Deepseek to secretly help China’s People’s Army

The chatbots of Chinese start-up Deepseek may be free of charge, but that comes at a high price, a leading US official has told Reuters. According to the official, Deepseek supports the military and intelligence services of the People’s Republic of China. This is said to go far beyond access to artificial intelligence.

On the one hand, according to the report, Deepseek is attempting to circumvent US sanctions against the People’s Republic via letterbox companies abroad. On the other hand, Deepseek freely passes on data about its users and statistics to Chinese intelligence services. “To the best of our knowledge, Deepseek has willfully supported China’s military and intelligence efforts and will likely continue to do so,” the news agency quotes the US State Department official as saying, without giving his name.

He accuses Deepseek of procuring large quantities of Nvidia H100 processors, even though their export to China has been banned since 2022. According to Reuters, this accusation has been confirmed by other sources, although the number of banned chips remains unclear. Nvidia claims that Deepseek only uses lower-performance H800 chips, which is legal. Nevertheless, it is suspected that Deepseek is still trying to obtain banned Nvidia chips via letterbox companies in third countries.

In addition, there is a much simpler, probably legal way to achieve the same goal: Chinese companies are apparently using data centers in third countries to remotely gain access to Nvidia chips that are not allowed to be exported to China. After all, the hardware itself is not as important as the results that can be calculated with it. Malaysia, for example, is currently investigating whether this procedure violates existing regulations. According to Reuters’ informants, Deepseek is also seeking to use such data centers in Southeast Asia.

Deepseek storms onto the AI stage

heise online has invited Deepseek to comment. The company caused a worldwide stir and a fall in Nvidia’s share price at the beginning of the year. Deepseek’s Large Language Model R1 performed better than ChatGPT-o1 – in demanding reasoning and math tests, while allegedly using human, technical and monetary resources much more sparingly.

In addition, Deepseek has published its AI models under an MIT license. This allows anyone to use distilled versions locally on relatively modest hardware or in their own cloud instances. To avoid angering Chinese censors, Deepseek’s models sometimes refuse to provide information on political issues.

(ds)