On March 4th, Huatai Securities pointed out that DeepSeek has been open sourced for 6 consecutive days since February 24th. Based on the previously released model parameters and technical reports, the core code of the Infra layer has been released again, involving module optimization such as MLA, communication computation, matrix multiplication, expert load, file access, etc., aiming to improve the efficiency of the model itself and hardware, and the adaptation of domestic GPUs is progressing smoothly. According to DeepSeek data, if all user requests for the web, app, and API are priced at R1, the daily total revenue will be $562027, with a cost profit margin of 545%. If we consider factors such as V3 pricing and nighttime discounts, with a 50% proportion of paid tokens, we estimate that the cost profit margin is expected to reach 108%, and the optimization effect is significant. Huatai Securities believes that continuous optimization of the model layer is expected to continuously reduce application layer costs and improve application performance. Suggest paying attention to companies with user, data, and scenario advantages in 2B and 2C applications.
Scan code to share