Nvidia readies Vera Rubin to replace Blackwell

Nvidia's upcoming advanced GPUs, Blackwell Ultra and Vera Rubin, give it as much as a two-year lead over competitors, an analyst says.

Nvidia plans to release a high-performance AI computing system powered by Blackwell Ultra GPUs this year and ship its next-generation AI superchip, Vera Rubin, in the second half of 2026 to keep pace with the AI industry's demand for more computing power.

On Tuesday, Nvidia CEO Jensen Huang also unveiled at the company's GPU Technology Conference(GTC) more powerful networking hardware for organizations heading to the next phase in AI -- reasoning models that enable agentic AI. Reasoning models, which are under development by major cloud providers, are AI systems that analyze data based on logical rules and previous learnings to make decisions.

"The computation requirement, the scaling law of AI, is more resilient and, in fact, hyper-accelerated," Huang said in his opening keynote. "The amount of computation we need at this point 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."

In 2026, Nvidia will respond to the expected power demand with Vera Rubin, comprising a next-generation CPU (Vera) and GPU (Rubin). The processor is named after an American astronomer who discovered dark matter, an invisible form of matter that plays a crucial role in the formation of galaxies.

The system running the Vera Rubin GPU architecture will include next-generation NVLink 6 interconnect technology delivering 3600 GB per second of bandwidth and a new high-bandwidth memory (HBM) 4 that reaches 13 TB per second, compared to 8 TB per second of HBM3e used in Blackwell Ultra. The Vera Rubin will also use a new network interface card, the CX9 SuperNIC.

"Basically, everything is brand new except for the chassis," Huang said, noting that each rack is 600 kWs and contains 2.5 million parts.

The Vera Rubin NVL144 system, available in the second half of next year, will be followed by Rubin Ultra NVL576 a year later, Huang said. The numbers correspond to the number of GPUs in each system.

Nvidia's focus on more powerful GPUs is warranted because reasoning models will strain AI infrastructure more than chatbots today ever could, according to Nick Patience, vice president and practice lead for AI at The Futurum Group. For now, Nvidia is ahead of all competitors in delivering the most powerful technology.

"Nvidia still has a window where it completely dominates the high-end GPU market, probably for the next 18 to 24 months at least," Patience said. "Nvidia will be challenged by the other chip vendors, but its build-out of a software stack -- with Dynamo being the latest evidence of that -- is a smart move to increase the size and scope of its market."

Huang introduced Nvidia Dynamo at GTC. The open source software orchestrates inference communications across thousands of GPUs powering reasoning models. Inference is the process of inputting new data into the models so they can draw logical conclusions, make predictions or solve problems.

In the second half of this year, Nvidia's hardware partners will release the company's DGX SuperPod systems built with Blackwell Ultra GPUs, renamed the B300 series. The GPU is a dual-die design with 208 billion transistors that provides a 50% performance increase over the B200 series, according to Nvidia.

The DGX systems will be available with either Blackwell Ultra (B300) or Grace Blackwell Ultra (GB300) chips. Systems with the latter are liquid-cooled and contain 36 Grace CPUs and 72 B300s. DGX B300 systems have air-cooled options.

More than a dozen hardware makers plan to start selling B300 servers in the second half of the year. They include Cisco, Dell Technologies, Hewlett Packard Enterprise and Lenovo.

Nvidia unveiled silicon photonics networking switches at GTC, including the Spectrum-X and Quantum-X. The switches can connect millions of GPUs across sites, according to Nvidia. The Spectrum-X is an Ethernet networking platform with multiple configuration options. The Quantum-X is an InfiniBand platform in a switch that provides 144 ports of 800 Gbs.

Nvidia partners will ship the Quantum-X system this year and the Spectrum-X switches in 2026.

Nvidia also introduced the DGX brand of personal AI supercomputers.

The DGX Station contains the GB300 desktop processor with 784 GB of coherent memory. The GPU is connected to a Grace CPU using the NVLink-C2C.

Nvidia also introduced a much smaller AI computer called the DGX Spark. It's powered by the GB10 Grace Blackwell processor.

The two computers are designed for AI developers, researchers, data scientists and students to prototype, fine-tune and inference models. Nvidia started taking reservations for the Spark on Monday. Manufacturing partners Asus, Boxx, Dell, HP, Lambda and Supermicro plan to offer the Station this year.

Antone Gonsalves is an editor at large for Informa TechTarget, reporting on industry trends critical to enterprise tech buyers. He has worked in tech journalism for 25 years and is based in San Francisco. Have a news tip? Please drop him an email.

Dig Deeper on AI infrastructure