Nvidia launches powerful new Rubin chip architecture

Today astatine the Consumer Electronics show, Nvidia CEO Jensen Huang officially launched the company’s caller Rubin computing architecture, which helium described arsenic the authorities of the creation successful AI hardware. The caller architecture is presently successful accumulation and is expected to ramp up further successful the 2nd fractional of the year.

“Vera Rubin is designed to code this cardinal situation that we have: The magnitude of computation indispensable for AI is skyrocketing.” Huang told the audience. “Today, I tin archer you that Vera Rubin is successful afloat production.”

The Rubin architecture, which was archetypal announced in 2024, is the latest effect of Nvidia’s relentless hardware improvement cycle, which has transformed Nvidia into the astir invaluable corp successful the world. The Rubin architecture volition regenerate the Blackwell architecture, which successful turn, replaced the Hopper and Lovelace architectures.

Rubin chips are already slated for usage by astir each large unreality provider, including high-profile Nvidia partnerships with Anthropic, OpenAI, and Amazon Web Services. Rubin systems volition besides beryllium utilized successful HPE’s Blue Lion supercomputer and the upcoming Doudna supercomputer astatine Lawrence Berkeley National Lab.

Named for the astronomer Vera Florence Cooper Rubin, the Rubin architecture consists of six abstracted chips designed to beryllium utilized successful concert. The Rubin GPU stands astatine the center, but the architecture besides addresses increasing bottlenecks successful retention and interconnection with caller improvements successful the Bluefield and NVLink systems respectively. The architecture besides includes a caller Vera CPU, designed for agentic reasoning.

Explaining the benefits of the caller storage, Nvidia’s elder manager of AI infrastructure solutions Dion Harris pointed to the increasing cache-related representation demands of modern AI systems.

“As you commencement to alteration caller types of workflows, similar agentic AI oregon semipermanent tasks, that puts a batch of accent and requirements connected your KV cache,” Harris told reporters connected a call, referring to a representation strategy utilized by AI models to condense inputs. “So we’ve introduced a caller tier of retention that connects externally to the compute device, which allows you to standard your retention excavation overmuch much efficiently.”

As expected, the caller architecture besides represents a important beforehand successful velocity and powerfulness efficiency. According to Nvidia’s tests, the Rubin architecture volition run 3 and a fractional times faster than the erstwhile Blackwell architecture connected model-training tasks and 5 times faster connected inference tasks, reaching arsenic precocious arsenic 50 petaflops. The caller level volition besides enactment 8 times much inference compute per watt.

Rubin’s caller capabilities travel amid aggravated contention to physique AI infrastructure, which has seen some AI labs and unreality providers scramble for Nvidia chips arsenic good arsenic the facilities indispensable to powerfulness them. On an net telephone successful October 2025, Huang estimated that between $3 trillion and $4 trillion volition beryllium spent connected AI infrastructure implicit the adjacent 5 years.