Nvidia unveils its advanced Rubin chip design

Nvidia Unveils Rubin: The Next Generation of AI Hardware

At the Consumer Electronics Show, Nvidia CEO Jensen Huang introduced the company's latest breakthrough in artificial intelligence hardware: the Rubin computing architecture. This cutting-edge platform is already in production, with plans to accelerate manufacturing in the latter half of the year.

"Vera Rubin is engineered to tackle a core issue we face today: the soaring computational demands of AI," Huang explained to attendees. "I'm pleased to announce that Vera Rubin has entered full-scale production."

Rubin Architecture: A New Era for Nvidia

First revealed in 2024, the Rubin architecture is the newest milestone in Nvidia’s ongoing innovation cycle, which has propelled the company to the top of the global tech industry. Rubin is set to succeed the Blackwell architecture, which itself replaced the earlier Hopper and Lovelace generations.

Widespread Adoption Across the Cloud and Supercomputing

Major cloud service providers are already planning to deploy Rubin chips, with notable collaborations involving Anthropic, OpenAI, and Amazon Web Services. Rubin-based systems will also power HPE’s Blue Lion supercomputer and the upcoming Doudna supercomputer at Lawrence Berkeley National Laboratory.

Innovative Design Inspired by Vera Rubin

Named after the renowned astronomer Vera Florence Cooper Rubin, the Rubin platform is built around six interconnected chips. At its core is the Rubin GPU, complemented by advancements in storage and networking through updated Bluefield and NVLink technologies. The lineup also features a new Vera CPU, optimized for agentic reasoning tasks.

Addressing AI’s Growing Memory Needs

Dion Harris, Nvidia’s senior director of AI infrastructure solutions, highlighted how the increasing complexity of AI workloads—such as agentic AI and long-duration tasks—places greater demands on memory systems. "As these new workflows emerge, the pressure on your KV cache intensifies," Harris noted, referencing a memory mechanism used by AI models to streamline input data. "To meet these needs, we’ve introduced an external storage tier that connects directly to the compute device, enabling much more scalable storage solutions."

Performance and Efficiency Leap Forward

The Rubin architecture delivers remarkable improvements in both speed and energy efficiency. Nvidia reports that Rubin outpaces the previous Blackwell generation by 3.5 times in model training and achieves fivefold gains in inference performance, reaching up to 50 petaflops. Additionally, the new platform offers eight times greater inference compute per watt.

Meeting the Demands of the AI Infrastructure Race

These advancements arrive at a time of fierce competition to build robust AI infrastructure, with both research labs and cloud providers vying for Nvidia’s latest chips and the resources to support them. During an earnings call in October 2025, Huang projected that global investment in AI infrastructure could reach between $3 trillion and $4 trillion over the next five years.

Upcoming Tech Event