Sources: Nvidia chips remain the norm for Chinese AI labs training LLMs, as switching from Nvidia's CUDA platform to Huawei's CANN requires major code rewriting
China's most advanced artificial intelligence models are still being trained on Nvidia chips, sources at major Chinese large language model …
Context & Ripple Effects
Chinese AI developers have already been pursuing several routes away from a single supplier: Alibaba and Baidu began using internally designed chips for model training, while Huawei has positioned Ascend chiefly for inference. The new reporting identifies the software migration burden, rather than chip availability alone, as a constraint on changing training platforms.
The divide is consistent with DeepSeek's reported split between Nvidia for training and Huawei for inference after Ascend training problems. It makes CUDA’s installed codebase and tooling a central part of Nvidia’s position in Chinese model development.
First-order effects
- Chinese AI labs continuing to train on Nvidia preserve existing CUDA-based workflows and avoid the major engineering work required to port them to CANN.
- Huawei faces a more immediate adoption barrier for Ascend in LLM training, even as its inference-oriented positioning remains distinct from the training market.
Second-order effects
- Alibaba and Baidu's move toward internally designed training chips becomes more consequential: building a viable alternative requires not only silicon but software stacks developers can use without extensive rewrites.
- A training-on-Nvidia, inference-on-Huawei configuration gains support as a practical response for labs that want to adopt Huawei hardware without migrating their full training codebase.
Third-order effects
- Chinese AI compute is moving toward heterogeneous deployments, where different chips serve training and inference rather than one domestic platform displacing Nvidia across both workloads.
- If CANN migration costs persist, AI-chip competition will be determined as much by developer ecosystems and porting tools as by processors, reinforcing CUDA as a software moat in training.
The trend: China's AI hardware strategy is fragmenting by workload, with software compatibility slowing the shift from Nvidia-led training to domestic alternatives while inference becomes a separate battleground.