TASS=Yonhap News

Chinese artificial intelligence (AI) company DeepSeek is reportedly developing its next-generation AI models using Nvidia's latest chips, which the United States has banned from export to China.

DeepSeek is developing a new model by securing thousands of graphics processing units (GPUs) equipped with Nvidia's latest architecture, 'Blackwell,' according to a report by U.S. tech publication The Information, citing multiple sources on January 10 (local time).

Sources said DeepSeek has been acquiring Nvidia chips over the past two years through indirect import routes via countries where purchases are permitted.

First, they secured non-Chinese data centers in Southeast Asia and procured Nvidia chips through official distributors.

Once the chips and servers were installed in these data centers, employees from Nvidia, Dell, and Supermicro would visit to inspect the equipment and verify compliance with export regulations.

However, once this inspection was completed, the servers were reportedly disassembled, and the components were smuggled into China piece by piece.

These components then passed through Chinese customs via false declarations, followed by reassembly and installation in Chinese data centers.

U.S. President Donald Trump announced on January 8 that he had allowed the export of Nvidia's H200 chips, based on the previous generation 'Hopper' architecture, to China, stating that the latest 'Blackwell' chips and the next-generation 'Rubin' chips were not included in this allowance.

An Nvidia spokesperson told The Information, "We have not encountered any substance or tip-off regarding 'ghost data centers' that were purportedly built and then dismantled to smuggle components by deceiving us and our partners. While such smuggling sounds absurd, we track all tips we receive."

However, Reuters recently reported that Nvidia has developed software features to track the location of its chips.

It is expected that once these features are activated, the smuggling of Nvidia chips into China will become more difficult.

Meanwhile, DeepSeek is known to be facing development complexities due to the use of 'Sparse Attention' technology in its new model development process.

This technology significantly reduces inference costs by utilizing only a portion of the model, rather than the entire model, when answering questions. However, sources said that DeepSeek is experiencing difficulties in applying it as the model size increases.

DeepSeek employees hope to launch the next-generation model before the Lunar New Year holiday in February of next year. However, founder Liang Wenpeng is reportedly prioritizing performance over schedule and has not set a deadline.

Earlier this year, DeepSeek shocked the global AI industry by releasing its high-performance open-source inference model 'R1'.

DeepSeek did not respond to The Information's request for comment.