SAN JOSE, CALIF.—As the energy and computing demands for AI keep shooting higher and higher, Nvidia is responding by previewing its next-generation GPU architectures, which promise to drastically increase performance while driving down costs.
At the company’s GTC conference on Tuesday in San Jose, Nvidia CEO Jensen Huang introduced “Rubin” and “Feynman,” Nvidia’s next AI-focused GPU architectures for 2026 and 2028. The company is also preparing an upgrade to the existing “Blackwell” architecture through a new GPU dubbed “Blackwell Ultra,” slated to arrive in the second half of this year.
Nvidia revealed the architectures to give companies time to plan and budget for their upcoming data centers as they race to develop new AI programs. The problem is that AI development is an expensive endeavor, requiring billions in investment to both buy the GPUs and to accommodate (as well as pay for) the mounting demands for electricity and cooling.
(Credit: Michael Kan)
Huang’s keynote at GTC also discussed how next-generation AI models require even more computing power to elevate their reasoning abilities, which can further drive up costs for companies. That’s because smarter AI models work by spending more time and compute resources to answer and verify the right solution to a user’s request through a “chain of thought” process.
Rubin and Feynman: Meeting the Accelerating Need for AI Power
According to Huang, the tech industry needs “100 times” more computing power than previously thought to deal with the rise of smarter “agentic AI.” The company’s roadmap tries to address the higher demands by upgrading every aspect of the GPU architecture, including boosting the transistor count, memory speeds, and interconnect.
As an example, Huang teased the development of a Rubin-powered AI server that will pack in a mind-boggling 1,300 trillion transistors, an immense increase from the 130 trillion transistors available in the existing Blackwell-powered GB200 NVL72 system.
(Credit: Michael Kan)
(Credit: Michael Kan)
In the near term, the upcoming Blackwell Ultra represents more of an incremental upgrade from the existing Blackwell architecture, which debuted a year ago. A single Blackwell Ultra GPU will be able to accommodate up to 288GB of HBM3e memory, a sizable increase from the maximum 192GB of memory in the earlier design.
Blackwell Ultra will be sold through a server unit called GB300 NVL72, which can offer a 50% performance increase over the existing Blackwell-powered GB200 NVL72 model. A single GB300 NVL72 unit will contain a total of 72 Blackwell Ultra GPUs, along with 36 Arm-based “Grace” CPUs.
(Credit: Michael Kan)
For companies looking for a bigger boost, Nvidia says its Rubin architecture will arrive alongside an upgraded “Vera” CPU to unleash even more performance. The first Vera Rubin NVL144 system promises to offer a performance increase of 3.3 times over the GB300. Meanwhile, the Rubin Ultra NVL576 for 2027 will boast a 14-times performance boost over the GB300. Huang projects that Rubin will massively bring down costs for AI providers.
Recommended by Our Editors
(Credit: Michael Kan)
(Nvidia GTC 2025: More Roadmap)
Nvidia’s CEO didn’t say much about the Feynman architecture. But on the software front, he did tout the release of Dynamo, an open source library, to further help companies streamline AI workloads at lower costs. “This allows each phase to be optimized independently for its specific needs and ensures maximum GPU resource utilization,” the company explained.
Huang: Blackwell Ultra Will Keep the Dollars Flowing
Despite the push to make next-generation AI more scalable, Huang still expects companies to spend plenty on acquiring GPUs from Nvidia. “The more you buy, the more you save,” he joked, the same quip that he’s delivered in past keynotes. In one presentation slide, he also showed that the tech industry is projected to spend a cool trillion dollars by 2028 on building new data centers.
(Credit: Michael Kan)
Nvidia didn’t reveal pricing for the Blackwell Ultra GB300. But a single unit will likely cost at least a few million, considering earlier estimates put the GB200 NVL72 at $3 million a pop. Nvidia is also marketing Blackwell Ultra as a sizable upgrade for users of its older “Hopper” architecture, promising a 50-times “increase in data center revenue opportunity.”
Amazon’s AWS, Google Cloud, and Microsoft Azure will be among the first cloud providers to offer access to Blackwell Ultra systems.
Get Our Best Stories!
This newsletter may contain advertising, deals, or affiliate links.
By clicking the button, you confirm you are 16+ and agree to our
Terms of Use and
Privacy Policy.
You may unsubscribe from the newsletters at any time.
About Michael Kan
Senior Reporter
