DeepSeek Claims $294K Training Cost for R1 AI Model, Stirring Global Debate

Chinese AI developer DeepSeek has revealed that training its reasoning-focused R1 model cost only $294,000, a fraction of what U.S. rivals reportedly spend. The disclosure, published in Nature, marks the first official estimate from the Hangzhou-based company and is expected to reignite debate over China’s role in the global AI race.

The R1 model, co-authored by founder Liang Wenfeng, was trained on 512 Nvidia H800 chips. This contrasts sharply with U.S. AI leaders, where OpenAI’s CEO Sam Altman has said training foundation models cost “much more” than $100 million. DeepSeek’s claims highlight its strategy of delivering lower-cost AI, which previously caused global investors to dump tech stocks over fears of disrupted market dominance.

DeepSeek acknowledged in supplementary documents that it also owns Nvidia A100 chips, used during early development stages before shifting to H800s for R1’s final 80-hour training run. While Nvidia confirmed the company’s lawful use of H800s, U.S. officials suggested DeepSeek may have had access to restricted H100 chips.

The company also addressed ongoing concerns about AI “model distillation,” a process in which one AI learns from another to reduce costs. Critics, including U.S. officials, have accused DeepSeek of distilling OpenAI’s systems. DeepSeek defended the practice, emphasizing efficiency and broader accessibility. It admitted some training data for its V3 model included web content with OpenAI-generated responses but said this was incidental, not intentional.

DeepSeek’s rare update offers new insight into its technology and cost-efficient methods, reinforcing China’s growing presence in AI innovation. The revelation is likely to intensify scrutiny from global rivals and regulators as competition for AI dominance accelerates.