DeepSeek Reports Impressive Daily Profitability
Chinese AI firm DeepSeek has reported a theoretical daily profit margin of 545% for its inference services. This figure comes despite offering some services for free and utilizing discounted pricing strategies. The company shared these details in a recent GitHub post, outlining the operational costs and potential revenue of its DeepSeek-V3 and R1 models.
Based on DeepSeek-R1’s pricing, the theoretical daily revenue generated clocks in at $562,027. The pricing structure involves charging $0.14 per million input tokens for cache hits, $0.55 per million for cache misses, and $2.19 per million output tokens. DeepSeek itself acknowledges that the real earnings are significantly lower, due to a combination of factors, including lower rates for DeepSeek-V3, free access to web and app services, and automatic nighttime discounts.
“Our pricing strategy prioritises accessibility and long-term adoption over immediate revenue maximisation,” DeepSeek stated in the GitHub post.
According to the company’s post, DeepSeek’s inference services utilize NVIDIA H800 GPUs. Matrix multiplications and dispatch transmissions function using the FP8 format, while core MLA computations and combine transmissions operate in BF16. The company adjusts GPU usage based on demand, deploying all nodes during peak hours and scaling them down at night to allocate computational resources for research and training.
The GitHub post revealed data from a 24-hour period, from February 27, 2025, to February 28, 2025. During this time, DeepSeek recorded peak node occupancy at 278, with an average of 226.75 nodes in operation. Considering that each node contains eight H800 GPUs and an estimated leasing cost of $2 per GPU per hour, the total daily expenditure reached nearly $87,072.
Market Reaction to DeepSeek’s Developments
The launch of DeepSeek’s R1 model, which reportedly required a $6 million training budget, resulted in a sharp market reaction. NVIDIA’s stock experienced a 17% tumble, erasing approximately $600 billion in value due to concerns about the model’s efficiency. However, during a recent earnings call, NVIDIA CEO Jensen Huang expressed optimism about the growing demand for inference, fueled by advanced scaling and new reasoning models.
“Models like OpenAI’s, Grok 3, and DeepSeek R1 are reasoning models that apply inference-time scaling. Reasoning models can consume 100 times more compute,” Huang said.
Huang also added, “DeepSeek-R1 has ignited global enthusiasm. It’s an excellent innovation. But even more importantly, it has open-sourced a world-class reasoning AI model.”
According to a recent report, DeepSeek is planning the release of its next reasoning model, the DeepSeek R2, ‘as early as possible.’ Initially, the model’s release was planned for early May; however, that timeline is subject to change. R2 is said to produce improved code and reason in languages beyond English.
