During the Paris Artificial Intelligence (AI) Global Summit on Monday, February 10th, a senior Google AI executive stated that many claims about the AI model recently released by the Chinese startup DeepSeek were “exaggerated and even misleading”.
During the summit, Demis Hassabis, the CEO and co-founder of DeepMind, Google’s AI research lab, told Bloomberg Technology in an interview that “When DeepSeek reported their training costs, they seemed to only report the cost of the final round of training. However, in reality, significant resources are needed for exploration, testing, and training before the final training. Therefore, we believe they underestimated the overall cost.”
Last month, DeepSeek released an AI model and claimed that the training costs were much lower than those of American competitors like DeepMind and OpenAI, shocking the world. However, it quickly sparked controversy and was rapidly banned by multiple countries due to serious privacy and security issues.
Hassabis pointed out that DeepSeek “seems to rely on some Western models for distillation, or basically fine-tuning AI by adjusting the outputs of other models.”
He mentioned that the DeepSeek team is the “most outstanding team from China” that he has seen, “but we haven’t seen any disruptive technological breakthroughs, nor have we seen any new technologies or inventions that we haven’t seen before.”
DeepSeek’s representatives did not immediately respond to Bloomberg’s request for comment.
The Chinese startup claims that training the model using older Nvidia chips cost $5.6 million. Many researchers have questioned this claim.
US authorities have launched an investigation to determine whether DeepSeek bypassed sanctions by purchasing Nvidia chips through Singapore.
Furthermore, in a research report released by the semiconductor research and consulting firm SemiAnalysis on January 31st, it was stated that DeepSeek spent over $500 million on GPU hardware alone, with total server capital expenditure of approximately $1.6 billion, of which up to $944 million was used to operate its chip clusters.