Chinese AI start-up DeepSeek on Saturday revealed some cost and revenue-related financial numbers that displayed its ‘theoretical profit’ to be more than five times the cost. The data associated with DeepSeek’s V-3 and R-1 models claimed a cost-profit ratio of up to 545% per day.
The start-up, however, clarified that these are hypothetical numbers and that the actual revenue could be significantly lower. This is because it has monetised only a small set of its services and offers discounts during off-peak hours. The costs also don’t factor in all the R&D and training expenses for building its models.
The Hangzhou-based start-up is known for developing powerful AI models at a significantly lower cost. This marks the first time it has elaborated on the profit margins it makes from less computationally intensive "inference" tasks. Inferencing refers to the computing power, electricity, data storage and other resources needed to make AI models work in real time.
This reveal comes at a time when the profitability of AI startups and their models is a buzzing topic among technology investors.
DeepSeek Operations Overview
DeepSeek also gave an overview of its operations that involved the process of compute power optimisation by load balancing. It means managing the traffic in a way that the work is evenly distributed between multiple servers and data centers. The start-up claimed that it innovated to optimise the amount of data processed by the AI model in a given time period, and managed latency (the wait time between a user prompting a query and receiving the response).
The sudden rise of China’s AI start-up DeepSeek took the entire tech industry by surprise, especially the giants sitting in California’s Bay area. Founded in 2023, the AI start-up launched its new large language model (LLM), DeepSeek- R1 in January 2025. This model shook the AI ecosystem, witnessing unprecedented and exponential growth.
DeepSeek gained this level of recognition because, while its competitors spent billions on building, training, and enhancing their models before offering them at a higher cost, the startup developed and trained its model at a fraction of the cost, kept it open-source, and made it available to the public at a significantly lower price. DeepSeek made its model more collaborative and economical without compromising the quality and efficiency of the model.
The pricing structure shows a significant difference, with DeepSeek-V3 being substantially more economical at $0.14 per mn input tokens and $0.28 per mn output tokens, compared to o1 Preview's $15.00 and $60.00 respectively.
