"DeepSeek Developing New AI Model Using Smuggled Nvidia Latest Chips"

"Brought in Through Indirect Channels as Individual Components"

According to The Information, an information technology (IT) media outlet, Chinese artificial intelligence (AI) company DeepSeek is developing its next-generation AI model using Nvidia's latest chips, which are banned from export to China by the United States, citing multiple sources on December 10 (local time).

AP Yonhap News

The Information reported that DeepSeek has secured thousands of graphics processing units (GPUs) featuring Nvidia's latest 'Blackwell' architecture and is currently developing a new model.

Sources stated that DeepSeek has been acquiring Nvidia chips through indirect channels over the past two years. When chips and servers are installed at non-Chinese data centers in regions such as Southeast Asia, the servers are then dismantled, and the components are brought into China individually.

On December 8, U.S. President Donald Trump announced that the export of Nvidia's 'H200' chips, which are based on the previous-generation 'Hopper' architecture, would be permitted to China. However, he clarified that the latest 'Blackwell' chips and the next-generation 'Rubin' chips would not be included in the export allowance.

A spokesperson for Nvidia stated, "We have not encountered any evidence or reports regarding so-called 'ghost data centers' that are built and then dismantled in order to smuggle components past us and our partners. While such smuggling seems highly implausible, we investigate every report we receive."

Meanwhile, DeepSeek is reportedly facing increased complexity in its new model development due to the use of 'Sparse Attention' technology. This technique allows the model to utilize only a portion, rather than the entirety, when answering questions, significantly reducing inference costs. However, as the model size increases, sources say DeepSeek is encountering difficulties in its application.

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.

This content was produced with the assistance of AI translation services.