China’s most advanced AI models remain dependent on Nvidia chips for training, as the complex task of migrating to local semiconductor architectures slows Beijing’s ambition for technological self-reliance.
- Nvidia’s CUDA ecosystem remains the AI training standard in China
- High rewriting costs slow adoption of Huawei Ascend and other local chips
- Some Chinese teams have begun training models fully on domestic hardware
What happened
China’s leading AI development teams continue to train their most advanced large language models (LLMs) on Nvidia chips, primarily due to the entrenched software ecosystem and high costs of switching to domestic semiconductor architectures. Nvidia’s CUDA platform has long been the industry standard for AI training, creating a significant barrier to fully adopting alternatives like Huawei’s Ascend chips, which require extensive code rewriting and optimization.
While domestic AI hardware capabilities are improving, major Chinese AI players report that transitioning training workflows involves considerable engineering effort and costs. For open-source models, this transition may take several engineers an additional month, but for more complex proprietary models with limited source code availability, the shift could span over six months and require large engineering teams.
Why it matters
China’s reliance on Nvidia chips highlights the challenges it faces in reducing dependence on foreign technology amid broader geopolitical and supply chain tensions. Training AI models demands highly specialized software tools and infrastructure, and Nvidia’s CUDA framework remains unmatched in ecosystem support, making it difficult for domestic chips to replace foreign hardware at scale.
This technological dependency sheds light on the broader hurdle of semiconductor self-sufficiency in China. Even though domestic companies have made strides, the practical realities of migrating extensive AI training infrastructure create a bottleneck for Beijing’s ambitions to fully localize critical AI development processes in the near term.
What to watch next
Observe the progress of Chinese tech firms like Meituan, which have reportedly started training large models entirely on domestic hardware clusters, signaling incremental success in this transition. These efforts may provide a blueprint for overcoming the software and engineering challenges holding back broader adoption.
Additionally, monitor the evolution of domestic AI software platforms like Huawei’s Compute Architecture for Neural Networks (CANN) and the development of optimized tools to reduce migration costs. Advances in these ecosystems will be critical to enabling a more widespread shift away from Nvidia-based training infrastructures.