This marks the first time the Hangzhou-based firm has shared details about the profitability of its inference operations—the phase after AI model training where models generate predictions or perform tasks, such as powering chatbots. The announcement could further unsettle global AI stocks, which experienced a sharp decline in January following the worldwide surge in popularity of DeepSeek’s R1 and V3-powered chatbots.
A Cost-Effective Approach to AI
DeepSeek’s revelation comes amid growing scrutiny of the massive expenditures by U.S. AI firms like OpenAI. The company previously stated that it spent less than $6 million on chips to train its models, a fraction of the budgets of its American counterparts. DeepSeek utilized Nvidia’s H800 chips, which are less powerful than the cutting-edge hardware available to U.S. firms, raising questions about the efficiency and cost-effectiveness of the AI industry outside China.
In a GitHub post published on Saturday, DeepSeek outlined its cost-profit calculations. Assuming a rental cost of $2 per hour for one H800 chip, the daily inference cost for its V3 and R1 models amounts to $87,072. In contrast, the theoretical daily revenue generated by these models is $562,027, resulting in a staggering cost-profit ratio of 545%. Over a year, this would translate to just over $200 million in revenue.
Real-World Revenue Constraints
However, DeepSeek emphasized that its actual revenue is substantially lower than the theoretical figure. Several factors contribute to this discrepancy: the cost of using the V3 model is lower than the R1 model, many services remain free for web and app users, and developers pay reduced rates during off-peak hours.
The company’s disclosure highlights the growing competitiveness of China’s AI sector, which is increasingly challenging the dominance of U.S. firms. By leveraging cost-efficient hardware and innovative monetization strategies, DeepSeek is positioning itself as a formidable player in the global AI landscape.