DeepSeek Discloses R1 Model Training Cost of US$294,000

Chinese AI developer DeepSeek disclosed in a peer-reviewed article in the journal Nature that it spent US$294,000 to train its reasoning-focused R1 model. The process took 80 hours using a cluster of 512 Nvidia H800 chips. This cost stands in stark contrast to U.S. competitors; OpenAI CEO Sam Altman previously stated that foundational model training costs exceed US$100 million.

In a supplementary document, the company acknowledged for the first time that it owns Nvidia A100 chips, which it used during preparatory stages for smaller models. This admission follows scrutiny from U.S. officials regarding whether the company had access to restricted H100 chips, though Nvidia has maintained that DeepSeek used lawfully acquired H800 chips.

DeepSeek also addressed allegations that it distilled OpenAI's models to develop its own technology. The company defended distillation as a method to increase technology access, admitting that while its V3 model's training data included a significant number of OpenAI-model-generated answers, the inclusion was incidental. DeepSeek further confirmed it used Meta's open-source Llama model for some of its distilled versions.