Beijing (Reuters) – The Chinese AI Startup Deepseek has unveiled some cost and turnover data on Saturday with regard to the HIT V3 and R1 models, which claim a theoretical costs of a maximum of 545% per day, although it warned that the actual income would be considerably lower.
This marks the first time that the company established in Hangzhou has unveiled information about its profit margins from less computational intensive “inference” tasks, the stage after training in which trained AI models are involved predictions or implementation tasks, such as via chatbots.
The revelation could further rattle AI shares outside of China that fell in January after web and app chatbots were driven by its R1 and V3 models that have risen in popularity worldwide.
The sale was partly caused by the claims of Deepseek that it spent less than $ 6 million on chips used to train the model, much less than some American rivals such as OpenAi.
The chips Deepseek claims that the used, Nvidia's H800, are also much less powerful than where OpenAi and other American AI companies have access, so that investors can further ask the commitments of the American AI companies to issue billions of dollars on advanced chips.
Deepseek said in a Github post that was published on Saturday that assuming the costs of renting one H800 chip is $ 2 per hour, the total daily inference costs for its V3 and R1 models are $ 87,072. The theoretical daily income generated by these models, on the other hand, are $ 562,027, which leads to a cost-gain ratio of 545%. In a year this would go up to just over $ 200 million in income.
However, the company added that the actual income of the actual income is considerably lower “because the costs for using the V3 model are lower than the R1 model, only some services are monitored because web and app access remain free and developers pay less during off-peak hours.
(Reporting by Eduardo Baptista; Edit by Daren Butler)
