Today at the Consumer Electronics Show, Nvidia CEO Jensen Huang officially announced the company’s new Rubin computing architecture, calling it the state-of-the-art in AI hardware. The new architecture is currently in production and will be further enhanced in the second half of this year.
“Vera Rubin is designed to address a fundamental challenge we have: the rapidly increasing amount of computation required for AI,” Huang told the audience. “Today, I can say that Vera Rubin is in full operation.”
First announced in 2024, the Rubin architecture is the latest in Nvidia’s continuous hardware development cycle that has turned it into the world’s most valuable company. The Rubin architecture replaces the Blackwell architecture, which replaces the Hopper and Lovelace architectures.
Rubin chips are already slated for use with nearly every major cloud provider, including Anthropic, OpenAI, and Nvidia’s famous partnership with Amazon Web Services. The Rubin system will also be used in HPE’s Blue Lion supercomputer and Lawrence Berkeley National Laboratory’s upcoming Doudna supercomputer.
The Rubin Architecture, named after astronomer Vera Florence Cooper Rubin, consists of six separate chips designed to work together. Although Rubin GPUs are at the center, the architecture also addresses growing bottlenecks in storage and interconnect with new improvements in the Bluefield and NVLink systems, respectively. The architecture also includes a new Vera CPU designed for agent inference.
Dion Harris, senior director of AI Infrastructure Solutions at Nvidia, explained the benefits of the new storage while pointing to the increased caching-related memory demands of modern AI systems.
“As new types of workflows start to take effect, such as agent AI and long-running tasks, they place a lot of stress and requirements on the KV cache,” Harris told reporters on a conference call, referring to the memory system used by AI models to compress input. “So we introduced a new storage layer that connects outside of the computing device, allowing us to scale our storage pool more efficiently.”
tech crunch event
san francisco
|
October 13-15, 2026
As expected, the new architecture also shows significant advances in speed and power efficiency. According to Nvidia testing, the Rubin architecture runs 3.5x faster than the previous Blackwell architecture for model training tasks and 5x faster for inference tasks, reaching up to 50 petaflops. The new platform also supports 8x more inference compute per watt.
Rubin’s new capabilities come amid a fierce race to build AI infrastructure. In this race, both AI labs and cloud providers are competing for Nvidia chips and the equipment needed to power them. During an October 2025 earnings call, Huang estimated that $3 trillion to $4 trillion will be spent on AI infrastructure over the next five years.
Read TechCrunch’s full coverage of the annual CES conference here.
