Beijing, 19 September 2025 — Chinese AI developer DeepSeek has revealed that training its reasoning-focused model R1 cost just US$294,000, a figure far lower than similar models built by many U.S. counterparts. The cost disclosure comes from a peer-reviewed article published in Nature, marking the first time the company has revealed R1’s training expenses.
DeepSeek, based in Hangzhou, said it trained R1 using 512 Nvidia H800 GPUs, a chip designed for the Chinese market after export restrictions limited access to higher-end chips like Nvidia’s H100 and A100. While the paper confirms use of H800s for the bulk of training, it also acknowledges that in earlier stages, DeepSeek had used A100 chips for preparatory experiments.
The article notes that DeepSeek’s earlier release of lower-cost AI systems in January 2025 stirred global concern, especially among major AI firms, because of the possibility that they could threaten established players in the AI arms race. Despite DeepSeek maintaining a low public profile since then, this new disclosure adds concrete detail to its claims.
According to the supplementary materials, after initial experimentation using A100 chips, the R1 model was trained for around 80 hours using the 512-chip cluster of H800 GPUs.
Analysts say DeepSeek’s ability to develop a reasoning model at such relatively modest expense could reshape perceptions of what it takes to build high-quality AI models. It raises questions about the scalability and cost efficiency of AI development, especially under constraints like export controls and limited access to cutting-edge chips. For the global AI community, DeepSeek’s example could prompt innovation not just in performance, but in how efficiently performance is delivered.
Source: Reuters
