or timetable. It does not say who will oversee the five platforms, which cities will host them, or whether funding will come from the central government or require local matching funds.

The document also leaves unexplained how the platforms relate to the existing group of pilot-scale bases. In July 2026, the National Development and Reform Commission said China had established more than 30 national pilot-scale bases for AI applications, while centrally administered and other state-owned enterprises had opened more than 1,000 use cases. The Ministry of Natural Resources now plans five “pilot-scale platforms.” Bases and platforms are different terms that could represent separate systems, but the document does not clarify whether the five will be newly built, established within existing bases, or form a subset of the NDRC-backed program.

That distinction matters. New construction would mean new investment and implementing entities. Designating existing facilities as platforms would use established channels and proceed much faster.

Closing a Loop Opened More Than Two Years Ago

The timeline shows that this document did not emerge from nowhere.

In September 2023, the General Office of the Ministry of Natural Resources issued the Work Plan for Building the National Territorial Spatial Planning Implementation and Monitoring Network (2023-2027), introducing the Territorial Information Model, or TIM, for the first time. It was conceived as an information model for territorial spatial planning with “nationwide coverage, dynamic updates, unified authoritative data, three-dimensional representation, spatiotemporal integration and appropriate precision.”

In June 2026, the State Council’s 20th special study session called for expanding the information-integration capabilities of the territorial spatial planning “One Map” system and building a foundational platform for territorial spatial governance.

On August 20, 2026, a forum guided by the Territorial Spatial Planning Bureau of the Ministry of Natural Resources was held in Shanghai under the theme “Intelligent TIM Development.” At the event, Wu Hongtao, deputy director of the ministry’s information center, said the “One Map” system had largely taken shape. An engineer from the ministry’s Territorial Spatial Planning Research Center said building a TIM for northern Hangzhou had previously taken more than a dozen experts over a month. With agents, one person using one computer completed the work in two to three hours, covering 19 map layers, 45,000 land parcels and more than 500,000 relationships.

In the new document, TIM occupies a different position. The main text says the program will be “based on the Territorial Information Model (TIM).” The first sub-scenario in Column 2 of the annex is titled “AI Plus Territorial Information Model (TIM)” and calls for optimizing TIM development to create an “object-relationship-rule-decision” chain.

TIM has gone from an innovation being tested in pilot projects to the foundation of the entire document. More than any individual use case, that shift shows what has happened over the past two years: TIM is no longer something that needs to prove its value, but infrastructure the system plans to rely on.