How Profitable is DeepSeek? Its Costs and Profits Are Unveiled: A Profit Margin of 545%!

TradingKey - From competing in capital expenditure, the number of chips, to the cost, the competition in the large AI model field is now being tested by its ability to monetize. While the U.S. tech giants are still burning money on AI, the DeepSeek has revealed astonishing details about its costs and profits, making over three million yuan in profit per day and boasting a profit margin of over 500%.

On Saturday, March 1st, the DeepSeek team released the "Overview of the DeepSeek-V3/R1 Inference System" on the Zhihu platform, disclosing the company's costs and revenues related to its large model. This makes it the first large model development company to disclose its cost structure.

According to the introduction, within the latest 24 hours counted by DeepSeek (from 12:00 on February 27th to 12:00 on February 28th, Beijing time), the company's GPU leasing cost for operating V3 and R1 was $87,072. If all tokens are calculated according to the pricing of the R1 model, the theoretical total revenue for one day is $562,027, and the cost profit margin is as high as 545%.

Calculated annually, the annual revenue of DeepSeek is expected to be over $200 million.

However, DeepSeek also added that the actual revenue is not that high because the pricing of V3 is lower, and the paid services only account for a part, and there are also discounts at night.

All services of DeepSeek V3 and R1 use H800 GPUs and adopt the same precision as that in training. The performance of this kind of chip is far inferior to the chips used by OpenAI and other U.S. tech giants for training AI, yet DeepSeek has developed AI models with comparable performance at a lower price.

The company stated that the optimization goals of the DeepSeek V3/R1 inference system are higher throughput and lower latency. To achieve this, DeepSeek's solution is to use large-scale cross-node expert parallelism (expert parallelism/EP):

• Firstly, EP significantly increases the batch size, thereby improving the efficiency of GPU matrix multiplication and enhancing the throughput.
• Secondly, EP distributes the experts across different GPUs, and each GPU only needs to calculate a small number of experts (therefore with less memory access requirements), thus reducing the latency.

According to the cost and profit situation disclosed by DeepSeek, some investment experts said that if DeepSeek were in the United States, it should be a company with a valuation of $10 billion.