Analysts revealed DeepSeek's actual expenses for training its AI system.

The Chinese AI startup DeepSeek has spent approximately $1.6 billion on training its large language model with reasoning capabilities, R1, whereas earlier reports mentioned only $6 million.

This is discussed in a report from SemiAnalysis.

It is reported that the training costs for DeepSeek-R1 have become a major topic of discussion around the Chinese startup, as its developers managed to reach OpenAI's level at significantly lower expenses. Contrary to various prior assumptions, analysts at SemiAnalysis note that DeepSeek is utilizing around 50,000 NVIDIA Hopper AI chips and is expecting an additional 10,000 to be delivered.

Due to U.S. export restrictions on China, DeepSeek is using not only H100 chips, which were considered the best in the industry before the release of BlackWell, but also less powerful H800 chips, specially designed for the Chinese market, H20, as well as A100 chips.

The largest share of available chips in the startup consists of H20 – approximately 30,000, including additional orders. Additionally, the company holds 10,000 each of H100, H800, and A100.

Background. Earlier, Mind reported that the Italian Data Protection Authority (Garante per la protezione dei dati personali) urgently restricted the processing of data of Italian users by Chinese companies providing the DeepSeek service.