China’s top AI is still trained on Nvidia chips. What is delaying a switch to local tech?
China’s most advanced artificial intelligence models are still being trained on Nvidia chips, sources at major Chinese large language model (LLM) developers say, as the prohibitively high cost of switching to local semiconductors continues to hamper Beijing’s push for self-sufficiency. While domestic hardware continues to advance, changing chip architecture presents a steep engineering…
China's most advanced AI models are still trained on Nvidia chips due to the prohibitive cost of switching to local semiconductor technology. Major Chinese large language model developers acknowledge that training these models on Nvidia chips remains the norm for now. The primary obstacle lies in the software ecosystem, specifically Nvidia's CUDA platform, which is the industry standard for AI development.
Huawei's alternative, CANN, requires developers to rewrite and optimize substantial amounts of code, which can significantly increase time and costs by at least 50 percent. The transformation of chip types depends on a model's openness and the maturity of the surrounding ecosystem. For open-source models like China's DeepSeek, training on Ascend chips could take an additional month compared to Nvidia-based systems.
However, training a model like Moonshot AI's Kimi K3, which only released its weights rather than its source code, could demand about six additional months of work. Despite these challenges, China's domestic labs continue to enhance their AI capabilities using foreign hardware. Some Chinese teams have begun utilizing local chips for AI training, like Meituan's 1.6-trillion-parameter model, LongCat-2.0, which was fully trained and run on a 50,000-card domestic computing power cluster.
Written by urgent.news from SCMP Tech's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
