The Hangzhou-based company claims that, under optimal conditions, the daily revenue generated by these models could significantly outweigh their operational costs. However, the firm was quick to caution that actual profits would be far less, given the complexities of real-world usage and the fact that only a portion of services are monetized.

This announcement marks the first time DeepSeek has revealed any figures related to its profit margins from “inference” tasks — the stage in AI model operation where trained models make predictions or perform tasks like powering chatbots. While these figures appear promising, the company emphasized that actual revenue would fall well below the theoretical maximum.

Impact on the Global AI Landscape

The news sent ripples through the global AI market, particularly among U.S. firms, whose stocks suffered significant drops earlier this year. DeepSeek’s claims add fuel to the fire, especially since the company reported spending less than $6 million on the chips used for training its models. This is significantly less than what U.S. giants like OpenAI have invested in advanced hardware.

DeepSeek’s use of Nvidia’s H800 chips, which are considered less powerful than those used by U.S. competitors, raises more questions about the efficiency of U.S.-based AI firms’ multi-billion dollar investments in cutting-edge hardware. Investors are increasingly skeptical about the high costs associated with AI training and whether these expenses truly translate into higher profit margins.

The Numbers Behind DeepSeek’s Claims

According to a GitHub post shared on Saturday, DeepSeek calculates that renting one H800 chip costs about $2 per hour, translating to a daily inference cost of $87,072 for its V3 and R1 models. The theoretical daily revenue generated from these models is much higher at $562,027. When factoring in these numbers, the company claims a remarkable 545% cost-profit ratio, which, over a year, could translate to over $200 million in revenue.

However, DeepSeek clarified that the actual revenue would be much lower due to various factors, including the lower operational costs of the V3 model, free access to certain services, and off-peak pricing for developers.

For more details, visit the full article on Reuters.