DeepSeek, obtaining Nvidia Blackwell chips through "Ghost Data Centers" ... rendering US export controls ineffective

Chinese artificial intelligence(AI) startup DeepSeek(DeepSeek) is reportedly bypassing U.S. export bans by using NVIDIA(NVIDIA)'s latest chips to develop the next-generation AI models. In the context of the U.S. strictly restricting high-performance semiconductors exports to China, this move raises questions about the actual effectiveness of technological controls.

According to The Information, a U.S.-based technology media outlet, on December 10(local time), citing multiple sources, DeepSeek has obtained thousands of graphics processing units(GPU) based on NVIDIA’s latest architecture “Blackwell” and has begun developing AI models. The Blackwell chips are NVIDIA’s next-generation GPUs announced this year, representing the latest product aimed at significantly enhancing AI learning and inference performance.

It is understood that over the past two years, DeepSeek has relied on data centers located in Southeast Asia that are not Chinese-based, using indirect methods to acquire NVIDIA chips. According to sources, DeepSeek purchases chips and servers through official distribution channels within permitted countries, then conducts installation and testing locally. Afterwards, they disassemble the devices into parts, transport them into China, and reassemble them at local data centers. This method is called “Ghost Data Center(Ghost Data Center)” and is regarded as a workaround to evade export restrictions.

The U.S. imposes strict controls on high-performance semiconductor exports, especially targeting NVIDIA’s latest AI chips, to curb China’s military buildup and technological rise. On December 8, President Donald Trump announced that some previous-generation chips, such as the “H200”(based on the Hopper architecture), would be allowed for export to China, but clearly stated that Blackwell and future versions like “Rubin” would remain under export bans.

NVIDIA responded to such smuggling suspicions by stating, “We have not heard of any construction of ghost data centers or smuggling of parts,” denying these allegations. However, the company also mentioned that it has recently developed software features capable of tracking the location of its chips, allowing for tighter monitoring of future chip outflows. In practice, once this feature is enabled, the likelihood of blocking such smuggling routes will increase.

On the other hand, DeepSeek’s development of its next-generation model incorporates “Sparse Attention” technology, which improves inference efficiency. This technique allows AI to generate responses using only parts of input data such as text or speech, rather than all information, thus reducing processing costs. However, as the model scales up, applying this technology becomes more complex, leading to delays in development. DeepSeek founder Liang Wenfeng emphasizes performance and has not set a specific release date, with staff hoping to complete it before the Spring Festival next year.

This situation indicates that the debate over how effective U.S. future technology restrictions on China will be may continue. Some analyses also suggest that if Chinese firms can persist in circumventing restrictions to acquire high-performance AI chips, the competition between the U.S. and China in the global AI industry could enter a new phase.