DeepSeek 'lied' about the cost of developing AI chatbot?

According to DeepSeek, they spent only $6 million to train the DeepSeek V3 neural network and used 2,048 Nvidia graphics processing units (GPUs). However, a report from SemiAnalysis shows that DeepSeek owns up to 50,000 Nvidia GPUs and has invested at least $1.6 billion in infrastructure.

Did DeepSeek cost around $1.6 billion to develop instead of the claimed $6 million?

Experts at SemiAnalysis said DeepSeek operates a large-scale computing system consisting of about 50,000 Hopper GPUs, including 10,000 H800 GPUs and 10,000 more powerful H100 GPUs, along with additional batches of H20 GPUs. These resources are distributed across multiple data centers and are used for AI training, scientific research, and financial modeling. The total capital expenditure for the servers is estimated at $1.6 billion, while operating expenses are up to $944 million, according to a report by Tom's Hardware.

DeepSeek originated from Chinese hedge fund High-Flyer but was spun off into an independent AI-focused company in 2023. Unlike many startups that typically outsource capacity to cloud providers, DeepSeek owns its own data centers, giving it control over the optimization of its AI models and accelerating innovation. Furthermore, being fully self-funded, the company is able to make decisions more quickly and flexibly than traditional competitors.

Estimates provided by SemiAnalysis.

DeepSeek's human resources policy is also notable, as the company only hires experts from China, focusing on skills and the ability to solve complex problems rather than formal qualifications. The salaries of some researchers here can exceed $1.3 million per year, helping them attract talent from top universities.

While DeepSeek positions itself as an innovator capable of challenging industry giants, analysts at SemiAnalysis warn that the company's success depends on huge investments, technological breakthroughs and a strong team of employees, however claims of a 'revolutionary budget' for DeepSeek's AI model appear to be exaggerated.

David Pac

Update 10 February 2025