At the end of last month, the user count surpassed Google Gemini. DeepSeek suspended API service recharge yesterday, raising external concerns about insufficient computing power

At the end of last month, the user count surpassed Google Gemini. DeepSeek suspended API service recharge yesterday, raising concerns about insufficient computing power.

According to research firm SimilarWeb, the number of users on the DeepSeek website has surpassed that of Google's Gemini chatbot. Notably, DeepSeek suspended API service recharge yesterday, leading to speculation about whether DeepSeek is experiencing insufficient computing power.

60% Higher Traffic than Gemini

Data from SimilarWeb shows that on January 31, DeepSeek.com had 2.4 million daily visits in the United States, which is 60% higher than Gemini's 1.5 million visits, despite Gemini being available for nearly two years. However, OpenAI's ChatGPT remains the most used, with a daily visit count of 19.3 million on the same day. On January 28, DeepSeek recorded a peak visit count of 4.9 million.

Outside the United States, the traffic gap between DeepSeek and Gemini is even larger. On January 31, DeepSeek's global (excluding China) visits reached 29.2 million, more than three times that of Gemini.

Additionally, statistics indicate that within 20 days of launching, the DeepSeek application has surpassed 20 million daily active users, exceeding the initial situation of ChatGPT and becoming the fastest-growing AI application globally.

"Current Server Resources are Tight"

However, following DeepSeek's explosive popularity, a surge of new users may have led to tight computing resources. Reports indicate that DeepSeek suddenly suspended its API service recharge on February 6, with the button showing a gray unavailable status. Subsequently, DeepSeek stated, "Current server resources are tight. To avoid impacting your business, we have suspended API service recharge. Existing recharge amounts can still be used, thank you for your understanding!"

The suspension of API service recharge by DeepSeek has raised speculation about whether DeepSeek is experiencing insufficient computing power. During the Spring Festival, DeepSeek still launched multiple products, causing a rapid increase in user numbers in a short period CITIC Securities analysts estimate that assuming DeepSeek has an average daily visit of 100 million times, with 10 questions per visit and each question's response using 1,000 tokens, which roughly corresponds to 750 English letters, the reasoning computing power demand for DeepSeek would be 1.6*10^19 TOPs per second. In this ordinary reasoning scenario, assuming DeepSeek uses FP8 precision H100 cards for reasoning with a utilization rate of 50%, the demand for H100 cards would be 16,177 units, and for A100 cards, it would be 51,282 units.

Increasing Demand Necessitates Infrastructure Investment

CITIC Securities further indicates that with the gradual popularization of low-cost reasoning models led by DeepSeek, reasoning costs and prices have been significantly reduced, which will lead to a booming development on the application side, in turn driving a multiple increase in total computing power demand. The analysis suggests that as the user base continues to grow, DeepSeek may struggle to meet the increasing demand. If this continues, DeepSeek will have to follow OpenAI's approach and increase investment in infrastructure.

Additionally, it is worth noting that some of DeepSeek's API services are currently still in a "promotional period." According to the official price list, the discount period for the deepseek-chat model ends at 24:00 Beijing time on February 8, after which the charges will be 2 yuan per million tokens for input and 8 yuan per million tokens for output