DeepSeek, a Chinese AI application that rapidly became the most downloaded free app on the App Store in early 2025, has recently been under scrutiny—not for its popularity but for allegations surrounding the technology behind its next-generation AI model. While the training cost of DeepSeek was approximately $6 million, far less than industry giants GPT-4 or Claude 3, concerns about the hardware powering its development have surfaced amid claims of potential illegal GPU imports.
Poor Accuracy and Controversial Bias Call DeepSeek's Performance into Question
Independent testing revealed DeepSeek's accuracy hovered at a disappointing 17%, ranking it near the bottom in comparisons against major chatbots like ChatGPT and Gemini. The AI generated false claims 39% of the time, and for news-related inquiries, over half of its responses were either vague or unhelpful. Furthermore, the bot frequently aligned answers with Beijing’s official stance on contentious topics—even when no connection to China was indicated in queries—raising questions about potential political bias in its responses.
Reports Suggest Smuggled Nvidia Blackwell GPUs Are Powering DeepSeek
A recent investigative report from The Information alleges that DeepSeek is utilizing thousands of Nvidia GPUs based on the advanced Blackwell microarchitecture, specifically the B100 and B200 models, to train its AI models. These GPUs are officially banned from export to China by the U.S. government due to national security concerns and efforts to maintain leadership in the global AI race. The report claims approximately 2,000 to 2,300 units were procured through a covert operation involving phantom data centers in Southeast Asia, subsequently dismantled and shipped to China disguised as other cargo.
Nvidia, however, dismisses these claims as "far-fetched," stating no credible evidence or tips have been received that support the existence of such phantom data centers. Despite this denial, Nvidia recently implemented a "digital leash" system to better track the physical location of its GPUs, suggesting the company takes smuggling concerns seriously. This skepticism is further compounded by recent Department of Justice enforcement actions, which disrupted illegal shipments of Nvidia’s H100 and H200 GPUs valued over $160 million destined for China.
DeepSeek Denies Blackwell GPU Usage, Points to Domestic and Alternative Hardware
DeepSeek insists that it is not employing smuggled Blackwell GPUs but instead relies on Nvidia’s H800 GPUs alongside Huawei’s domestically produced Ascend 910C AI chips for training its next-generation AI models. Huawei’s Ascend 910 series, built on SMIC’s 7nm (N+2) manufacturing process, serves as China’s high-end AI chip benchmark but still lags behind Nvidia’s advanced 4nm H200 GPUs fabricated by TSMC.
US-China Chip Restrictions and the Complexities of AI Hardware Trade
In a complex geopolitical and economic move, former U.S. President Donald J. Trump authorized Nvidia to supply its H200 GPUs to select Chinese customers, with the U.S. government taking a 25% cut of sales revenue. Beijing, however, continues to restrict widespread access to these GPUs domestically, favoring the advancement of Chinese chipmakers such as Huawei despite limitations posed by restricted access to cutting-edge lithography equipment.
The ongoing tensions illustrate the difficulties in controlling high-tech hardware flows amid global competition and the strategic importance of AI in future technological dominance.
