DeepSeek, a Hangzhou-based AI developer, reported that training its R1 model cost US$294,000, according to a paper published in Nature on September 19.

This figure is much lower than estimates for similar efforts by US firms, though OpenAI has not disclosed detailed costs for its models.

DeepSeek trained the R1 model using 512 Nvidia H800 chips, hardware designed for the Chinese market after US export controls restricted access to more advanced chips.

The company also acknowledged for the first time that it owns Nvidia A100 chips and had used them for early development of the R1 model.

US officials told Reuters in June that DeepSeek has access to large volumes of H100 chips, but Nvidia said DeepSeek used lawfully acquired H800 chips.

Training costs for large-language models typically cover expenses for running clusters of powerful chips over extended periods.

DeepSeek had previously attracted top talent in China by operating an A100 supercomputing cluster.

🔗 Source: Reuters

Recent DeepSeek developments