DeepSeek, a Chinese AI firm based in Hangzhou, has revealed that its R1 reasoning-focused artificial intelligence model cost just two hundred ninety-four thousand dollars to train, a figure far lower than those reported for leading U.S. foundational models. The cost estimate appeared in a peer-reviewed paper in Nature, along with technical details showing R1 was trained for eighty hours using a cluster of five hundred twelve Nvidia H800 chips. This is the first time DeepSeek has published such specifics.

The article also discloses that DeepSeek used Nvidia A100 chips during early preparatory experiments for smaller models. The company says the core R1 training used H800 chips, which Nvidia designed for the Chinese market following export restrictions imposed in October 2022 that banned exports of more powerful H100 and A100 models to China.

U.S. officials and companies have questioned some of DeepSeek’s claims. There is an ongoing investigation by the U.S. Commerce Department into whether DeepSeek accessed restricted-export Nvidia chips through intermediaries in Southeast Asia. Nvidia has stated its review indicates DeepSeek used lawfully acquired H800 products and denied use of H100s for the R1 model itself.

The Nature publication follows earlier financial disclosures by DeepSeek, such as claims of high profit margins if users paid for all services. The low training cost for R1 has revived discussions about how export controls, hardware access, and cost efficiencies are shaping global AI competition. Analysts warn that while cost is important, performance, scale, data quality, and compute infrastructure also remain critical for determining impact in the AI field.

The debate now centers on whether DeepSeek’s methods and hardware disclosures will withstand further scrutiny, and what this means for U.S. export policy, China’s AI ambitions, and the strategic leverage that cost-efficient AI models may offer in the fast-evolving global tech race.