Chinese AI Firm DeepSeek Discloses $294,000 Training Cost for R1 Model, Challenging U.S. Benchmarks

Chinese artificial intelligence developer DeepSeek has disclosed for the first time that training its R1 AI model cost approximately $294,000—a figure significantly lower than training expenses reported by leading U.S. AI firms. This information, published in a peer-reviewed paper in the journal Nature, is expected to intensify debate regarding China’s position in the ongoing global AI technology competition.

The Hangzhou-based company, which had remained largely discreet since early 2023, included the training cost estimate in the recent article co-authored by its founder, Liang Wenfeng. The paper detailed that the R1 model training utilized 512 Nvidia H800 graphics processing units (GPUs) over 80 hours. Earlier drafts of the paper released in January omitted cost details.

By comparison, Sam Altman, CEO of OpenAI, previously estimated that foundational AI model training expenses had surpassed $100 million, although specific spending figures have not been publicly disclosed.

The cost of training such large-language models encompasses operational expenses over weeks or months, involving extensive computing power to process vast datasets of text and code. DeepSeek’s claim of substantially lower costs has generated skepticism among U.S. industry experts and officials, particularly regarding the hardware deployed.

The Nvidia H800 chips referenced by DeepSeek are tailored for the Chinese market following U.S. export controls enacted in October 2022, which prohibit sales of Nvidia’s more advanced H100 and A100 AI chips to China. While U.S. authorities previously indicated that DeepSeek had accessed large quantities of H100 GPUs despite restrictions, Nvidia has maintained that the company has only used legally obtained H800 units.

In supplementary materials accompanying the Nature publication, DeepSeek acknowledged owning A100 GPUs, which it confirmed were used during early development phases involving smaller model experiments before transitioning exclusively to the H800 cluster for the full R1 training.

Sources have noted that DeepSeek attracted top Chinese AI talent partly due to its unique possession of A100 supercomputing infrastructure, a rarity among domestic competitors amid tightening geopolitical constraints.

This revelation underscores the complexities of technological development and export regulation enforcement in the AI sector, as the U.S. and China continue to vie for leadership in artificial intelligence innovation.