Recently, Chinese company DeepSeek threw the multi-billion dollar artificial intelligence industry into chaos by releasing its R1 model for just $5.576 million. However, analytics firm SemiAnalysis estimated that DeepSeek's true costs were well over $1 billion.
Recently, Chinese company DeepSeek threw the multi-billion dollar artificial intelligence industry into chaos by releasing its R1 model for just $5.576 million. However, analytics firm SemiAnalysis estimated that DeepSeek's true costs were well over $1 billion.
The announcement that DeepSeek was able to train R1 using only a fraction of the resources needed by large tech companies investing in AI sent Nvidia's stock price down by a record $600 billion in one day. If a Chinese startup could create such a powerful model without spending billions on the most powerful GPUs, what's to stop everyone else from doing the same?
But did DeepSeek really build its model, which still tops the Apple App Store charts, for such a low price? Analyst firm SemiAnalysis says no.
According to its analysis, DeepSeek has access to about 50,000 Hopper GPUs, including 10,000 H800s and 10,000 H100s. It also has orders for many more H20s destined for China. The GPUs are distributed between High-Flyer, the hedge fund behind DeepSeek, and the startup. They are spread across multiple geographies and are used for trading, inference, training, and research, TechSpot reports .
SemiAnalysis writes that DeepSeek has invested much more than the declared $5.5 million. This is the cost of pre-training the AI, which is only a fraction of the total. The company's total investment in servers is about $1.6 billion, of which about $944 million is spent on operating expenses. Investment in graphics processors, meanwhile, is more than $500 million.
DeepSeek is said to be recruiting all of its talent exclusively from China. This contrasts with reports that other Chinese tech companies, such as Huawei, are trying to lure workers from abroad, with Taiwanese TSMC workers being the most sought-after targets. DeepSeek is reportedly offering salaries of more than $1.3 million to promising candidates, far more than what rival Chinese AI companies are paying.
DeepSeek also has the advantage of largely using its own data centers rather than relying on external cloud providers. This allows for more experimentation and innovation in the AI product stack. SemiAnalysis writes that it is the best lab in this “open weight class” today, ahead of Meta’s Llama, Mistral, and others.
