The US is investigating whether DeepSeek may have imported embargoed NVIDIA high-performance GPUs via Singapore



Chinese AI company DeepSeek released the inference model '

DeepSeek-R1 ' as open source on January 20, 2025. DeepSeek-R1, developed amid strict semiconductor export restrictions imposed by the United States, is said to perform at the same level as OpenAI's inference model 'o1' and has attracted a lot of attention, including winning first place in the App Store's free app rankings . However, it has been reported that the US Department of Commerce is investigating the matter due to concerns that 'DeepSeek may have illegally obtained high-performance AI chips subject to US export restrictions through a Singaporean intermediary.'

US Probing If DeepSeek Got Nvidia Chips From Firms in Singapore - Bloomberg
https://www.bloomberg.com/news/articles/2025-01-31/us-probing-whether-deepseek-got-nvidia-chips-through-singapore



US reported investigating whether China's DeepSeek used restricted AI chips | Fox Business
https://www.foxbusiness.com/technology/us-reportedly-investigating-whether-chinas-deepseek-used-restricted-ai-chips

US investigates whether DeepSeek smuggled Nvidia AI GPUs via Singapore | Tom's Hardware
https://www.tomshardware.com/tech-industry/artificial-intelligence/us-investigates-whether-deepseek-smuggled-nvidia-ai-gpus-via-singapore

DeepSeek claims that DeepSeek-R1 was developed using a 'reduced-performance NVIDIA GPU' that complies with the US semiconductor export restrictions. DeepSeek has not disclosed the specific model name of the GPU used to train DeepSeek-R1, but when it released 'DeepSeek-V3' with 671 billion parameters in December 2024, it said that 'it took about 2.8 million GPU hours to train with 2048 NVIDIA H800s.'

Chinese AI company DeepSeek releases AI model 'DeepSeek-V3' comparable to GPT-4o, with a threatening 671 billion parameters - GIGAZINE



In contrast, Meta reported that Llama 3 training had 405 billion parameters, using a supercomputer equipped with 16,384 NVIDIA high-performance GPUs and H100s . It took 30.8 million GPU hours to complete the training.

Because DeepSeek-R1 shows better performance than Llama 3, suspicions have arisen that it may have been trained on a more powerful GPU cluster than the one used for Llama 3, raising the possibility that it may have been trained on high-performance NVIDIA GPUs, which are subject to import restrictions.



The U.S. Department of Commerce and the Federal Bureau of Investigation (FBI) are investigating the possibility that DeepSeek may have acquired high-performance NVIDIA GPUs through a third-party company in Singapore. Singapore is not subject to export restrictions on high-performance GPUs, and NVIDIA's revenue from Singapore has been increasing rapidly since the third quarter of 2023, when China introduced restrictions on AI chips.




NVIDIA also issued a statement saying, 'Revenue related to Singapore is not indicative of a diversion to China. Many of our customers have entities in Singapore that use those entities for products marketed to the U.S. and Western countries. We insist that our partners comply with all applicable laws, and we take immediate action when we receive information to the contrary.'

'It appears that DeepSeek has been successful in circumventing U.S. export restrictions,' Commerce Secretary Howard Lutnick said. 'China should be able to compete fairly without using U.S. hardware, and if approved, we will step up enforcement of restrictions on the sale of AI chips.' DeepSeek has not commented on the report.

Related Posts:

in Software,   Hardware, Posted by log1r_ut