China’s top clouds expand Huawei Ascend AI instances in mainland regions as Nvidia supply stays tight

China’s largest cloud providers are expanding Huawei Ascend AI compute across mainland regions this week, according to Chinese language service catalogs and developer notices. Alibaba Cloud, Baidu AI Cloud, and Tencent Cloud have added or widened access to Ascend 910 series instances for model training and inference in Beijing, Shanghai, and Shenzhen, citing availability and support for domestic frameworks. The move reflects persistent constraints on Nvidia GPUs under United States export rules and signals a structural tilt of China’s enterprise AI stack.

The shift matters because it pushes developers toward Huawei’s CANN runtime and MindSpore ecosystem, along with vendor adaptations of PyTorch and PaddlePaddle for NPUs. Toolchain divergence raises portability costs for multinational firms that operate in China and rely on CUDA outside the mainland.

Baidu has promoted PaddlePaddle support on Ascend and has offered migration guides that map CUDA operators to Ascend kernels. Alibaba engineers have published documentation for training Qwen‑class models on Ascend through PyTorch NPU backends and custom optimizers. Tencent’s developer portals highlight inference services on Ascend with prebuilt images for mainstream large language models and vision tasks, along with connectors to its data lake products.

Local media including Caixin and 36Kr previously reported large 910B procurement by Baidu and ByteDance, and Baidu and Alibaba both disclosed work on non‑CUDA training pipelines in 2023 and 2024. Huawei’s own cloud continues to market Pangu‑series model services on Ascend, but the notable change this quarter is third party clouds positioning Ascend compute alongside their Nvidia offerings for enterprise customers with compliance requirements.

Regulatory pressure adds momentum. The Ministry of Industry and Information Technology has urged adoption of “Xinchuang” indigenous IT stacks in government and critical sectors, and the Cyberspace Administration’s model filing regime favors deployers that can meet data residency and security reviews on certified domestic infrastructure. Cloud sellers are bundling Ascend with on‑premise appliance options for finance and telecom clients that face audits.

For global technology teams, the practical outcome is a two‑track AI infrastructure map. Inside China, production workloads increasingly target Ascend, CANN, and domestically supported frameworks. Outside China, CUDA remains standard. Vendors of MLOps platforms, model hubs, and inference engines that want China revenue will need to maintain first class Ascend support, or risk ceding the market to local providers that already do.