China expands AI subsidies with computing power vouchers to boost SME adoption

China's local governments are rolling out "computing power vouchers" that heavily subsidize AI training and inference for small and medium-sized enterprises. Cities that have already rolled out these vouchers are Beijing, Shanghai, Henan, Shandong, Chengdu, Shenzhen, and Ningbo. Shanghai leads the group with about CN¥600 million (about $84 million) in vouchers, covering up to 80 percent of AI rental fees plus CN¥100 million (about $14 million) for data and LLM training. These programs aim to fill underused data centers and push broader AI adoption among smaller firms.

The voucher fits into China's December 2024 policy, called "Implementation Opinions on Promoting the High-Quality Development of the Data Labeling Industry." The program aims to lower R&D costs for smaller companies. Vouchers work by allowing SMEs to redeem credits at their local or national data centers to access computing power below market pricing. Chengdu's local government has expanded their pilot program, now allocating CN¥100 million (about $14 million) for research institutions, while Shandong's provincial government has allocated CN¥30 million (about $4 million) with plans for an additional CN¥1 billion (about $140 million). Meanwhile, Beijing's local government is already taking applications.

China's "Eastern Data, Western Computing" strategy placed many facilities in western regions with cheaper power to serve coastal demand. The rapid expansion left pockets of low utilization, with some sites running at just 20 to 30 percent capacity. The voucher program aims to create a unified, nation-spanning high-power compute network that can route workloads across the country, thereby increasing utilization for underused data centers.

For SMEs, these vouchers allow them to access cheaper training cycles, which in turn enable faster prototyping and broader access to AI resources that were previously priced out of reach. These vouchers also open up a path for local governments to monetize idle capacity and justify the capital expenditure. The effectiveness and take-up rates, however, remain unknown as administrative friction and legal/logistical concerns hamper progress.

Separately, there are also reports about equipping 39 new data centers with 115,000 "illegal" Nvidia Hopper GPUs, though these reports are still unconfirmed. Broader industry market context points to supporting AI adoption, but results will vary by region and by how quickly local governments streamline voucher access and scheduling.