Link CopyTwitterLinkedinWhatsappTop Chinese AI Models Trained On Nvidia Despite Self Sufficiency PushChina's most advanced large language models are still being built on Nvidia silicon, four years after Washington first restricted chip exports to Beijing. The reason is not a shortage of domestic hardware. It is the sheer cost of rewriting the software that makes AI training possible.Industry sources at major Chinese LLM developers confirmed that training on Nvidia chips remains the norm across the country's AI labs. Domestic chips from Huawei and others have improved to the point where they can handle inference — the lighter task of running a trained model — and, in some cases, even pre-training at scale. The bottleneck is the ecosystem. Nvidia's Compute Unified Device Architecture (CUDA) has been the industry standard for AI development since 2007. Huawei's alternative, Compute Architecture for Neural Networks (CANN), requires developers to rewrite and optimise large amounts of code. One AI researcher involved in model development put the additional time and cost of migrating existing workflows to Huawei's Ascend chips at 50 per cent or more.The gap between "can train" and "does train" on domestic chips is the story behind the headlines. It explains why Moonshot AI's Kimi K3, a 2.8-trillion-parameter model released in July 2026, was reportedly trained partly on Nvidia Blackwell processors obtained through Chinese cloud providers. It explains why Alibaba's 2.4-trillion-parameter Qwen3.8-Max also relied on Nvidia hardware, including Blackwell-class chips. And it explains why DeepSeek, the Hangzhou lab that shook markets in early 2025 with models rivalling US offerings, still used foreign silicon for training its V4-pro model even as it ran inference on domestic chips.About The AuthorAt heart, I am a storyteller drawn to the watershed moments that bend the technology landscape. I braid narrative with data, humanise statistics, and trace the arc from first spark to world-changing impact. My reportage, features and reviews are witty, sardonic, visual and vivid, using anecdote to illuminate rather than eviscerate.
Top Chinese AI Models Trained On Nvidia Despite Self Sufficiency Push
China's most advanced large language models are still being built on Nvidia silicon, four years after Washington first restricted chip exports to Beijing. The reason is not a shortage of domestic hardware. It is the sheer cost of rewriting the software that makes AI training possible.
TL;DRAI
Moonshot, Alibaba and DeepSeek train models on Nvidia Blackwell, because CUDA-to-CANN migration costs 50%+ more. Software is the bottleneck, not chips; this Nvidia dependency persists, constraining tech roadmaps for enterprise IT managers seeking alternatives.
2,468 words~11 min read
