DeepSeek’s Hardware Cost $500 million
SemiAnalysis, a semiconductor research and consulting firm, recently released a report that provided additional context for DeepSeek’s costs.
The firm calculated that DeepSeek has spent “well over $500 million on hardware over the company’s history,” noting that total cost of ownership and R&D expenses are high. It would take “a considerable amount of computation” to create “synthetic data” for the model to train on, according to SemiAnalysis.
Anthropic raised billions of dollars from Amazon and Google, which shows how much more money is needed to run the models and the business, despite the report stating that the Claude 3.5 Sonnet from Anthropic cost “tens of millions to train.”
SemiAnalysis explained they needed to pay staff, collect and clean data, develop new architectures, and experiment.
The compute costs for DeepSeek are not estimated in the paper. The analysts stated that DeepSeek is unique since they were the first to attain such cost and capability. DeepSeek’s R1 “is a good model,” the company continued, adding that “catching up to the reasoning edge this quickly is objectively impressive.”
DeepSeek became the most talked-about topic in the tech industry, with many Wall Street and industry insiders concentrating on just one figure: $6 million.
According to DeepSeek’s report on its most recent AI model, the company’s entire training expenses, which were calculated using the rental price of Nvidia’s GPUs, came to $5.576 million.
DeepSeek stated that the figure only included the model’s “official training” and did not include the expenses associated with “prior research and ablation experiments on architectures, algorithms, or data.”.
ABOUT THE AUTHOR
See More
Join 350 000+ traders receiving Free Trading Signals
Alerts
Rates
Calendar
Signals
Top Broker