A batch of national standards took effect in China on 1 March 2026. The two most directly relevant to the AI industry are GB/T 46069.1-2025 (Artificial Intelligence — Operator interfaces — Part 1: Basic mathematics) and GB/T 46069.2-2025 (Part 2: Neural networks).

What the standards govern

Operator interfaces are the seam between frameworks, compilers and chips. Whether the same model code runs reliably across different vendors’ AI processors depends heavily on whether that layer is consistent.

The two standards set out requirements for parameter information, programming languages, interface consistency and data structures for basic-mathematics and neural-network operators respectively. The stated aim is to improve compatibility between software and mainstream AI processors and to support a healthier hardware-software ecosystem.

GB/T standards are recommended, not mandatory. Breaching them does not in itself trigger an administrative penalty. But recommended standards shape industry behaviour through three other routes:

First, procurement and tendering. Government and state-owned customers typically score compliance with national standards as a technical criterion, turning the standard into a de facto entry threshold.

Second, integration testing and acceptance. Once interface consistency has a common definition, disputes between integrators and chipmakers narrow and acceptance criteria become clearer.

Third, regulatory interfaces. The closer a company’s description of model architecture, training framework and deployment environment is to standardised language, the smoother formal review tends to be.

For companies, the question is not “are we compliant” but “has the standard already become the yardstick our customers use”.

What it means for three types of company

Companies with in-house frameworks or inference engines. Interface consistency directly affects the cost of porting across chips. If a product must support both domestic and mainstream compute, aligning early reduces duplicated work — see our analysis of the first trillion-parameter model trained and run entirely on domestic compute.

Companies building agents and interoperability products. Operator interfaces address the lower layers; interoperability between agents is governed by a separate standards family — see our note on China’s seven national standards on AI agent interconnection. The two lines running in parallel show that standardisation is shifting from single-point technology to system coordination.

Companies that need algorithm filing. The more standardised the technical documentation, the smoother filing tends to be — see our algorithm filing service.

Standards and safety governance run in parallel

Technical standards answer “how do things connect”; safety governance answers “who is responsible when something goes wrong”. One does not substitute for the other. Companies aligning with technical standards should also track the governance framework, which continues to evolve — see our note on China’s AI governance and development arrangements.

For AI companies, the standards effective on 1 March are less a compliance event than a signal: the second half of industry competition is not about single-point capability but about delivering reliably on agreed interfaces and rules.

This article is based on public information from China’s national standards disclosure system and the Standardisation Administration for March 2026. It is for general reference only and does not constitute legal advice.