Chinese artificial intelligence (AI) startup DeepSeek has disclosed some financial details regarding its popular V3 and R1 models. In a statement on Saturday, the company stated that, theoretically, its daily cost-to-profit ratio could reach up to 545%, although the actual revenue remains significantly lower.
The Hangzhou-based company has, for the first time, shared information about its profitability, particularly in relation to the inference phase, which follows model training.
DeepSeek claimed that it used the less powerful H800 chips from Nvidia for model training, incurring a cost of only $6 million. This expenditure is considerably lower compared to the substantial investments made by U.S.-based AI competitors in AI chips.
According to DeepSeek, if the rental cost of each H800 chip is $2 per hour, the daily inference cost for its V3 and R1 models amounts to $87,072. Meanwhile, the estimated daily revenue stands at $562,027, translating to over $200 million annually.
However, the company acknowledged that the actual revenue is lower than these estimates, as some services are still offered for free and developers receive cost reductions during off-peak hours. DeepSeek’s disclosed figures could prompt U.S. AI investors to reassess their strategies in the sector.
