The Chinese company Deepseek first published data on the costs and income related to the work of its popular artificial intelligence models V3 and R1. According to published information, theoretical profitability from the use of these models can href=”https://github.com/deepseek-ai/open-infra-index/blob/main/2025022opensourceweek/day_6_one_more_thing_deepseekv3r1_inferanste_system_oVerview.md”> achieve 545% per day. However, the company warned that the actual income is much lower.
Deepseek has become known thanks to the development of effective and inexpensive AI models. In January, her technology caused a commotion in the world market, when chat bots based on R1 and V3 rapidly gained popularity. This led to the fall of the quotations of American companies engaged in the development of AI, since investors began to doubt the justification of billions of investments in powerful chips and infrastructure.
The main reason for the unrest was the deepseek that the cost of training its models was less than $ 6 million. For comparison, Openai and other American companies spend many times more. In addition, DeepSeek used the NVIDIA H800 relatively weak graphic processors, while American companies are based on more powerful solutions.
In her report on Github Deepseek, she revealed that at the cost of renting one H800 chip $ 2 per hour, daily costs for the use of models V3 and R1 are $ 87,072. At the same time, theoretical income from their work reaches $ 562 027, which gives the expected annual revenue of more than $ 200 millions.
Despite the high potential indicators, the company clarified that the actual income is significantly lower. This is due to the fact that the V3 model is cheaper to use, many services remain free, and tariffs for developers vary depending on the time of day.