DeepSeek's AI training cost claims exaggerated, says Google's AI chief
What's the story
Demis Hassabis, the head of Google's artificial intelligence (AI) division DeepMind, has taken a dig at the reported development costs of DeepSeek's AI model. In an interview with Bloomberg Television, Hassabis hinted that the claim of spending less than $6 million on its AI system is "exaggerated and a little bit misleading." This comes after DeepSeek made global headlines last month by launching a widely-used chatbot and AI model at a much lower cost than its American counterparts.
Cost scrutiny
Hassabis questions DeepSeek's cost reporting
Hassabis was skeptical about DeepSeek's cost reporting, saying the company "seems to have only reported the cost of the final training round, which is a fraction of the total cost." He further dismissed suggestions that DeepSeek's emergence could disrupt the economics of AI development. At the AI Action Summit in Paris, he said, "We don't see any new silver bullet technologies," and "DeepSeek is not an outlier on the efficiency curve."
Cost investigation
DeepSeek's computing costs and chip usage under scrutiny
DeepSeek claimed to have spent $5.6 million on computing costs to train its model with older NVIDIA Corp. chips, a claim several researchers have doubted. To address these doubts, US authorities have initiated an investigation into whether DeepSeek circumvented a chip ban by purchasing them through Singapore. Bloomberg also reported that OpenAI and Microsoft Corp. are investigating if a DeepSeek-linked group accessed OpenAI's data through distillation - a process where one AI model uses another's outputs for training.
Model efficiency
Hassabis highlights Google's AI model efficiency
Despite the market unrest triggered by DeepSeek, several Silicon Valley giants, including Alphabet, have stuck to their massive spending plans. Just last week, Alphabet announced plans for $75 billion in capital expenditures in 2025. The funding will go into its cloud-computing division and services like Gemini - an AI model Google is embedding into search and other products. "Gemini is more efficient than DeepSeek in terms of its training to performance or cost to performance," Hassabis said, stressing Google's unwavering commitment to AI development.