A technical change in the latest model from Chinese artificial intelligence start-up DeepSeek could be a big step towards achieving China’s goal of AI self-sufficiency, as it shows a new level of coordination between local model developers and hardware makers, according to analysts and industry insiders.
This week, the Hangzhou-based AI lab said its new V3.1 model was trained using a UE8M0 FP8 scale data format, which was suitable for the “home-grown chips soon to be released”.
While DeepSeek did not specify the vendor of the implied chips or whether their use would be in training or inferencing, the wording elicited enthusiasm about an upcoming tech breakthrough that could enhance China’s prospects of cutting reliance on imported AI chips such as the graphic processing units (GPUs) from Nvidia.
Do you have questions about the biggest topics and trends from around the world? Get the answers with SCMP Knowledge, our new platform of curated content with explainers, FAQs, analyses and infographics brought to you by our award-winning team.
DeepSeek did not respond to a request for comment on Friday.
Separately, Shanghai-listed shares of Cambricon Technologies, a local GPU designer that is a potential challenger to Nvidia, gained 20 per cent on Friday. The stock has more than doubled from a July low, as mainland investors bet on its growing role in supplying domestic AI chips.
In Hong Kong on Friday, shares of Hua Hong Semiconductor gained 18 per cent, while Semiconductor Manufacturing International Corp was also up 10 per cent amid hopes that the two chip foundries could do the heavy lifting in producing China’s own GPUs.
A DeepSeek sign is seen near its office in Beijing, February 19, 2025. Photo: Reuters alt=A DeepSeek sign is seen near its office in Beijing, February 19, 2025. Photo: Reuters>
The investor optimism appears to be warranted. Zhang Ruiwang, a Beijing-based information technology system architect working in the internet sector, said the UE8M0 FP8 technique, when paired with complementary methods, could open the door for training AI models at even lower costs.
The cost of training new AI models has soared over the past few years, according to research by Epoch AI, a non-profit AI research institute. The combined cost of hardware, including AI accelerator chips, other server components and interconnect hardware, accounted for nearly 70 per cent of the total costs.
The use of the UE8M0 FP8 technique comes with some major improvements for AI stacking, Zhang added. It further reduced the amount of graphics storage and computing power needed to run AI systems, speeding up both training and inferencing of the systems, and it was more “engineering-friendly” for deployment, he said.
