(Yicai) Jan. 9 -- Many Chinese artificial intelligence companies began seeking ways to slash AI models' training costs last year due to supply restrictions of high-end computing power resources, with some developers cutting relevant costs by as much as 80 percent.

Jijia Technology's visual model training cost fell below USD1 million last year from USD5 million in 2023, co-founder and Chief Scientist Zhu Zheng told Yicai. With the improvement of training efficiency, the Shenzhen-based firm's demand for Nvidia graphics cards is no longer as urgent as before, Zhu added.

The training cost of Hangzhou-based DeepSeek's large language model DeepSeek-V3 was only around USD5.6 million thanks to optimizing the co-design of algorithms, frameworks, and hardware, according to research released at the end of last year. The cost was significantly less than the hundreds of millions of US dollars usually spent on training similar models overseas.

There are many ways to reduce training costs, including cooperating with cloud computing service providers to speed up training and optimize communication and setting, Zhu noted. In addition, data optimization can also provide a large space for improving training efficiency, he said.

However, reducing training costs from USD1 million will be more difficult, according to Zhu. It may be something that AI companies cannot handle alone and may require more support from partners, he pointed out.

Editors: Tang Shihua, Martin Kadiev