China’s DeepSeek Shakes Up AI World with Ultra-Low-Cost Training Model

In a groundbreaking disclosure published Wednesday in Nature, Hangzhou-based AI company DeepSeek revealed, for the first time, the costs and technical details behind its large-language model, R1. Co-authored by founder Liang Wenfeng, the article states the model was trained using 512 Nvidia H800 chips over 80 hours, marking a rare glimpse into the financial and computational resources behind cutting-edge AI in China.

Earlier this year, DeepSeek’s release of lower-cost AI systems sent ripples through global tech markets, raising fears among investors that US giants like Nvidia could face competitive pressure. The Nature article also clarified that previous versions of the paper had omitted these critical cost details.

Training large-language models usually takes weeks of computation on powerful processors, often costing tens or hundreds of millions of dollars. OpenAI CEO Sam Altman noted in 2023 that training foundational models exceeded $100 million, though specifics were never shared.

US authorities have scrutinized DeepSeek’s chip usage. While the company lawfully employed Nvidia H800 chips, it confirmed for the first time possession of A100 chips used in early development. Reports suggest that access to advanced processors has enabled DeepSeek to attract top Chinese AI researchers.

The firm also addressed allegations of copying OpenAI models. While US officials claimed DeepSeek “distilled” OpenAI’s tech, the company defended the approach as a cost-cutting and performance-enhancing method, making AI more accessible. DeepSeek acknowledged using Meta’s open-source Llama and incidental OpenAI-generated content in its V3 training data, emphasizing it was not intentional replication.

This first-of-its-kind disclosure provides critical transparency on AI training costs and methods, highlighting China’s growing influence in global AI innovation while fueling debates on ethics, competitiveness, and technological access.