DeepSeek Reveals Theoretical 545% Daily Profit Potential for AI Models

Chinese artificial intelligence startup DeepSeek has disclosed detailed cost and revenue figures for its popular V3 and R1 models, claiming a theoretical cost-profit ratio of up to 545% per day. However, the company emphasized that actual revenues would be significantly lower due to various operational factors.

According to the disclosure, DeepSeek rents Nvidia H800 chips at roughly $2 per hour. Based on this cost, the daily inference expense; covering the phase where AI models perform tasks like powering chatbots, is estimated at about $87,072. In contrast, the models’ theoretical daily revenue could reach approximately $562,027, translating to a striking 545% profit margin. If these figures held true over a full year, the annual revenue could potentially exceed $200 million.

Despite the impressive theoretical numbers, DeepSeek warned that real-world earnings would be much lower. The company noted that several factors contribute to this discrepancy, including:

• Variations in Model Costs: The V3 model operates at lower costs compared to R1.
• Access Provisions: Some services offer free web and app access, which affects revenue.
• Developer Fees: Reduced charges during off-peak hours further lower potential income.

This financial snapshot marks the first time the Hangzhou-based company has publicly detailed its profit margins derived from less computationally intensive inference tasks, shedding light on an area that has long been a point of interest in the AI industry.

The announcement comes at a time when the global AI sector is closely examining cost efficiency and hardware investment strategies. Earlier in the year, the popularity of DeepSeek’s R1 and V3 models was linked to a decline in AI stock values outside China. Investors have also raised questions about the company’s reliance on Nvidia’s H800 chips, which are considered less powerful than the more advanced hardware employed by some U.S. competitors.

DeepSeek’s disclosure offers valuable insights into the financial dynamics of AI inference tasks and may prompt further discussion on balancing operational costs with revenue potential in the competitive AI landscape.