Google AI Chief Questions DeepSeek’s $5.6M Training Claim as ‘Exaggerated’

The AI world has been buzzing with DeepSeek’s claim that it trained its R1 model—rivaling OpenAI’s GPT-4—for just $5.6 million. This figure is a stark contrast to the $100+ million reportedly spent by OpenAI on GPT-4, and it sent shockwaves through the tech industry. However, at the Artificial Intelligence Action Summit in Paris on Monday, Google DeepMind CEO Demis Hassabis cast doubt on DeepSeek’s bold assertion, calling it “exaggerated” and “misleading.”

A Powerful Model with Questionable Claims

While Hassabis acknowledged DeepSeek’s technical achievements, he questioned the legitimacy of its cost-efficiency claims.

“It’s a very impressive model, very impressive piece of work, and I think the team is probably the best team I’ve seen come out of China,” Hassabis told Bloomberg. “That said, I think a lot of the claims are exaggerated and a little bit misleading.”

Hassabis argued that the $5.6 million figure likely represents only the cost of DeepSeek’s final training run, which is just a fraction of the total cost associated with developing an AI model from scratch. The overall expense would also include data acquisition, model architecture development, multiple training iterations, and fine-tuning.

Accusations of Model Distillation

Hassabis also suggested that DeepSeek might have relied heavily on Western AI models to distill from during development, a claim echoed by OpenAI.

“We know PRC-based companies—and others—are constantly trying to distill the models of leading US AI companies,” OpenAI stated to Bloomberg in the wake of DeepSeek’s release.

Model distillation involves using existing models to inform or refine new ones, potentially reducing both the time and resources needed to train from scratch. If DeepSeek did employ this strategy, it could explain the dramatically lower costs without representing a true breakthrough in cost efficiency.

How Does DeepSeek Compare to Google’s AI?

Despite DeepSeek’s impressive capabilities, Hassabis was clear that Google doesn’t view it as a revolutionary outlier. He compared DeepSeek’s efficiency to Google’s own Gemini model, claiming that Gemini offers better performance relative to its training and cost.

“It’s impressive, but it isn’t some new outlier on the efficiency curve,” Hassabis said. “For example, Gemini is more efficient than DeepSeek in terms of its training to performance or its cost to performance. We just don’t talk about that very much.”

Hassabis’s comments highlight a broader point in the AI industry: cost alone doesn’t tell the full story. Efficiency, scalability, and the quality of the model’s outputs are just as important as how much it costs to train.

Looking Ahead: DeepSeek’s Place in the AI Landscape

Whether or not DeepSeek’s $5.6 million training claim holds up under scrutiny, the model’s emergence signals China’s growing influence in the global AI race. Even if DeepSeek relied on Western models or disclosed only partial costs, its ability to produce a model that rivals OpenAI’s outputs is noteworthy.

For companies like Google and OpenAI, DeepSeek’s rise serves as both competition and motivation. As the AI landscape becomes increasingly international, efficiency claims will continue to be tested—not just in terms of cost, but also in transparency, innovation, and real-world performance.

You can watch Demis Hassabis’s full remarks from the AI Summit below.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing, image, and idea-generation support from ChatGPT, an AI assistant. However, the final perspective and editorial choices are solely Alicia Shapiro’s. Special thanks to ChatGPT for assistance with research and editorial support in crafting this article.