Chinese AI startup DeepSeek has made waves in the AI industry by revealing a theoretical daily profit margin of 545% for its inference services. However, the company admits that real earnings are much lower due to discounted pricing, free web and app access, and nighttime cost reductions.
Breaking Down DeepSeek’s AI Business Model
In a recent GitHub post, DeepSeek shared insights into the operational costs and revenue potential of its DeepSeek-V3 and DeepSeek-R1 models. Based on the pricing structure of DeepSeek-R1, the revenue breakdown per million tokens is:
- $0.14 per million input tokens (cache hits)
- $0.55 per million input tokens (cache misses)
- $2.19 per million output tokens
Using these rates, DeepSeek’s theoretical daily revenue stands at $562,027. However, the company stated that actual revenue is much lower due to factors like free access for some users, deep discounts, and a focus on long-term adoption over immediate profits.
High-Cost Operations: Running on NVIDIA H800 GPUs
DeepSeek’s AI inference services rely on NVIDIA H800 GPUs, a high-performance AI computing solution. The company optimizes costs by:
- Scaling GPU usage based on demand (all nodes running during peak hours, reduced usage at night for research and training).
- Using FP8 format for matrix multiplications and dispatch transmissions while employing BF16 for MLA computations and combined transmissions.
Between February 27, 2025, and 12:00 PM on February 28, 2025, DeepSeek recorded:
- 278 peak nodes in operation.
- 226.75 average nodes in use.
- Each node containing 8 NVIDIA H800 GPUs.
- An estimated leasing cost of $2 per GPU per hour, leading to total daily operational expenses of $87,072.
DeepSeek’s Impact on the Stock Market & NVIDIA’s Response
DeepSeek’s latest AI model, DeepSeek-R1, has not only drawn industry attention but also triggered a sharp reaction in the US stock market. Following its launch, NVIDIA’s stock fell by 17%, wiping out nearly $600 billion in market value. Investors expressed concerns over the efficiency of DeepSeek’s models, which could impact the demand for NVIDIA’s high-end chips.
However, NVIDIA CEO Jensen Huang addressed these concerns during the company’s recent earnings call. Huang noted that inference demand is accelerating, driven by test-time scaling and advanced reasoning models.
“Models like OpenAI’s Grok 3 and DeepSeek-R1 are reasoning models that apply inference-time scaling,” Huang explained. “Reasoning models can consume 100 times more compute, which continues to drive GPU demand.”
Despite the stock market drop, Huang praised DeepSeek-R1, calling it a world-class open-source reasoning AI model that has ignited global enthusiasm.
What’s Next? DeepSeek-R2 on the Horizon
DeepSeek has hinted at the early release of its next-generation AI model, DeepSeek-R2. Originally slated for May 2025, the company is now considering launching it ahead of schedule.
According to DeepSeek, the upcoming model will feature:
- More advanced reasoning capabilities.
- Improved coding performance.
- Support for multiple languages beyond English.
Analysis & Fact-Check
- 545% Profit Margin vs. Real Earnings
- The reported 545% theoretical profit margin is based on ideal revenue scenarios. However, DeepSeek itself admits that real earnings are much lower due to its pricing strategy, which prioritizes adoption over short-term revenue.
- Stock Market Impact on NVIDIA
- NVIDIA’s 17% stock decline following DeepSeek-R1’s launch suggests that investors perceive potential competition from efficient AI models. However, NVIDIA’s assertion that reasoning models drive higher GPU consumption could mitigate long-term concerns.
- DeepSeek’s R&D Efficiency
- DeepSeek claims it trained DeepSeek-R1 on a $6 million budget, which is remarkably low compared to OpenAI’s GPT-4 training costs (estimated to be in the hundreds of millions). If these figures hold, DeepSeek’s efficiency could reshape the AI industry.
DeepSeek is rapidly emerging as a major AI player, with its open-source reasoning models and efficient GPU usage strategy disrupting industry norms. While its theoretical profit margins appear impressive, the company’s pricing approach prioritizes long-term market adoption over immediate financial gain.
The broader AI market will closely watch DeepSeek-R2’s release, as its performance could influence global AI development strategies and further impact AI chipmakers like NVIDIA. Whether DeepSeek’s rapid expansion will pose a serious challenge to established players remains to be seen, but for now, it has undeniably made its mark.
