DeepSeek, a Chinese AI developer, has revealed that its R1 model cost a mere $294,000 to train, a significantly lower figure than those reported by its American counterparts. This revelation, published in the academic journal Nature, is likely to reignite the discussion surrounding China’s position in the global AI race. The company, which had previously remained largely silent following the January release of its lower-cost AI systems, provided this training cost information in a peer-reviewed article. The model utilized 512 Nvidia H800 chips for 80 hours of training.
This low cost stands in stark contrast to statements made by OpenAI CEO Sam Altman, who indicated that training costs for foundational models exceeded $100 million. This significant disparity has raised questions and fueled speculation about DeepSeek’s methods and the true cost of developing advanced AI systems. The company’s use of Nvidia’s H800 chips, designed for the Chinese market after US export restrictions were implemented, has also drawn scrutiny. While Nvidia maintains that DeepSeek uses lawfully acquired chips, US officials previously alleged that DeepSeek’s access to a large number of H100 chips —a more powerful model restricted from export to China —was unlawful.
The Nature article also addressed previous accusations from US officials suggesting DeepSeek had “distilled” OpenAI’s models. DeepSeek acknowledged using the A100 chips in the preparatory stages and that its V3 model training data included content generated by OpenAI models. However, the company insisted this was unintentional and a byproduct of using crawled web pages, asserting that the distillation process—learning from pre-existing models—is a legitimate technique that results in more efficient and affordable AI development, ultimately broadening access to AI technology.
DeepSeek’s low training costs, combined with its use of both A100 and H800 chips, have sparked significant debate about the competitive landscape of AI development. The company’s transparent accounting of training expenses, albeit delayed, invites further examination into the methodologies employed and the potential implications for the future of AI development globally. Further investigations are needed to fully understand the implications of DeepSeek’s approach and its impact on the global AI landscape.
