Chinese AI startup DeepSeek has made a bold claim about the profitability of its AI models, albeit with certain caveats.

In a post on X, DeepSeek revealed that its online services boast a theoretical “cost profit margin” of 545%. However, this margin is based on hypothetical income, rather than actual earnings.

The company provided more detailed information about these figures in a lengthy GitHub post that outlined its strategy for achieving higher throughput and lower latency. According to the post, if the usage of its V3 and R1 models during a 24-hour period were billed using R1 pricing, DeepSeek’s daily revenue would be $562,027.

In contrast, the cost of leasing the necessary graphics processing units (GPUs) would be a mere $87,072.

DeepSeek conceded that its actual revenue is significantly lower due to various factors, such as nighttime discounts, lower pricing for V3, and the fact that only a subset of its services are monetized, with web and app access remaining free.

It is worth noting that if the app and website were not free and discounts were not available, usage would likely be substantially lower. As such, these calculations appear to be speculative and are more indicative of potential future profit margins rather than a reflection of DeepSeek’s current financial situation.

DeepSeek’s disclosure of these figures comes amidst ongoing debates about the cost and potential profitability of AI. The company gained significant attention in January with the release of a new model that allegedly matched OpenAI’s o1 on certain benchmarks, despite being developed at a lower cost and in the face of US trade restrictions that limit Chinese companies’ access to high-performance chips. This development led to a decline in tech stocks and raised questions among analysts about AI spending.

DeepSeek’s technology has not only impacted the financial markets but also briefly displaced OpenAI’s ChatGPT as the top app in Apple’s App Store, although it has since dropped in the rankings and is currently ranked #6 in productivity, behind ChatGPT, Grok, and Google Gemini.