TAIPEI (Taiwan News) — An independent research and analysis company estimates that DeepSeek costs far more than advertised.
According to a report from the tech website Tom’s Hardware, DeepSeek may not be as disruptive as some believe and “there are no wonders” in AI. The report said DeepSeek had gained attention for its AI model, R1.
DeepSeek said that its training expenses were under US$6 million (NT$196 million) and required only 2,048 GPUs, while its latest R1 model can rival OpenAI’s o1. However, industry research firm SemiAnalysis disputes these claims, saying that the company behind DeepSeek has invested over NT$32.8 billion (US$1 billion) in server infrastructure.
The report said that DeepSeek operates an extensive computing infrastructure with around 50,000 Hopper GPUs, which include 10,000 H800 units, 10,000 H100 units, and additional purchases of H20 chips. These resources are distributed across multiple locations and are used for AI training, research, and financial modeling.
According to SemiAnalysis, the company's total investment in servers is approximately US$1.6 billion, with an estimated operating cost of US$944 million.
SemiAnalysis said that the US$6 million figure only accounts for the GPU cost of the pre-training run, which represents just a small fraction of the model’s total cost. It does not include critical expenses such as R&D and total cost of ownership of the hardware.
Far from what was being packaged as a “side project,” SemiAnalysis says DeepSeek took substantial, sustained effort. “We are confident that their GPU investments account for more than US$500 million, even after considering export controls,” said SemiAnalysis.
Tom's Hardware said that DeepSeek's rise demonstrates how a well-funded independent AI company can challenge industry leaders. However, it cautioned “the public discourse might have been driven by hype. Reality is more complex.”