Huawei is seeking to capture a larger share of China’s artificial intelligence chip market dominated by Nvidia, helping local companies adopt their rival’s silicon for so-called “inference” tasks.
China’s leading AI companies rely on graphics processing units (GPUs) made by Nvidia to “train” large language models, the US chipmaker’s products worth $3.4 trillion being considered essential to the development of technology.
Instead of challenging Nvidia on training, Huawei is positioning its latest Ascend AI processors as the hardware of choice for Chinese groups running “inference,” the computation undertaken by LLMs to generate a response to a prompt.
The Chinese tech giant is betting that inference will become a bigger source of future demand if the pace of model training slows and AI applications such as chatbots become more widespread.
“Training is important, but it only happens a few times,” said Georgios Zacharopoulos, a senior AI researcher working on accelerating inference at Huawei’s Zurich lab. “Huawei mainly focuses on inference, which will ultimately serve more customers.”
It focuses on the less technically difficult but potentially lucrative route of retrofitting AI models trained on Nvidia products to run on Ascend chips, according to company employees and Ascend customers. Since Nvidia and Ascend GPUs run on different softwareHuawei helps companies use another software tool to make the two systems compatible.
Huawei’s efforts are accompanied by government support. Chinese authorities have urged local tech giants to buy more of Huawei’s artificial intelligence chips and move away from Nvidia.
A person close to Nvidia’s China operations said Huawei was seen internally as the country’s most serious competitor, adding that its chip design capability was “advanced.”
Washington has sought to curb Beijing’s development of AI by controlling exports aimed at hindering the development of sensitive technologies in China.
Unlike their American rivals like OpenAI and Google, companies cannot access the most advanced GPUs in China. But even if Chinese groups can only acquire Nvidia H20 chips, less suited to export controls, the less powerful GPUs remain in high demand because they are considered better than local alternatives.
Analysts and Huawei Researchers said Ascend was not yet ready to replace Nvidia for model training due to technical issues, such as a breakdown in how chips interact with each other within a “cluster.” larger range of AI chips when training larger and larger models.
“Although Ascend chips perform well on a chip-by-chip basis, there is a bottleneck in inter-chip connectivity,” said Lin Qingyuan, China semiconductor analyst at Bernstein. “When you train a big model, you have to break it down into smaller tasks. If one chip fails, the software must find a way for the other chips to take over without delay.
The other challenge for Huawei is convincing developers to abandon Nvidia’s Cuda software, known as the company’s “secret sauce” for its ease of use for developers and its ability to dramatically speed up data processing. .
But Huawei’s soon-to-be-released updated version of its AI chip, the Ascend 910C, should also address these concerns. “We expect this new generation of hardware to come with improved software that makes it more accessible to developers,” said a Huawei employee, who requested anonymity.
Huawei and Nvidia face fierce competition. Chinese Internet Group Baidu and chip designer Cambricon have made great progress in developing AI chips. Meanwhile, in the United States, Amazon and Microsoft are also betting that they can capture more market share in chips for inference purposes as AI applications become more widespread.
Estimates from SemiAnalysis, a chip consultancy, suggest that Nvidia made $12 billion in sales in China last year by shipping 1 million of its H20 chips to the country, selling twice as many AI chips as Huawei with its Ascend 910B.
“Nvidia’s China-specific H20 GPUs make up the majority of AI chips sold in China. But the lead is rapidly shrinking as Huawei increases its manufacturing capacity,” said Dylan Patel, chief analyst at SemiAnalysis.
Industry insiders have warned that Huawei’s AI chip push is also limited by insufficient supply, with two potential customers telling the Financial Times they were unable to obtain the chips.
Huawei did not respond to a request for comment. Nvidia declined to comment.
Analysts said Huawei manufacturing likely faced challenges because of U.S. export controls that have made Chinese manufacturing plants dependent on outdated chipmaking equipment.
The emphasis on inference also indicates an evolutionary dynamic of Chinese AI that differs from that of the United States. Washington’s export controls mean Chinese AI players are not in the same race as their Silicon Valley rivals Meta, x.AI and Elon Musk’s OpenAI to build large mega-clusters of Nvidia’s most advanced GPUs.
“Chinese companies are playing a different game. They pay much more attention to inference than the United States, because it is possible to achieve large efficiency gains even with less powerful chips, which also means they can achieve more commercialization. fast,” said Bernstein analyst Lin.
Chinese companies are betting they can stay competitive in AI by reducing the cost of inference, making it less expensive to run AI applications, he said.
Last month, Hangzhou and Beijing-based startup DeepSeek released its V3 model, which gained attention due to its low training and inference costs compared to comparable models in the United States. .
The company proposed a new way for an AI model to selectively focus on specific parts of the input data to reduce model operating costs. It also used the “expert mixing” technique popular with other Chinese AIs. start-upwhich also helps speed up inference since only part of the model is used to generate an answer.
DeepSeek said Huawei successfully adapted V3 to Ascend, providing detailed instructions to developers on how to use the chip. The FT has already reported that Huawei had sent engineers to help customers migrate from Nvidia to Ascend.
Additional reporting by Zijing Wu in Hong Kong