China's DeepSeek claims theoretical cost-profit ratio of 545% per day

BEIJING (Reuters) - Chinese AI startup DeepSeek connected Saturday disclosed immoderate outgo and gross information related to its deed V3 and R1 models, claiming a theoretical cost-profit ratio of up to 545% per day, though it cautioned that existent gross would beryllium importantly lower.

This marks the archetypal clip the Hangzhou-based institution has revealed immoderate accusation astir its nett margins from little computationally intensive "inference" tasks, the signifier aft grooming that involves trained AI models making predictions oregon performing tasks, specified arsenic done chatbots.

The revelation could further rattle AI stocks extracurricular China that plunged successful January aft web and app chatbots powered by its R1 and V3 models surged successful popularity worldwide.

The sell-off was partially caused by DeepSeek's claims that it spent little than $6 cardinal connected chips utilized to bid the model, overmuch little than what U.S. rivals similar OpenAI person spent.

The chips DeepSeek claims it used, Nvidia's H800, are besides overmuch little almighty than what OpenAI and different U.S. AI firms person entree to, making investors question adjacent further U.S. AI firms' pledges to walk billions of dollars connected cutting-edge chips.

DeepSeek said successful a GitHub station published connected Saturday that assuming the outgo of renting 1 H800 spot is $2 per hour, the full regular inference outgo for its V3 and R1 models is $87,072. In contrast, the theoretical regular gross generated by these models is $562,027, starring to a cost-profit ratio of 545%. In a twelvemonth this would adhd up to conscionable implicit $200 cardinal successful revenue.

However, the steadfast added that its "actual gross is substantially lower" due to the fact that the outgo of utilizing its V3 exemplary is little than the R1 model, lone immoderate services are monetized arsenic web and app entree stay free, and developers wage little during off-peak hours.

(Reporting by Eduardo Baptista; Editing by Daren Butler)