Amid attention to the low-cost artificial intelligence (AI) model from the Chinese startup DeepSeek, analyses indicate that the actual development expense could be ten times more than initially reported. DeepSeek has stated it spent approximately $5.6 million (about 810 million won) to develop the AI model.
On 1st (local time), U.S. economic media CNBC reported that the semiconductor research and consulting firm SemiAnalysis estimated that the hardware expenditure required for DeepSeek's AI model development would "far exceed the $500 million (about 730 billion won) already invested." The hardware expenditure refers to costs for AI chips, servers, and other necessities for running the AI model.
This amounts to about 90 times the total training expense originally claimed by DeepSeek. In its report on the latest AI models, DeepSeek calculated the expense of renting NVIDIA's inexpensive AI chip, the 'H800 graphics processing unit (GPU),' at $2 per hour for two months, resulting in an expense of $5.576 million (approximately 813 million won).
SemiAnalysis noted that "considerable expense goes into research and development and operations and maintenance, and generating 'synthetic data' for training the AI model requires enormous computing resources."
Furthermore, SemiAnalysis mentioned in its report that the training expense for Anthropic's latest AI model, Claude 3.5, one of OpenAI's main competitors, "amounts to tens of millions of dollars," explaining that "Anthropic has raised billions of dollars in investments from Amazon and Google, which illustrates how much funding is needed for operating AI models and corporate operations."
In response, it was explained that "various expenses arise due to numerous experiments, new architecture development, data collection, and employee salaries required to develop the AI model."
SemiAnalysis further stated, "What is clear is that DeepSeek is unprecedented in achieving such a level of expense and performance," adding that "DeepSeek's R1 model is exceptional, and reaching this degree of logical reasoning capability so quickly is objectively impressive." R1 is DeepSeek's most recently released reasoning model. DeepSeek has claimed that the R1 surpasses OpenAI's 'O1' (Ohwon), which was launched in September last year, in several AI model tests.