DeepSeek boasts estimated ROI of 545% for its AI services

Chinese language model developers DeepSeek this week released some interesting data on the estimated profitability of their V3 and R1 language models over a 24-hour period. According to the authors of the calculations, these models can earn six and a half times more than they spend on renting computing power.

Image source: Unsplash, Solen Feyissa

In fact, based on the information published by DeepSeek representatives on their GitHub pages, the company spent $87,072 on renting computing accelerators in a randomly selected day, while the potential monetization of its V3 and R1 models could have brought it $562,027 over the same period. By correlating these values, the authors of the calculations obtained a conditional profitability of 545%.

However, it is important to understand that calculations based on this methodology imply a number of assumptions. First of all, potential income was calculated without discounts, and the pricing policy for the more expensive R1 model was taken as a basis. Secondly, not all publicly available DeepSeek services are monetized and are paid for by users. If the fee for access to them was charged at a commercial price, the number of users could decrease, and this would reduce the revenue received.

Finally, the calculations in this example do not take into account DeepSeek’s expenses for electricity, rent for data storage, or for research and development as such. In any case, this attempt to demonstrate its prospects and viability to potential investors should inspire representatives of other startups to publish similar calculations. For now, the field of artificial intelligence requires huge expenses from investors, and the financial return is very ephemeral and distant in time.

DeepSeek explains that the company has achieved the high efficiency of its services through a number of optimizations. Firstly, traffic is distributed between several data processing centers as evenly as possible. Secondly, the time for processing a user’s request is flexibly regulated. Thirdly, the processed data is sorted into batches for optimal load on the infrastructure.