appConfigEntityId

bodyType

customContents

entityId

entityUuid

isVideo

articlePageQuery

article

articleSchemaContent

content

hooksArticlePageAdvertisementConfigArticle

hooksArticlePageFloatingButtonConfigArticle

appConfig

filter

json

liveArticle

Minxiao Chang

Ann Cao

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 bottleneck.
“Training LLMs on Nvidia chips for now remains the norm among Chinese AI developers,” said a person familiar with the industry.
One of the core hurdles lies in the software ecosystem. Nvidia’s Compute Unified Device Architecture (CUDA) platform has long been the industry standard for AI development.
By contrast, Huawei Technologies’ alternative – Compute Architecture for Neural Networks (CANN) – requires developers to rewrite and optimise large amounts of code, according to an AI researcher involved in model development.
“Our existing training pipelines are reliant on CUDA,” said James Wang, who develops AI models at a research institute affiliated with a Shanghai-based university. “CUDA code cannot run directly on Ascend and requires extensive rewriting.”
Wang estimated that migrating existing workflows to Huawei’s Ascend chips could add at least 50 per cent in time and costs for his team.
The length of time required to change the type of chip in a system would depend on a model’s openness and the maturity of the surrounding ecosystem, said a Beijing-based engineer involved in transitioning LLM workloads to Ascend.

For open-source models such as China’s DeepSeek and its distilled versions, training on Ascend chips could take two to three engineers about one more month than on Nvidia-based systems, as developers could leverage existing ecosystem support, the engineer said.
By contrast, training a model such as Moonshot AI’s Kimi K3, which has only released its weights – the underlying parameters that encode its intelligence – rather than its source code, could require a team of about 10 engineers more than six months of additional work, the engineer said.
The continued reliance on Nvidia highlights the challenges facing China’s efforts to reduce dependence on foreign supply, while domestic labs rush to advance their AI capabilities.
The AI development lifecycle involves training and inference, the use of a trained model to make predictions using new data. While training an AI model requires complex and intensive resources, the process for inference is much easier. Systems such as DeepSeek-V4 and Moonshot’s Kimi K3 have been successfully adapted for Chinese platforms such as Huawei and Alibaba Group Holding, the owner of the South China Morning Post.
Neither DeepSeek nor Moonshot has disclosed which AI chips it used to train its models.
Kimi K3 was reportedly trained on Nvidia chips, including advanced Blackwell processors, according to a July report by the US outlet The Information, which also said that Chinese AI companies were continuing to rely on Nvidia hardware for training frontier models, despite efforts to expand domestic alternatives.
Some Chinese teams have started using local chips for AI training. In June, Meituan released LongCat-2.0, a 1.6-trillion-parameter model that it said was fully trained and run on a 50,000-card domestic computing power cluster, without naming the supplier.


Philippine tribe vows to stand ground against US-led Pax Silica: ‘listen to us’

As the US weighs restrictions on cloud computing, how will China’s AI sector adapt?

Chinese AI chipmaker Hygon plots expansion from data centres to robotics

China’s CXMT touts new chip manufacturing platform as close to ‘world’s most advanced’

Semiconductors

China’s top AI is still trained on Nvidia chips. What is delaying a switch to local tech?

Huawei Technologies’ Compute Architecture for Neural Networks requires developers to rewrite and optimise large amounts of code. Photo: Shutterstock

Visitors learn about Huawei’s Ascend AI-based CloudMatrix 384 at the 2025 World AI Conference in Shanghai. Photo: Xinhua

China tightens chip-design protection and clarifies rights to spur innovation

People learn about a reconfigurable AI chip at the permanent exhibition at the Zhongguancun Exhibition Center in Beijing. Photo: Xinhua

Cambricon posts 108% surge in first-half revenue amid China’s massive AI chip drive

People visit the booth of Chinese AI chipmaker Cambricon Technologies at an expo in Shanghai. Photo: Getty Images

Unrelenting AI demand spawns new plant for Chinese PCB maker Victory Giant

A production line at Victory Giant Technology in Huizhou, in southern China’s Guangdong province. Photo: Handout

Victory Giant raised US$2.6 billion in its Hong Kong listing in April. Photo: AFP

China launches probe into Palo Alto Networks as US trade tensions intensify

The flags of the US and China on a motherboard. Photo: Shutterstock

Top Huawei chip scientist opens up on strategy, ‘blind arrogance’ and his war-hero mindset

An image of a semiconductor displays the name of Huawei Technology’s chip-design unit, HiSilicon. Photo: handout

Liao Heng joined Huawei’s chip-design unit, HiSilicon, in 2016. Photo: handout

Tech

Big Tech

Tech Trends

Tech War

High transition costs are keeping AI developers in China reliant on Nvidia chips, industry sources say


High transition costs are keeping AI developers in China reliant on Nvidia chips, industry sources say.


China’s top AI is still trained on Nvidia chips. What is delaying a switch to local tech?

.css-1c6uqr6{color:inherit;font-weight:inherit;font-size:inherit;font-family:inherit;line-height:inherit;overflow-wrap:break-word;}High transition costs are keeping AI developers in China reliant on Nvidia chips, industry sources say

High transition costs are keeping AI developers in China reliant on Nvidia chips, industry sources say