Institution: DeepSeek will drive a multiple increase in total computing power demand

After an overnight surge in popularity, new users flocked in, and DeepSeek may be facing a shortage of computing power.

On Thursday, February 6, DeepSeek suddenly suspended its API service recharge, with the button showing a gray unavailable status.

Subsequently, DeepSeek issued a statement saying, “Current server resources are tight, and to avoid impacting your business, we have suspended API service recharge. The existing recharge amount can still be used, thank you for your understanding!”

As of this morning, DeepSeek's API service recharge remains suspended. This has led to speculation from the outside: Is DeepSeek running out of computing power?

During the Spring Festival, DeepSeek's product update caused it to "explode into popularity," with its user base skyrocketing in a short period.

According to media reports citing domestic AI product ranking statistics, DeepSeek's application reached over 20 million daily active users within 20 days of launch. The daily active users of the DeepSeek application (excluding website data) surpassed those of ChatGPT within the first five days of launch, making it the fastest-growing AI application globally.

According to estimates by Guotai Junan Securities analysts Shu Di and Li Qi, assuming DeepSeek has an average daily visit of 100 million times, with each inquiry being 10 times and each response using 1000 tokens, 1000 tokens roughly correspond to 750 English letters, thus DeepSeek's inference computing power demand per second would be 1.6*10^19 TOPs.

In this ordinary inference scenario, assuming DeepSeek uses FP8 precision H100 cards for inference with a utilization rate of 50%, the demand for H100 cards on the inference side would be 16,177, and for A100 cards, it would be 51,282.

Guotai Junan further stated that with the gradual popularity of low-cost inference models led by DeepSeek, the significant reduction in inference costs and prices will inevitably lead to a boom in application measurement, which in turn will drive a multiple increase in total computing power demand.

Some analysts pointed out that as the user base continues to grow, DeepSeek may struggle to meet the increasing demand. If this situation persists, DeepSeek may have no choice but to follow OpenAI's example and increase investment in infrastructure.

Currently, some of DeepSeek's API services are still within a promotional period The official price list shows that the discount period for the deepseek-chat model ends at 24:00 Beijing time on February 8, 2025. After the discount ends, it will be charged at 2 yuan per million input tokens and 8 yuan per million output tokens. The deepseek-reasoner model will be charged at 4 yuan per million input tokens and 16 yuan per million output tokens from its launch.

Original title: Is DeepSeek's computing power insufficient?