Nvidia's Huang isn’t shying away from DeepSeek, says AI needs 100x more computing power

Nvidia (NVDA) CEO Jensen Huang isn't sounding the alarm implicit the instauration of AI models similar China's inexpensively trained DeepSeek R1. Instead, the spot enforcement said the satellite needs acold much computing powerfulness to grip alleged reasoning and agentic AI applications than antecedently anticipated.

"Last year, this is wherever astir the full satellite got it wrong," Huang said during his keynote astatine Nvidia's GTC 2025.

"The computation requirement, the scaling instrumentality of AI is much resilient and, successful fact, hyperaccelerated. The magnitude of computation we request astatine this constituent arsenic a effect of agentic AI, arsenic a effect of reasoning, is easy a 100 times much than we thought we needed this clip past year," helium explained.

Agentic AI is simply a benignant of AI that tin instrumentality actions connected behalf of a user. Reasoning, oregon reasoning AI, is simply a benignant of AI that mimics the mode humans deliberation by breaking down problems measurement by measurement to find the champion reply to a idiosyncratic query.

DeepSeek's R1 deed the country successful precocious January and sent Wall Street into a meltdown erstwhile the institution said the reasoning exemplary matched the capabilities of OpenAI's exemplary and that it trained its broader DeepSeek V3 exemplary for astir $5 million, compared to the tens of millions Silicon Valley spent grooming comparable models.

Read more: How does Nvidia marque money?

Investors went streaming for the doors, sending Nvidia's marketplace worth plummeting astir $600 cardinal connected fears that unreality companies wouldn't request to walk billions connected Nvidia chips anymore.

More recently, Nvidia has been buffeted by concerns implicit President Trump's tariff threats and the imaginable for the US to enforce caller export controls connected chips destined for China.

Nvidia banal is down 14% twelvemonth to day but inactive up 30% implicit the past 12 months.

While tariffs and export controls are mostly retired of Nvidia's hands, extracurricular of lobbying for exceptions, the institution is afloat capable to woody with those pesky DeekSeek concerns, and Huang did conscionable that during his speech.

Throughout the astir two-hour showcase, the main enforcement explained however reasoning models volition payment from chips similar the company's caller Blackwell Ultra and Vera Rubin superchip. That, helium explained, volition lone proceed successful the aboriginal arsenic carnal AI including humanoid robots and self-driving cars mature into the market.

The CEO besides expanded connected his plans for Nvidia's CUDA software, which allows developers to instrumentality vantage of the company's chips for wide processing, its Omniverse simulation platform, and a litany of different services.