By Steven Rosenbush

SAN JOSE, Calif. -- Nvidia CEO Jensen Huang has laid out a new product road map that shows the company's AI infrastructure and reasoning models cycling through upgrades at breakneck speed, but some customers may not be ready to move quite so fast.

Huang on Tuesday delivered a two-hour keynote at Nvidia GTC in which he said an updated version of its current leading-edge Blackwell artificial intelligence system, known as Ultra, would be available later this year. He also unveiled a new line of an even more powerful system called Rubin that is expected to be available in the second half of 2026. And an Ultra version of the Rubin will offer 14 times the performance of the Blackwell, he said.

Nvidia's graphics processing units and infrastructure have been wildly in demand for tasks such as training leading-edge AI models. That, along with the company's many other technologies, have made Nvidia, with a current valuation of more than $2.8 trillion, among the most valuable in the world -- albeit no longer No. 1.

Traditionally, computers have been about retrieving information, but many people believe this will be the year computing moves into a new era of reasoning, which requires a lot more computing power. "The amount of computation needed is easily 100 times more than we thought we needed at this time last year," Huang said, reiterating a point he has made in recent months.

As a result, he argues that the Blackwell systems are rapidly eclipsing the prior generation known as Hopper. "There are circumstances where Hopper is fine...not many," he quipped from the stage. "When technology is moving this fast and because the workload is so intense, we really like you to invest in the right versions."

Not so fast

Plenty of cloud service providers, companies and researchers will want the latest versions of chips, AI, software and other infrastructure, and Nvidia says demand for Blackwell "AI factories" is strong.

But there are companies that are perfectly content running their businesses on systems of an older vintage.

"My enterprise, internally, I have only 250 GPUs, that's it. And I have plenty," Hewlett Packard Enterprise CEO Antonio Neri said in an interview at GTC. Those chips, of pre-Blackwell vintage, including some Hoppers, are sufficient. "For me, for what I do, that's plenty," he said.

He thinks Nvidia's software is crucial to its success. He has focused on helping customers quickly build and deploy agentic AI with Nvidia NIM software and HPE's private cloud AI. It's all about speed. "Time to market is what really matters for enterprises," Neri said.

Ford Motor director of AI Bryan Goodman said that from his perspective, Hopper chips remain relevant, at least for now. "I don't quite see it as they're useless," Goodman said. "He was, of course, trying to say that his new product is so much better than his old product, that he's making his old product obsolete. But I suspect we'll get a lot of use out of our Hopper GPUs the next few years. And we're working on Blackwell."

The case for perpetual upgrades

Nvidia has a lot riding on the case for upgrades, given that its shares were recently trading for less than 27 times this year's projected earnings, a 23% discount to the multiple it was commanding at last year's conference.

The logic of Nvidia's pitch to cloud-service providers and enterprise customers is that they can't not want the latest technology, not only because performance keeps improving, but because price-performance is continually getting better. Overall spending on AI is going up, but the unit cost of analyzing "tokens" of data on the latest hardware keeps going down. Customers must stay ahead of that curve for economic reasons, to keep their costs from exploding along with consumption.

That argument may well be persuasive. Even so, the time frame for proving it out is unclear. Not all customers are willing or able to keep pace with the cadence of yearly upgrades.

Huang at a press conference Wednesday said he didn't expect instant adoption of new products. "We're not building chips anymore; those are the good old days," Huang said. "What we do now is we build AI infrastructure that are deployed hundreds of billions of dollars at a time, so you better do a good job planning. And AI infrastructure isn't something you decide to buy today and deploy tomorrow, you decide two years in advance...plan and hopefully stand it up quickly."

Write to Steven Rosenbush at steven.rosenbush@wsj.com

(END) Dow Jones Newswires

March 20, 2025 12:36 ET (16:36 GMT)

Copyright (c) 2025 Dow Jones & Company, Inc.