DeepSeek didn't train its model for $6 million. A new semiconductor report shows the company spent over $1 billion on its compute cluster. The $6 million figure conflates capital costs with R&D and likely covers only the final training run, not the full investment.
DeepSeek's $6 million training headline was fake marketing. The headline was, oh wow, this company spent $6 million and it's basically at par with OpenAI's O1 model, which is their best one right now. There's a new report by leading semiconductor analysts. So DeepSeek actually spent over $1 billion on its compute cluster. The widely reported $6 million number is highly misleading as it includes CapEx and R&D, which might be a part of that Singapore piece, right? At best, this describes the cost of the final training run only.
