Google AI chief says DeepSeek’s cost claims are ‘exaggerated’

Last month, DeepSeek shocked the world by releasing a popular chatbot and AI model using far less money than US rivals, such as DeepMind and OpenAI. Hassabis, who runs the AI unit of Alphabet Inc.’s Google, told Bloomberg Television that DeepSeek “seems to have only reported the cost of the final training round, which is a fraction of the total cost.”

The executive also shot down the idea that DeepSeek’s emergence upends the economics of AI development. “We don’t see any new silver bullet technologies,” Hassabis said on Monday, in Paris, at the Artificial Intelligence Action Summit. “DeepSeek is not an outlier on the efficiency curve.”

The Chinese startup reported spending $5.6 million on computing costs to train its model using older Nvidia Corp. Chips. Several researchers have questioned those claims. US authorities opened a probe to see if DeepSeek circumvented a chip ban by purchasing through Singapore. 

Bloomberg News reported earlier that OpenAI and Microsoft Corp. Are investigating whether a group tied to DeepSeek obtained data from OpenAI using a process known as distillation, where one AI model pilfers outputs from another for training purposes.

Hassabis said DeepSeek seems “to have relied on some Western models to distill from,” without offering specifics. 

A representative for DeepSeek didn’t immediately respond to a request for comment.

Alphabet is one of several Silicon Valley giants that stuck with major spending plans after DeepSeek caused turmoil in public markets. Last week, Alphabet reported plans for $75 billion in capital expenditures in 2025. That money will go to its cloud-computing division and services like Gemini, an AI model Google is weaving into search and other products.

“Gemini is more efficient than DeepSeek in terms of its training to performance or cost to performance,” Hassabis said. “We just don’t talk about it very much.”