DeepSeek spent $1.6bn on training its language model

Chinese artificial intelligence startup DeepSeek spent approximately $1.6bn on training its large language model R1, which possesses reasoning capabilities, despite earlier reports of much lower costs of just $6m. This is according to a report by the analytics firm SemiAnalysis.

The cost of training the DeepSeek-R1 model has become a subject of active discussion, as the startup managed to reach the level of OpenAI while spending significantly less. According to analyst information, DeepSeek uses around 50,000 Hopper AI chips from NVIDIA and is also expecting a delivery of another 10,000 such chips.

Due to US export restrictions on technology for China, the startup is forced to use not only H100 chips, which were considered industry leaders before the emergence of Blackwell, but also less powerful H800 chips, as well as H20 and A100 chips specifically developed for the Chinese market.

Among the chips used, H20s account for the largest share — around 30,000, taking additional orders into account. The company also has approximately 10,000 H100, H800, and A100 chips each.