China’s DeepSeek Reveals $294K AI Training Cost, Raising Global Eyebrows

Chinese AI developer DeepSeek has disclosed that training its R1 model cost just $294,000, far lower than the hundreds of millions reportedly spent by U.S. rivals like OpenAI. The revelation, published Wednesday in Nature, is expected to rekindle debates about Beijing’s role in the global AI race.

DeepSeek’s $294K AI

The Hangzhou-based company, led by founder Liang Wenfeng, said R1 was trained using 512 Nvidia H800 chips over 80 hours. While the chips were designed specifically for China following U.S. export restrictions, DeepSeek also admitted for the first time that it had used A100 GPUs during preparatory stages of development.

Training costs for AI large-language models typically run into the tens or hundreds of millions, with OpenAI’s Sam Altman previously noting that foundation model development has cost “much more” than $100 million. DeepSeek’s claim of drastically lower costs may challenge Western AI dominance and intensify scrutiny over its methods.

The company has faced criticism for allegedly relying on “model distillation” — a process where one AI system learns from another — raising concerns it may have indirectly leveraged U.S. models like OpenAI’s GPT. DeepSeek maintains distillation makes models cheaper and more efficient, broadening access to AI technologies.

The Nature article also acknowledged that some training data for its V3 model contained OpenAI-generated responses from crawled web pages, though the company insisted this was “incidental.”

DeepSeek’s transparency marks a rare move after months of silence since its low-cost models rattled global markets in January, sparking investor fears about new competition for U.S. chip giants like Nvidia.

With Nvidia confirming DeepSeek lawfully used H800 chips, and U.S. officials alleging it may have obtained restricted hardware, the company now sits at the heart of geopolitical and technological tensions shaping the future of AI.