Nvidia CEO Jensen Huang officially unveiled the company's new Rubin computing architecture, hailing it as the "state of the art in AI hardware." Launched at the Consumer Electronics Show, the Rubin platform is already in full production and is set to significantly ramp up later this year.
"Vera Rubin is designed to address this fundamental challenge that we have: The amount of computation necessary for AI is skyrocketing," Huang told the audience. "Today, I can tell you that Vera Rubin is in full production."
First announced in 2024, the Rubin architecture represents the latest evolution in Nvidia's aggressive hardware development roadmap, a strategy that has propelled the company to become the world's most valuable corporation. Rubin is poised to succeed the Blackwell architecture, continuing a rapid upgrade cycle that previously saw Blackwell replace the Hopper and Lovelace platforms.
The Rubin chips are already earmarked for deployment by virtually every major cloud provider. This includes high-profile partnerships with leading AI developers like Anthropic and OpenAI, as well as cloud giants such as Amazon Web Services. Beyond cloud infrastructure, Rubin systems are also slated for integration into advanced supercomputing projects, including HPE's Blue Lion supercomputer and the forthcoming Doudna supercomputer at Lawrence Berkeley National Lab.
Named in honor of the pioneering astronomer Vera Florence Cooper Rubin, the new architecture integrates six distinct chips designed to operate cohesively. At its core is the powerful Rubin GPU. However, the architecture also tackles critical bottlenecks in data storage and interconnection through significant enhancements to the Bluefield and NVLink systems, respectively. A new Vera CPU, specifically engineered for "agentic reasoning," is also part of the comprehensive design.
Dion Harris, Nvidia's senior director of AI infrastructure solutions, highlighted the critical role of the new storage system. He explained that modern AI systems, particularly those handling "agentic AI" or long-term tasks, face escalating cache-related memory demands.
"As you start to enable new types of workflows, like agentic AI or long-term tasks, that puts a lot of stress and requirements on your KV cache," Harris noted, referring to a memory system that AI models utilize to condense inputs. "So we've introduced a new tier of storage that connects externally to the compute device, which allows you to scale your storage pool much more efficiently."
The Rubin architecture delivers substantial advancements in both speed and power efficiency. Nvidia's internal tests indicate that Rubin will perform 3.5 times faster than its predecessor, Blackwell, for model-training tasks. For inference tasks, it boasts a five-fold speed increase, reaching up to 50 petaflops. Furthermore, the new platform achieves an impressive eight times more inference compute per watt, underscoring its energy efficiency.
The introduction of Rubin comes at a time of intense global competition to build out robust AI infrastructure. AI research labs and major cloud providers are aggressively vying for Nvidia's cutting-edge chips and the extensive facilities required to power them. Reflecting this massive investment, CEO Jensen Huang projected in an October 2025 earnings call that an astounding $3 trillion to $4 trillion will be invested in AI infrastructure over the next five years.
