China's DeepSeek reveals massive 'theoretical' margins
Make us a preferred source
Make us a preferred source on Google
The news: Chinese AI startup DeepSeek on Saturday disclosed some cost and revenue data for its V3 and R1 models, revealing its online service had a cost profit margin of 545% over a 24-hour period.
The numbers: The Hangzhou-based company said in a GitHub post that assuming the cost of renting one Nvidia H800 chip is USD2 ($3.2) per hour, the total daily inference cost for its models would be about USD87,072.
Using user statistics, its theoretical daily revenue is USD562,027, it said, amounting to just over USD200 million annually.
However, DeepSeek clarified its actual revenue was “substantially lower” because only some services are monetised, web and app access remain free, and developers pay less during off-peak hours.
Costs included in the calculations also did not include R&D and training expenses, Bloomberg noted. DeepSeek previously said it spent under USD6 million on chips to train its models, a small fraction compared to what US rivals spend.
The context: AI startups rarely disclose financial details, and DeepSeek's transparency, in contrast with US rivals like OpenAI, comes as investors struggle to quantify the return on investment for AI companies.
The sources: DeepSeek GitHub entry , Bloomberg
