Archive · 19 Sep 2025
DeepSeek AI Model Training Cost: $294,000 Using H800 Chips
DeepSeek, a Chinese company, said it trained its R1 AI model for $294,000, significantly less than what OpenAI estimates.
DeepSeek used 512 Nvidia H800 chips for training.
U.S. officials have questioned DeepSeek's access to these chips.
The H800 chips were designed for the Chinese market.
DeepSeek also acknowledged using A100 chips in initial development stages.
This sparked debate about AI development costs and competition, and the possibility that DeepSeek may have used other AI models in their training.
DeepSeek, a Chinese AI developer, reported a $294,000 training cost for its R1 model.
The information appeared in a peer-reviewed article published Wednesday.
DeepSeek used 512 Nvidia H800 chips for training.
OpenAI's CEO stated training costs for foundational models exceed $100 million.
U.S. officials have questioned DeepSeek’s access to certain AI chips and potential model distillation.
- Who
- DeepSeek, a Chinese AI developer
- What
- Reported a $294,000 training cost for its R1 AI model
- When
- The information appeared in a peer-reviewed article on Wednesday
- Where
- Hangzhou, China
- Why
- To provide transparency on AI model development costs and prompt debate on China's AI development
DeepSeek's Stance
U.S. Concerns
Model Distillation
DeepSeek's Stance
DeepSeek defends distillation as improving performance at lower costs, promoting broader AI access.
U.S. Concerns
Some US officials and AI figures have questioned whether DeepSeek deliberately distilled OpenAI's models into its own.
- Company
- DeepSeek
- Model
- R1
- Training Cost
- $294,000
- Chips Used
- 512 Nvidia H800 chips
- CEO Estimate
- OpenAI CEO estimated training costs over $100 million
