DeepSeek has announced that it spent $294,000 on developing (or training) its popular R1 model. This amount is significantly lower compared to the expenses of U.S. rivals, highlighting China’s advantage in the field of artificial intelligence.
To build the R1 model, DeepSeek used 512 Nvidia H800 chips. These chips are specifically manufactured by Nvidia for the Chinese market due to U.S. export restrictions.
In 2023, Sam Altman stated that developing large AI models costs “over $100 million”, though he did not provide an exact figure. This makes DeepSeek’s disclosure even more intriguing.
The U.S. banned the export of Nvidia’s most powerful AI chips — the H100 and A100 — to China in 2022. Despite this, there were suspicions that DeepSeek possessed A100 chips. However, the company explained that it only used them for “small-scale testing”, while the main training process was completed within 80 hours using H800 chips.
