Chinese AI startup ‘DeepSeek’ which already wrecked havoc across the Western AI markets — both financially and technologically — has shared it’s financials for the first time, data which could further bring down AI stocks globally. Sharing financial details related to its AI models, DeepSeek-V3 and DeepSeek-R1, the company highlights how much profit it could theoretically make from running these models, particularly from inference tasks. Notably, inference tasks are those when the AI is used to generate responses rather than being trained.
DeepSeek claims that the daily cost-to-profit ratio for its AI models could reach a staggering 545%, thanks to much cheaper Nvidia GPUs as well as lower tasks needed compared with other AI models. It means that for every $1 spent on running the models, the company could potentially earn $5.45 in revenue.
However, this number is based on ideal conditions, assuming high demand and optimal efficiency. In fact, the two-year-old Hangzhou-based startup DeepSeek itself clarifies the same. DeepSeek mentioned that actual revenue could be much lower due to various factors, including server costs and electricity expenses, fluctuating demand for AI services, pricing strategies and customer acquisition costs.
DeepSeek also revealed some more details in its recent GitHub post. According to the company, the daily cost of running DeepSeek V3 and R1 models is $87,072. Actually, over 24 hours (from February 27, 2025, to February 28, 2025), DeepSeek recorded a peak node occupancy of 278, with an average of 226.75 nodes in operation. Each node is equipped with eight H800 GPUs, and the estimated leasing cost per GPU is $2 per hour. These (V3 and R1) models could theoretically generate $562,027 in daily revenue. At this rate, the yearly revenue potential exceeds $200 million.
Such disclosure from DeepSeek suggests that the company is confident in its business model and wants to attract investors to compete with global AI firms. Compared to US rivals like OpenAI and Anthropic, which reportedly spend hundreds of millions on GPU resources, DeepSeek appears to have optimized its training process to be much cheaper. In fact, the DeepSeek-R1 model is said to be 20 to 50 times more cost-efficient than similar offerings from industry giants.
Interestingly, the company revealed its revenue projection a day after announcing a discounted pricing strategy for using its AI models through an API during off-peak hours. The discount applies between 12:30 AM and 8:30 AM (China time). Under this move, if developers or businesses use DeepSeek’s AI models during these hours (as mentioned above), they will have to pay significantly less amount. While the V3 model will be available at a 50% discount, the R1 reasoning model will receive a 75% discount. This becomes more significant as DeepSeek had over 22.15 million daily users worldwide as of January 2025.
