On Saturday, a Chinese AI startup revealed cost and revenue details regarding its popular V3 and R1 models, citing a potential daily cost-profit ratio of up to 545%. However, it cautioned that actual revenue would be considerably lower.
This disclosure is the first time the Hangzhou-based company has shared information on profit margins from less computationally intensive "inference" tasks, where trained AI models provide predictions or perform tasks like chatbots.
The news might unsettle AI stocks internationally, especially after the success of web and app chatbots powered by the R1 and V3 models.
The recent drop in stock value was partly due to DeepSeek's claims of spending less than $6 million on training chips, significantly less than U.S. competitors such as OpenAI.
DeepSeek stated in a GitHub post that the daily inference cost for the V3 and R1 models using the Nvidia H800 chip amounts to $87,072. Conversely, the theoretical daily revenue is $562,027, resulting in a cost-profit ratio of 545% and a potential annual revenue exceeding $200 million.
Nevertheless, the company acknowledged that the "actual revenue is substantially lower" due to lower costs for the V3 model, some free services, and remaining areas of monetization.