By Stephen Nellis and Max A. Cherney
(Reuters) – When Nvidia CEO Jensen Huang takes the stage this week for the company’s annual software developer conference, he will defend his nearly $3 trillion chip company’s dominance as pressure mounts on its biggest customers to rein in the costs of artificial intelligence.
Nvidia’s conference comes after China’s DeepSeek spooked U.S. markets with a competitive chatbot it alleged took less computing power than rivals to create. Nvidia’s stock dropped because selling computing power in the form of chips that cost tens of thousands of dollars apiece is what helped Nvidia’s revenue more than quadruple over the past three years to $130.5 billion.
At the conference, Nvidia is expected to reveal details of a chip system called Vera Rubin, named for the American astronomer who pioneered the concept of dark matter, with the system expected to go into mass production later this year. Those details will come even as Rubin’s predecessor, a chip named after mathematician David Blackwell announced this time last year, is trickling onto the market after production delays that have eaten into Nvidia’s margins.
Nvidia’s big moneymakers face pressure from technological change as AI markets shift from “training,” which is the process of feeding AI models such as chatbots huge troves of data to make them smart, to “inference,” which is when the model uses those smarts to produce answers for users. Nvidia, with a market share exceeding 90%, owns the training market but faces competition in inference – and how much market share those competitors take will depend on how inference computing is carried out.
‘BIGGER HAMMERS’
Inference computing comes in many forms, from a smartphone that rewords emails to a data center churning out complex analysis of financial documents. Scores of startup companies in Silicon Valley and beyond, as well as Nvidia’s traditional rivals such as Advanced Micro Devices , are betting that they can sell chips that will get the job done at lower overall cost – especially electricity costs, where Nvidia’s chips consume so much power that AI companies are investigating nuclear reactors to power them.
“They have a hammer, and they’re just making bigger hammers,” said Bob Beachler, vice president at Untether AI, one of the at least 60 startups trying to unseat Nvidia in inference markets. “They own the (training) market. And so every new chip they come out with has a lot of training baggage.”
But Nvidia has argued that a new kind of AI called “reasoning” plays in its favor. Reasoning chatbots think aloud, generating a few lines of text and then reading that text back to themselves to think on the problem more – all of which uses more of the computing power that Nvidia’s chips excel at.
“The market for inference is going to be many times bigger than the training market,” said Jay Goldberg, chief executive of D2D Advisory, a finance and strategy consulting firm. “As inference becomes more important, their percentage share will be lower, but the total market size and the pool of revenues could be much, much larger.”
BEYOND CHATBOTS
Nvidia is also expected to hint at its plans in other computing markets, such as using new AI techniques that improve chatbots to make robots more useful.
One big area of focus will be quantum computing. In January, comments by Huang that the technology was decades away helped crash shares of companies betting on it and spurred Microsoft and Alphabet’s Google to come out with claims that the technology is much closer to usefulness. That in turn prompted Nvidia to announce it would devote a full day of its conference to the state of the quantum industry and its own plans.
Huang will deliver the keynote address on Tuesday.
Also on deck is Nvidia’s efforts to build a personal computer central processor chip, an endeavor first reported by Reuters and revealed by Nvidia in January.
“It could eat into what’s left of the Intel market,” said Maribel Lopez, an independent technology industry analyst.
(Reporting by Stephen Nellis and Max Cherney in San Francisco; Editing by Andrea Ricci)
Comments