- Joined
- Aug 20, 2022
- Messages
- 16,156
- Points
- 113
DeepSeek Could Have 50,000 NVIDIA Hopper GPUs, Spent US$1.6 Billion Building Servers
US$6 million may be a fraction of the company's total training cost.
BY JOHN LAWFEBRUARY 4, 2025
Image: Bloomberg
Chinese AI startup Deepseek took the tech world by storm last week, after it debuted a low-cost AI model that it claimed was trained on a budget of US$6 million (~RM26.79 million), and cause the stock value of hardware tech giant to plunge by more than US$600 billion (~RM2.68 trillion). However, new information suggests that its training cost might not be the full story.
According to SemiAnalysis, an industry analyst firm, the parent company behind DeepSeek had incurred US$1.6 billion (~RM7.14 billion) in hardware costs and supposedly has a massive stock of 50,000 NVIDIA Hopper GPUs. The inventory includes 10,000 H800 and 10,000 H100 GPUs, plus additional purchases of H20 units.
Given the massive number of NVIDIA GPUs in its arsenal, SemiAnalysis believes that DeepSeek’s alleged US$6 million training costs for its latest V3 Mixture-of-Experts (Moe) AI model is itself just a portion of the actual total training costs. That amount is estimated to be more than US$500 million (~RM2.23 billion), and that the US$6 million does not account for research, model refinement, data processing, and infrastructure expenses.
A more important point that is being investigated is how exactly DeekSeek, a Chinese company, managed to get their hands on so many NVIDIA GPUs, despite the current trade restrictions in place against China. Multiple reports point to the GPUs being purchased via third parties in Singapore, prompting the country’s Ministry of Trade and Industry (MTI) to launch an investigation.
(Source: SemiAnalysis, Tom’s Hardware, CNBC)
