Was the actual development cost of DeepSeek $1,6 billion?

Chinese startup DeepSeek has recently become the focus of attention in the tech world with its surprisingly low computing resource usage for its advanced AI model called R1. This model is touted as a potential competitor to Open AI's o1, even though the company claims DeepSeek only spent $6 million and 2.048 GPUs to train it.

However, industry analytics firm SemiAnalysis estimates that the company behind DeepSeek incurred $1,6 billion in hardware costs, with a "team" of 50.000 Nvidia Hopper GPUs. If this is confirmed, the assertion that DeepSeek can replicate AI inference training with significantly lower investment than industry leaders will have to change.

The analysis report suggests that DeepSeek operates a vast computing infrastructure comprising approximately 10.000 H800 GPUs and 10.000 H100 GPUs, in addition to H20 GPUs. This hardware is distributed across multiple locations and serves purposes such as AI training, research, and financial modeling. The company's total investment in servers is approximately $1,6 billion, with an estimated $944 million spent on operating costs, according to SemiAnalysis.

DeepSeek originated from High-Flyer, a Chinese hedge fund that adopted AI early and invested heavily in GPUs. In 2023, High-Flyer launched DeepSeek as a separate venture focused solely on AI. Unlike many competitors, DeepSeek remains self-funded, giving the company flexibility and speed in decision-making. Despite claiming to be a small arm, the company has invested over $500 million in its technology, according to SemiAnalysis.

The Chinese company also recruits talent from mainland China, with no poaching from elsewhere. According to SemiAnalysis, DeepSeek focuses on skills and problem-solving abilities rather than formal degrees, recruiting from Peking University and Zhejiang University, and offering very competitive salaries. Some AI researchers at DeepSeek are said to earn salaries exceeding those at other leading Chinese AI companies such as Moonshot.

DeepSeek emphasizes efficiency and algorithmic improvement rather than scaling, reshaping expectations surrounding AI model development. For various reasons, this approach has led some to believe that the Chinese company's rapid advancements could reduce demand for high-end GPUs, impacting companies like Nvidia.

The figure of $6 million only represents a portion of the total training costs, not including research costs, model refinement, data processing, or general infrastructure costs.