DeepSeek, the Chinese AI startup, recently sparked controversy in the AI industry by claiming that its R1 model was trained for just $5.5 million.
However, a new report by SemiAnalysis suggests the actual costs are significantly higher—around $1.6 billion—and that DeepSeek has access to 50,000 Nvidia GPUs.
The $5.5 Million Claim vs. Reality
DeepSeek’s initial claim about the costs of developing its Mixture-of-Experts model sent Nvidia’s stock plummeting by $600 billion.
The $5.5 million figure seemed too low compared to the vast investments by major AI companies.
According to SemiAnalysis, the true investment includes $944 million in operating costs and over $500 million on GPUs alone.
DeepSeek’s GPU Power
The firm has access to a huge array of Hopper GPUs, including 10,000 H800s and 10,000 H100s, along with H20s for Chinese-specific operations.
These GPUs are shared with High-Flyer, a quantitative hedge fund backing DeepSeek, distributed across multiple locations for trading, research, and training.
Talent and Data Centers
DeepSeek exclusively hires talent from China and offers salaries up to $1.3 million for top candidates.
Unlike other tech giants, DeepSeek runs its own data centers, giving it more control and flexibility for AI development.
