Is China's DeepSeek using smuggled GPUs to train its next-gen AI model?

Report alleges that DeepSeek uses 2,000-2,300 powerful Nvidia B100 and B200 GPUs.

0comments
DeepSeek's logo and wordmark in front of an opaque background.
Remember Chinese AI firm DeepSeek? Early this year it became the most downloaded free app in the App Store replacing ChatGPT at the top. But that is not what made DeepSeek such a big story early in 2025. Typically, it was all about the money since the cost to train DeepSeek was $6 million, which is only 1% to 10% of the money required to train GPT-4 or Claude 3.

DeepSeek tested poorly for accuracy and often took Beijing's position when answering a query


However, later that month we passed along the results of tests that showed DeepSeek's accuracy to be a rather alarming 17% placing it 10th out of 11 chatbots including ChatGPT and Gemini. 39% of the time, DeepSeek responded to a query by giving a fake claim. When it came to news related prompts, 53% of the time the AI app gave a vague response or an answer that was not useful. But even more upsetting was that when responding to false online claims, DeepSeek took Beijing's position on the matter even though the prompts didn't mention China.

The Information recently claimed in a report that DeepSeek is using several thousand smuggled AI chips designed by Nvidia using its advanced Blackwell GPU microarchitecture to train its next-gen AI Large Language Model (LLM). Nvidia's Blackwell AI chips are not allowed to be shipped to China due to U.S. export restrictions. The U.S. is preventing Nvidia from exporting these GPUs to China for a couple of reasons. One is to keep the U.S. ahead of the pack when it comes to the global AI race. The second reason is to prevent such chips from getting into the hands of China's military.


More specifically, the report alleges that DeepSeek is using 2,000 to 2,300 Nvidia B200 and B100 Blackwell-based GPUs. Reportedly, the Nvidia chips were ordered by phantom data centers in Southeast Asia that were allowed to purchase them. Then, they were dismantled and shipped to China disguised as other items. 

Nvidia says reports of smuggled Blackwell chips is "far-fetched


A Nvidia spokesperson said, "We haven’t seen any substantiation or received tips of ‘phantom data centers’ constructed to deceive us and our OEM partners, then deconstructed, smuggled and reconstructed somewhere else. While such smuggling seems far-fetched, we pursue any tip we receive." 

Considering that the Department of Justice (DOJ) recently busted an operation that smuggled into China over $160 million worth of Nvidia's H100 and H200 GPU's, you have to wonder why Nvidia would call reports of smuggled GPUs into China "far-fetched." Additionally, after the initial reports of smuggling were made public, Nvidia added what it calls a "digital leash" to verify a chip's current location. That sounds like Nvidia took the reports of smuggled chips more seriously than it let on.

DeepSeek denies that it is using advanced Blackwell GPUs and claims that it is training its next gen AI model using Nvidia H800 and Huawei Ascend 910C chips. The top domestically designed and manufactured AI chip in China is Huawei's Ascend 910B. It is a workhorse, same as the Nvidia chip it is often compared to, Nvidia's A100. Last week, looking to generate more business for Nvidia and take domestic sales away from China, President Donald J. Trump gave Nvidia approval to ship its H200 AI chip to approved customers. 

How Trump made the U.S. partners with Nvidia in a brilliant, but ultimately flawed, move


Always thinking about ways to generate some income for the U.S., Trump made the U.S. government a partner with Nvidia as it will receive a 25% cut of H200 sales in China. However, Beijing reportedly will limit access of Chinese companies to the H200 which led Nvidia to be very cautious when forecasting sales of the GPU in China. Thus, Trump's move to secure a 25% cut of Chinese sales of the H200, despite actually being a brilliant idea on his part, might not result in a big payday for Uncle Sam.

Recommended For You

Is DeepSeek smuggling powerful Nvidia GPUs to train its next-gen AI model?


The reason why China would limit local firms from buying Nvidia's H200 is pretty simple. The CCP would prefer to see Chinese companies like Huawei create their own powerful GPUs. However, like with its Kirin application processors, Huawei's silicon is held back by the lack of advanced lithography machines, which are not allowed to be shipped to China.  

Huawei's Ascend 910B and Ascend 910C AI chips are built using SMIC's 7nm (N+2) process. Since Nvidia's H200 is built using TSMC's superior 4nm process node, it means that if China did not limit interest in the H200, it probably would be a huge seller in the country.
Google News Follow
Follow us on Google News
COMMENTS (0)

Latest Discussions

by 30zpark • 2
by RxCourier9534 • 13

Recommended For You

FCC OKs Cingular\'s purchase of AT&T Wireless