Chinese AI developer DeepSeek revealed its training costs for the R1 model in a recent Nature article. The company spent just $294,000 on training its model, significantly less than what US rivals have reported. This update is likely to reignite debate over Beijing’s place in the AI race.
DeepSeek has faced skepticism from global investors and US companies over its claims of lower-cost AI systems. However, the new information suggests that the company may be developing more efficient models at a lower cost.
The training costs refer to the expenses incurred from running powerful chips for weeks or months to process vast amounts of text and code. DeepSeek’s R1 model used 512 Nvidia H800 chips, which were designed specifically for the Chinese market due to US export controls.
US officials have previously stated that DeepSeek had access to “large volumes” of more powerful H100 chips after export controls were implemented. However, Nvidia has denied this claim, stating that DeepSeek used lawfully acquired H800 chips.
In response to criticisms from US AI figures, DeepSeek has defended its use of model distillation, a technique where one AI system learns from another, allowing the newer model to reap benefits without incurring associated costs. The company claims that this approach enables broader access to AI-powered technologies while maintaining better performance.
Source: https://edition.cnn.com/2025/09/19/business/deepseek-ai-training-cost-china-intl