Staff Reporter
On Saturday, Chinese AI startup DeepSeek revealed some financial data related to its popular V3 and R1 models, claiming a theoretical cost-profit ratio of up to 545% per day. However, the company cautioned that actual revenue would likely be much lower.
This announcement marks the first time the Hangzhou-based firm has shared insights into its profit margins from less computationally intensive “inference” tasks. These tasks involve trained AI models making predictions or performing functions, such as those used in chatbots.
This news could further impact AI stocks outside of China, which saw a significant decline in January after chatbots powered by DeepSeek’s R1 and V3 models gained widespread popularity.
The recent sell-off in AI stocks can be partly attributed to DeepSeek’s assertion that it spent less than $6 million on chips for training its models, a figure significantly lower than what U.S. competitors like OpenAI have invested.
DeepSeek claims to use Nvidia’s H800 chips, which are less powerful than those available to OpenAI and other American AI companies. This has raised concerns among investors about the validity of U.S. firms’ promises to invest billions in advanced chip technology.
In a GitHub post on Saturday, DeepSeek estimated that, assuming a rental cost of $2 per hour for one H800 chip, the total daily inference cost for its V3 and R1 models would be $87,072. In contrast, the projected daily revenue from these models is $562,027, resulting in a striking cost-profit ratio of 545%. This could translate to over $200 million in revenue annually.
However, the company noted that its “actual revenue is substantially lower.” This is due to the lower usage costs of the V3 model compared to the R1, limited monetization of certain services, and reduced payments during off-peak hours when access remains free for users.
