China’s draft artificial general intelligence policy focuses on three practical needs: computing power, training data and applications. Beijing is seeking public opinion on the proposal, which aims to strengthen support for AI firms through cloud providers and data companies.
Coordinating access to computing power
Training large language models and other AI systems requires substantial processing power. The draft calls for closer collaboration among cloud providers, universities and companies, linking the organizations that supply computing resources with those that use them.
It also proposes a state-backed centralized platform to allocate public cloud resources according to demand. In the policy’s design, cloud capacity would be treated as a resource that users can access through a shared system, rather than relying only on separate arrangements between individual providers and customers.
The proposal arrives in a market where domestic cloud providers already have different positions. Alibaba accounted for over a third of China’s cloud infrastructure services spending last year, according to Canalys. Huawei, Tencent and Baidu trailed behind.
For AI developers, the policy puts cloud access alongside the other inputs needed to build and train models. Its approach is to connect providers and users more closely, while using a centralized platform to direct public cloud resources based on demand.
Building a supply of Chinese-language data
The second part of the plan addresses the availability of quality Chinese-language data. It encourages the compliant cleansing of datasets, including anonymization. This points to the work involved in preparing data for AI training while accounting for privacy requirements.
Data preparation can involve substantial manual effort. The source article notes that OpenAI relies on Kenyan workers to label training data and remove toxic text. Beijing’s proposal likewise recognizes data processing as a necessary part of developing AI, though its stated focus is on Chinese-language datasets and compliant cleansing.
The government’s big data exchange is also expected to aid data sourcing. Launched in 2021 to facilitate data trading across facets of society, the exchange is presented as a way to help connect organizations with data.
Together, data sourcing and cleansing form a distinct strand of the proposal. Having datasets available is only part of the challenge described: the policy also points to improving their quality and anonymizing them.
Putting AI to work in pilot projects
The third area is the use of AI in specific applications. The draft lists potential pilots in medical diagnosis, drug making, financial risk control, transportation and urban management.
These examples span different kinds of work, from health and drug development to financial decisions and city operations. The policy presents them as potential pilot applications, connecting the development of AI capabilities with opportunities to use them in practical settings.
That focus on applications sits alongside the proposal’s attention to computing power and data. The three areas—resources for training, suitable datasets and uses for AI—describe the main components of the policy’s plan.
Hardware remains part of the plan
The draft also discusses software and hardware infrastructure for AI training. That emphasis comes amid escalating U.S.-China competition, as China seeks to strengthen innovation in key technologies such as semiconductors.
The United States restricts exports of Nvidia’s H100 AI chip to China. Nvidia responded with a less powerful processor for China to circumvent export controls, while domestic companies Huawei and startup Biren are working on alternatives to Nvidia.
Those details put the policy’s infrastructure proposals in a wider context. Cloud platforms and training data are central to the draft, but computing capacity also depends on the hardware available to AI developers. The proposal addresses several parts of that ecosystem as Beijing seeks to support home-grown AI firms.