Nvidia (NVDA) CEO Jensen Huang isn’t sounding the alarm over the introduction of AI models like China’s inexpensively trained DeepSeek R1. Instead, the chip executive said the world needs far more computing power to handle so-called reasoning and agentic AI applications than previously anticipated.
“Last year, this is where almost the entire world got it wrong,” Huang said during his keynote at Nvidia’s GTC 2025.
“The computation requirement, the scaling law of AI is more resilient and, in fact, hyperaccelerated. The amount of computation we need at this point as a result of agentic AI, as a result of reasoning, is easily a hundred times more than we thought we needed this time last year,” he explained.
Agentic AI is a type of AI that can take actions on behalf of a user. Reasoning, or thinking AI, is a type of AI that mimics the way humans think by breaking down problems step by step to find the best answer to a user query.
DeepSeek’s R1 hit the scene in late January and sent Wall Street into a meltdown when the company said the reasoning model matched the capabilities of OpenAI’s model and that it trained its broader DeepSeek V3 model for roughly $5 million, compared to the tens of millions Silicon Valley spent training comparable models.
Read more: How does Nvidia make money?
Investors went streaming for the doors, sending Nvidia’s market value plummeting nearly $600 billion on fears that cloud companies wouldn’t need to spend billions on Nvidia chips anymore.
More recently, Nvidia has been buffeted by concerns over President Trump’s tariff threats and the potential for the US to impose fresh export controls on chips destined for China.
Nvidia stock is down 14% year to date but still up 30% over the past 12 months.
While tariffs and export controls are largely out of Nvidia’s hands, outside of lobbying for exceptions, the company is fully able to deal with those pesky DeekSeek concerns, and Huang did just that during his speech.
Throughout the roughly two-hour showcase, the chief executive explained how reasoning models will benefit from chips like the company’s new Blackwell Ultra and Vera Rubin superchip. That, he explained, will only continue in the future as physical AI including humanoid robots and self-driving cars mature into the market.
The CEO also expanded on his plans for Nvidia’s CUDA software, which allows developers to take advantage of the company’s chips for general processing, its Omniverse simulation platform, and a litany of other services.
