English 
Spanish 
3 min read
0 comments
New research has revealed that DeepSeek’s AI didn’t cost $6 million to train, but much more. Find out!
A recent analysis of the SemiAnalysis revealed that the actual cost of training DeepSeek's artificial intelligence (AI) is significantly higher than previously thought.
Although initial estimates suggested an investment of approximately $6M monthly, the report indicates that the real value reaches impressive $1,3 billion.
Debunking the $6 Million Myth
The initial estimate of $6M monthly considered only the expenses of pre-training on GPUs, neglecting substantial investments in research and development, infrastructure and other essential costs accumulated by the company.
- Revolution in education: free artificial intelligence courses that are transforming careers in Brazil
- Technological advances in healthcare: The role of artificial intelligence, RNA therapies and robotics in the evolution of diagnostics and treatments in 2025
- Startup Harumi Transforms Processes into Power for Large Companies with AI and Mathematics
- Why does Gen Z trust AI more than humans at work? Discover the reasons behind this preference!
The report highlights that DeepSeek’s total server capital expenditure (CapEx) amounts to approximately $1,6 billion, with a considerable portion of that going toward operating and maintaining its extensive GPU clusters.
Robust infrastructure and hardware investments
DeepSeek has access to approximately 50.000 Hopper series GPUs, including models such as the country-specific H800s, H100s, and H20s produced by NVIDIA in response to U.S. export restrictions.
This diversification in hardware inventory reflects the company's strategic sourcing decisions and operational efficiency.
Organizational structure and operational efficiency
Unlike some of the largest AI labs, DeepSeek operates its own data centers and adopts a streamlined model that contributes to its agility and efficiency. This ability to adapt quickly is vital in an increasingly competitive AI landscape.
In terms of performance, DeepSeek’s R1 model demonstrates reasoning capabilities comparable to OpenAI’s o1.
However, it is not considered the undisputed leader in all performance metrics. While DeepSeek's pricing strategy has received praise, it is worth noting that Google's Gemini Flash 2.0, with similar capabilities, proves to be even more cost-effective when accessed via API services.
This presents DeepSeek with the challenge of balancing performance and cost to ensure its future success.
One notable innovation highlighted in the report is Multi-Head Latent Attention (MLA) technology, which significantly reduces inference costs by an impressive 93,3% by reducing the use of key-value (KV) caches. This approach represents a major step forward towards more cost-effective AI solutions.
Experts suggest that DeepSeek's innovations are likely to be quickly adopted by Western AI labs looking to stay competitive.
The Future of Chinese AI
Although there is optimism regarding possible improvements and efficiency gains, SemiAnalysis warns of external challenges.
The report speculates that operating costs could fall further, driven by DeepSeek's ability to adapt quickly compared to its larger, more bureaucratic counterparts.
However, scaling up operations amid increased U.S. export controls presents a significant hurdle that DeepSeek must navigate with caution.
With information from SemiAnalysis.
