Bloomberg reports that the U.S. government is investigating whether DeepSeek acquired Nvidia's restricted GPUs for AI workloads through intermediaries in Singapore, bypassing U.S. export restrictions. Concerns have grown as DeepSeek's AI model R1 shows capabilities on par with leading OpenAI and Google models. Adding to the concerns, Singapore's share of Nvidia's revenue increased from 9% to 22% in two years.
DeepSeek has not disclosed the specific hardware used to train its R1 model. However, it previously indicated that it used a limited number of H800 GPUs — 2,048 — to train its V3 model with 671 billion parameters in just two months, or 2.8 million GPU hours. By contrast, it took Meta 11 times more compute resources (30.8 million GPU hours) to train its Llama 3 model with 405 billion parameters using a supercomputer featuring 16,384 H100 GPUs over 54 days. Observers believe that R1 also consumes fewer resources than competing models. However, R1 was likely trained on a more powerful cluster than the one used for V3.
