The cost of training DeepSeek-R1 has become one of the main topics around the Chinese startup, as the developers managed to reach the OpenAI level at a much lower cost. Although there were different assumptions before, SemiAnalysis analysts report that DeepSeek uses about 50 thousand Hopper AI chips from NVIDIA and expects to deliver another 10 thousand.
Due to US export restrictions to China, DeepSeek uses not only H100 chips, which were the best in the industry before BlackWell's release, but also the less powerful H800, H20, and A100 chips specially designed for China. The Chinese startup has the most H20 chips - about 30 thousand, including additional orders. This is followed by 10 thousand H100, H800, and A100.
According to the SemiAnalysis report, the startup's total investment in servers is approximately $1.6 billion, of which about $944 million was spent on operational needs. The analysts also note that the $6 million for which the V3 model was allegedly trained is only a part of the preliminary training costs and a small part of the total expenses. This amount covers only the cost of the processors used for pre-training.