Chinese DeepSeek, which made a stunning debut with its chatbot in January, claims to have spent only $6 million on training its artificial intelligence system. However, SemiAnalysis researchers found that the announced amount is greatly underestimated – the real costs are $1.3 billion.
As Ineteresting Ingineering writes, their analysis showed that the figure of $6 million only takes into account the basic costs of preliminary training on graphics processors. DeepSeek's total capital expenditure on server infrastructure – which analysts estimated at $1.3 billion. – includes:
- investments in research and development,
- infrastructure costs,
- operation and maintenance of equipment,
- technical infrastructure.
DeepSeek has access to about 50,000 Hopper GPUs, which has caused some confusion in the industry. SemiAnalysis explains: this does not mean the presence of 50,000 H100s, as previously believed. Instead, the set includes various models, including the H800, H100, and H20 for specific countries, released by NVIDIA in response to US export restrictions.
Another aspect that analysts focused on is DeepSeek's unique organizational structure. Unlike other large AI labs, the Chinese startup independently manages its data centers. This model provides greater flexibility and efficiency, which is becoming critical in an increasingly competitive market.