Chinas Deepseek claims theoretical cost-profit ratio of 545% per day

Beijing (Reuters) – Chinese Ki -Startup Deepseek announced some cost and sales data in connection with its HIT V3 and R1 models on Saturday and claimed a theoretical relationship between costs of up to 545% per day, although it warned that the actual income would become significantly lower.

This is the first time that the company based in Hangzhou reveals information about its profit margins from less arithmetically intensive “inference” tasks, the stage after training, in which trained AI models are present, which make predictions or carry out tasks, e.g. B. through chatbots.

The revelation could continue to rattle the AI ​​shares outside of China, which were popular worldwide in January according to web and app chatbots that were powered by the R1 and V3 models.

The sale was partly caused by Deepseek's claims to spend less than $ 6 million for chips used to train the model, let alone what the US rival like Openaai spent.

The chips deepseek claim that it is used, the H800 from NVIDIA, are also much less powerful than what Openai and other US ACI companies have access, so that investors further question the promises of the US ACI companies to issue billions of dollars for state-of-the-art chips.

Deepseek said in a Github post published on Saturday that the assumption that the cost of renting an H800 chip is $ 2 per hour that is daily daily costs for its V3 and R1 models 87,072 USD. In contrast, the theoretical daily income generated by these models is $ 562,027, which leads to a cost-profit ratio of 545%. In one year this would achieve sales of a little more than 200 million US dollars.

However, the company added that the “actual sales are much lower”, since the costs for the use of its V3 model are lower than the R1 model, only a few services are monetized as web and app access, and the developers pay less during the expiry times.

(Reporting by Eduardo Baptista; Editor of Daren Butler)