At the Consumer Electronics Show, Jensen Huang, CEO of Nvidia, officially unveiled the company’s new Rubin computing architecture, which he described as the latest achievement in artificial intelligence hardware. The new architecture is currently in production and is expected to further expand production capacity in the second half of this year.
Jensen Huang told attendees, “Vera Rubin aims to address a fundamental challenge we face: the amount of computation required for artificial intelligence is growing rapidly. Today, I can tell everyone that Vera Rubin has entered the full production stage. ”The Rubin architecture was initially announced in 2024 and is the latest achievement in Nvidia’s wireless hardware development cycle, which has transformed Nvidia into the world’s most valuable company by market capitalization. The Rubin architecture will replace the Blackwell architecture, which previously replaced the Hopper and Lovelace architectures.
Widely adopted by various industries
The Rubin chip is planned to be used by almost all major cloud service providers, including Nvidia’s high-profile partnerships with Anthropic, OpenAI, and Amazon Web Services. The Rubin system will also be used for HPE’s Blue Lion supercomputer and the upcoming DouDNA supercomputer at Lawrence Berkeley National Laboratory.
Architecture composition and technological innovation
This architecture is named after astronomer Vera Florence Cooper Rubin and consists of six independent chips designed for collaborative use. Rubin GPU is at the core, but this architecture also addresses the growing bottlenecks in storage and interconnection through new improvements to Bluefield and NVLink systems, respectively. The architecture also includes a brand new Vera CPU designed specifically for intelligent agent inference.
When explaining the advantages of the new storage system, Dion Harris, Senior Director of Artificial Intelligence Infrastructure Solutions at NVIDIA, pointed out the increasing demand for cache related memory in modern AI systems. Harris told reporters during a conference call, “When you start implementing new workflows such as intelligent AI or long-term tasks, it puts a lot of pressure and demand on your KV cache.” He refers to a memory system used by AI models to compress input content. “Therefore, we have introduced a new storage hierarchy that is externally connected to computing devices, allowing you to expand storage pools more efficiently.”
Significant performance leap
As expected, the new architecture also represents significant progress in speed and energy efficiency. According to Nvidia’s testing, the Rubin architecture will run 3.5 times faster than the previous generation Blackwell architecture in model training tasks and 5 times faster in inference tasks, with a maximum of 50 petaflops. The new platform will also support inference computing power that is 8 times higher per watt.
Industry background of publication
The release of Rubin’s new features comes at a time of intense competition in the construction of artificial intelligence infrastructure. In this competition, both AI labs and cloud service providers are vying for Nvidia chips and the facilities needed to power them. During the October 2025 earnings conference call, Jensen Huang estimated that $3 trillion to $4 trillion will be invested in AI infrastructure over the next five years.