China’s major clouds push developers onto domestic AI chips as Nvidia supply stays tight
Kai Tanaka ·
China’s largest cloud providers are steering customers toward instances powered by domestic AI accelerators, according to new Chinese language release notes and pricing pages published this month. Alibaba Cloud, Baidu AI Cloud, Tencent Cloud and Huawei Cloud are promoting training and inference on Huawei Ascend, Baidu Kunlun and Alibaba’s Hanguang chips, citing immediate availability and lower queue times than Nvidia based options.
The timing matters because developer demand for GPT class model fine tuning and high throughput inference remains elevated, while US export controls keep supply of Nvidia alternatives constrained. Beijing has also pressed state owned enterprises and regulated sectors to prioritize indigenous compute, a message repeated by MIIT and provincial governments that fund data center build outs under the East Data, West Compute initiative.
Clouds push domestic accelerators
Huawei Cloud and partners are expanding Ascend 910B based clusters with updated MindSpore and Ascend CANN SDKs, along with PyTorch and TensorFlow adapters that target popular models such as Qwen and Baichuan. Baidu AI Cloud has refreshed PaddlePaddle support for Kunlun 2 cards and published migration guides that map CUDA dependent ops to Kunlun kernels for Llama family weights.
Alibaba Cloud is promoting Hanguang based inference for e commerce and search workloads, bundling its PAI platform with quantization toolkits that target low bit serving on domestic chips. Tencent Cloud lists Ascend and third party domestic GPUs such as Cambricon and Enflame on developer consoles, with prebuilt images for retrieval augmented generation, video generation and vector databases.
Customers report shorter wait times for domestic chip instances compared with Nvidia compatible queues, according to posts on Chinese developer forums tied to these clouds. Hourly prices on Ascend and Kunlun nodes are being discounted against scarce H20 class instances, and several providers have added credits for code migration.
Why this matters for AI supply chains
The shift locks in a parallel AI stack inside China, from chips to compilers and frameworks. It reduces reliance on Nvidia and AMD in the near term, but it also fragments the global developer experience, since kernels, operator libraries and performance tooling differ across Ascend, Kunlun, Cambricon and Enflame.
For global vendors, this creates a larger market for porting layers, inference runtimes and model optimization toolchains that can bridge CUDA centric code to China specific back ends. For policy makers, it underscores how export controls are reshaping compute footprints, not eliminating demand, which will influence where data centers, packaging capacity and memory supply get allocated in the next two years.