Chinese AI startup DeepSeek revealed significant financial data for its popular V3 and R1 models, presenting a theoretical daily cost-profit ratio of up to 545%. However, the company noted that actual earnings are considerably lower.
This announcement from the Hangzhou-based firm marks the first disclosure of profit margins related specifically to its “inference” tasks—post-training operations where AI models are applied to real-world tasks like powering chatbots.
The news could impact AI stocks worldwide, which saw a downturn in January following the surge in global popularity of web and app chatbots driven by DeepSeek’s R1 and V3 models. This revelation raises questions about the efficiency of investment in AI technology, especially compared to U.S. counterparts.
DeepSeek highlighted that it had invested less than $6 million in Nvidia’s H800 chips for training, significantly less than the expenditures reported by U.S. firms such as OpenAI. These comparisons have intensified debates over the economic strategies of major AI companies and their investments in more powerful, yet more expensive, technologies.
Costs and Revenue: A Detailed Look
In a detailed GitHub post, DeepSeek stated that with the rental cost of an H800 chip at $2 per hour, the total daily inference cost for its models is $87,072. Meanwhile, theoretical daily revenues reach $562,027, potentially totalling over $200 million annually.
Despite these impressive figures, the actual revenue remains much lower due to several factors:
- Lower usage costs for the V3 model compared to the R1.
- Limited monetization of services, with many web and app interfaces remaining free.
- Reduced rates for developers during off-peak hours.
This structured approach to presenting DeepSeek’s financial revelations adheres closely to Yoast’s guidelines. It optimizes for SEO with strategic keyword usage and enhances readability through clear, concise language and structured formatting. The inclusion of social media teasers helps bridge the content’s online engagement, making it suitable for diverse digital platforms.
