DeepSek Claims Theoretical Profit Ratio 545% Per Day

JAKARTA An artificial intelligence (AI) startup from China, DeepSeek, on Saturday March 1, revealed data related to the costs and revenues of its flagship AI models, V3 and R1. The company claims that its theoretical profit ratio could reach 545% per day, although they admit that actual revenue is much lower.

This is the first time the Hangzhou-based company has shared information about its profit margins from computationally lighter "inference" tasks. Inference is a stage after training, where an AI model that has been trained is used to make predictions or carry out tasks such as chatbots.

This disclosure has the potential to further shake up the AI stock market outside China, which experienced a sharp decline in January 2025. The fall was partly due to the rapidly increasing popularity of R1 and V3 DeepSek model-based chatbots globally.

One of the factors that surprised investors was DeepSeeek's claim that they only spent less than $6 million to buy chips to train their AI model. This amount is much smaller than competing spending from the United States such as OpenAI.

In addition, DeepSek claims that they only use the Nvidia H800 chip in their model training process. This chip has a lower performance than the chips used by OpenAI and other US AI companies, sparking questions about the needs of US companies to invest billions of dollars in advanced chips.

In a post on GitHub on Saturday, DeepSek explained that assuming the cost of renting one Nvidia H800 chip is 2 US dollars per hour, the total daily inference costs for V3 and R1 models only amounted to 87,072 US dollars (Rp1.4 billion). Meanwhile, the theoretical daily income generated by these two models is estimated at 562,027 US dollars (Rp9.3 billion).

With these calculations, the theoretical profit ratio claimed by DeepSee reaches 545%. If calculated within a year, revenue from this model could reach more than 200 million US dollars (Rp 3.3 trillion).

However, DeepSek also insists that its actual revenue is much lower. This is due to several factors, such as lower V3 model usage costs compared to R1, only a few monetized services, as well as free access to web services and apps that are still available. In addition, developers who use DeepSek services are also subject to cheaper rates at hours beyond usage peaks.