TLDR: Chinese AI firm DeepSeek has announced that its R1 reasoning-focused AI model was trained for a mere $294,000, a figure significantly lower than those reported by its U.S. counterparts. This revelation, published in the scientific journal Nature, utilized 512 Nvidia H800 chips and is poised to intensify discussions about China’s position in the global artificial intelligence competition.
Hangzhou-based AI developer DeepSeek has made a striking claim regarding the cost-efficiency of its R1 artificial intelligence model, stating that its training incurred an expense of just $294,000. This figure, detailed in a peer-reviewed article published in the academic journal Nature on Wednesday, September 18th or 19th, 2025, stands in stark contrast to the significantly higher training costs cited by leading U.S. AI companies.
The R1 model, described as reasoning-focused, was trained over 80 hours using a cluster of 512 Nvidia H800 chips. These H800 chips were specifically designed by Nvidia for the Chinese market following U.S. export restrictions imposed in October 2022, which prohibited the sale of more powerful H100 and A100 AI processors to China. DeepSeek’s founder, Liang Wenfeng, was listed as one of the co-authors on the Nature article, marking a rare public update from the company regarding its R1 model’s training economics.
This disclosure is expected to reignite the ongoing debate about China’s advancements and competitive standing in the artificial intelligence race. For comparison, Sam Altman, CEO of U.S. AI giant OpenAI, indicated in 2023 that ‘foundational model training’ had cost ‘much more’ than $100 million, though OpenAI has not released specific figures for its models. The substantial difference in reported costs has previously caused market ripples, with DeepSeek’s announcement of lower-cost AI systems in January prompting global investors to sell off tech stocks due to concerns about potential threats to the dominance of AI leaders like Nvidia.
Also Read:
- Alibaba Unveils Open-Source Tongyi DeepResearch AI Agent, Outperforming Industry Leaders
- Nvidia Invests $5 Billion in Intel, Forging Strategic AI and PC Partnership
DeepSeek has also addressed questions regarding its development methodology. The company stated that any similarities observed between its models and those from OpenAI are incidental. This is attributed to its V3 model’s training data, which relied on crawled web pages that inadvertently contained a ‘significant number of OpenAI-model-generated answers,’ leading the base model to ‘acquire knowledge from other powerful models indirectly.’ Despite these explanations, U.S. companies and officials have reportedly questioned some of DeepSeek’s assertions about its development costs and the technology employed.
