DeepSeek Claims "Theoretically" 545% Daily Cost-Profit Margin, but Reality May Be Lower

DeepSeek has made a significant financial and technological impact on the Western AI market. The company recently disclosed its financial data for the first time, which could further influence the global AI stock market. DeepSeek shared financial insights related to its AI models, DeepSeek-V3 and DeepSeek-R1, emphasizing the potential profitability of running these models, particularly in inference tasks. Notably, inference tasks refer to AI generating responses rather than undergoing training.

According to DeepSeek, due to lower Nvidia GPU prices and reduced task requirements, its AI models could achieve an impressive daily cost-to-profit ratio of up to 545%. This means that for every dollar spent on operating the models, the company might generate $5.45 in revenue.

However, this figure is based on ideal conditions, assuming high demand and optimal efficiency. In reality, DeepSeek, the two-year-old Hangzhou-based startup, clarified that actual earnings could be much lower due to various factors such as server costs, electricity expenses, fluctuations in AI service demand, pricing strategies, and customer acquisition expenses.

In a recent GitHub post, DeepSeek revealed additional details. Based on company data, the daily operational cost for running DeepSeek V3 and R1 models stands at $87,072. Over a 24-hour period (from February 27 to February 28, 2025), DeepSeek recorded a peak node usage of 278, with an average of 226.75 nodes running. Each node is equipped with eight H800 GPUs, with an estimated rental cost of $2 per GPU per hour. These models (V3 and R1) could theoretically generate daily revenues of $562,027, translating into an annual revenue potential exceeding $200 million.

This disclosure by DeepSeek highlights the company's confidence in its business model and its ambition to attract investors while competing with global AI firms. Compared to U.S. rivals like OpenAI and Anthropic, which reportedly spend hundreds of millions on GPU resources, DeepSeek appears to have optimized its training processes to significantly lower costs. In fact, the DeepSeek-R1 model is said to be 20 to 50 times more cost-efficient than similar offerings from industry leaders.

Interestingly, just a day after announcing discounted pricing strategies for off-peak API usage of its AI models, the company unveiled its revenue projections. Discounts apply during non-peak hours between 12:30 AM and 8:30 AM China Time. Under this initiative, developers or businesses using DeepSeek's AI models during these periods will pay significantly less. The V3 model will receive a 50% discount, while the R1 inference model will enjoy a 75% discount. This move becomes even more crucial given that, as of January 2025, DeepSeek boasts over 22.15 million daily active users globally.