Shanghai and Shenzhen expand AI compute vouchers, nudge developers to Huawei Ascend and other domestic chips
Kai Tanaka ·
Shanghai’s Municipal Commission of Economy and Informatization and Shenzhen’s Development and Reform Commission this week published updates to their AI compute voucher programs, adding explicit preference for domestic accelerators such as Huawei’s Ascend series, according to official notices. The vouchers subsidize GPU hour purchases by startups, universities, and industrial research groups on partner clouds in each city.
The changes matter because they tie public money to the underlying chips used, which will push workloads onto non NVIDIA hardware at scale. That will affect which Chinese large model stacks get optimized and where training happens through the end of the year.
What changed this week
Both cities expanded eligible use cases to include multi modal model training for manufacturing and healthcare, raised per applicant caps for SMEs, and added audit requirements for model provenance and dataset handling cited in the notices. Preference points and higher subsidy ratios apply when applicants choose domestic accelerators, including Huawei Ascend 910B and 310P, Cambricon MLU based cards, and Enflame DTU platforms.
Approved clouds include Alibaba Cloud, Huawei Cloud, Tencent Cloud, Baidu AI Cloud, and state linked providers like Inspur, which are required to publish available capacity by chip family and framework support. The notices state that supported software stacks must include MindSpore and Ascend CANN for training on Ascend, alongside PyTorch and PaddlePaddle.
Policy context and industry response
Local compute vouchers, often called suanli coupons in the documents, have been used since 2023 to lower costs for AI development. The updated terms align with national guidance from MIIT and the National Data Bureau that encourages domestic chip adoption and standardized model risk controls.
Cloud providers have been adding Ascend capacity to meet demand and subsidy rules. Huawei Cloud markets Ascend instances across several regions in China, and Baidu and Alibaba offer Ascend based options alongside constrained NVIDIA inventory. IDC and Canalys data show Alibaba Cloud, Huawei Cloud, and Tencent Cloud remain the top three in China by revenue, so their instance mix will shape developer choices.
For global suppliers, the subsidy design weakens NVIDIA’s pull on new Chinese AI workloads and increases the likelihood that Chinese foundation and industry models are tuned first for Ascend and other local accelerators. That raises the barrier for cross border model portability and creates a larger installed base for China specific AI toolchains.