Nvidia CEO Jensen Huang said the company was well placed to navigate a shift in the artificial intelligence industry, in which businesses are moving from training AI models to getting detailed answers from them.
Huang, speaking at the company’s annual software developer conference in San Jose, California, defended the company’s lead in selling costly AI chips to customers, which has recently been questioned by investors after China’s DeepSeek made a competitive chatbot with allegedly fewer AI chips.
“Almost the entire world got it wrong,” Huang said on stage at the conference, dressed in his usual black leather jacket and jeans. He called the conference “the Super Bowl of AI.”
“The amount of computation we need as a result of agentic AI, as a result of reasoning, is easily 100 times more than we thought we needed this time last year,” he said, referring to autonomous AI agents that require little human intervention for routine tasks.
But his presentation positioning the company’s existing chips as well suited for changes sweeping the AI market failed to reassure investors. Nvidia shares fell 2.5 percent.
Nvidia’s big moneymaking chips face pressure from technological change as AI markets shift from “training,” AI models such as chatbots on huge troves of data to make them smart, to “inference,” which is when the model uses its intelligence to produce answers for users.
New chips
Huang announced new chips, including its next GPU chip Blackwell Ultra, which will be available in the second half of this year.
He said Nvidia’s chips have two main purposes: helping AI systems respond smartly to a huge number of users, but also giving those responses as fast as possible.
Huang argued that Nvidia’s chips are the only ones that can do both.
“If you take too long to answer a question, the customer is not going to come back. This is like web search,” he said.
He announced new software called Dynamo, which Nvidia released for free and is meant to speed up the process of reasoning.
Huang also announced that automaker General Motors has selected Nvidia to build its self-driving car fleet.
Nvidia stock has more than quadrupled in value over the past three years as the company powered the rise of advanced AI systems such as ChatGPT, Claude and many others.
Much of Nvidia’s success stemmed from the decade that the Santa Clara, California-based company spent building software tools to woo AI researchers and developers – but it was Nvidia’s data centre chips, which sell for tens of thousands of dollars each, that accounted for the bulk of its US$130.5 billion ($205 billion) in sales last year.
Huang hinted last year the new flagship offering will be named Rubin and consist of a family of chips – including a graphics processing unit, a central processing unit and networking chips – all designed to work in huge data centers that train AI systems.
Analysts expect the chips to go into production this year and roll out in high volumes starting next year.
Nvidia is trying to establish a new pattern of introducing a flagship chip every year, but has so far hit both internal and external obstacles.
The company’s current flagship chip, called Blackwell, is coming to market slower than expected after a design flaw caused manufacturing problems.
The broader AI industry last year grappled with delays in which the prior methods of feeding expanding troves of data into ever-larger data centres full of Nvidia chips had started to show diminishing returns.
