On Monday, forward of the Client Electronics Present, Huang formally launched the Vera Rubin structure, which is now in manufacturing and anticipated to ramp up in quantity within the second half of the 12 months. This transfer follows a blockbuster 12 months for its Blackwell chip, as demand for AI infrastructure continued to surge.
In a press briefing forward of Huang’s keynote, Dion Harris, Nvidia’s senior director of HPC and AI infrastructure options, described Vera Rubin as “six chips that make one AI supercomputer.”
“Vera Rubin is designed to handle this elementary problem that now we have: The quantity of computation vital for AI is skyrocketing,” Huang advised the viewers throughout a presentation on the CES.
Huang added that in comparison with the Blackwell mannequin, Rubin marks a leap in efficiency, with greater than triple the pace, might run inference 5 instances quicker, and may ship considerably extra inference compute per watt of vitality.
Rubin was first introduced in 2024 and has been slated to exchange Blackwell ever since. The early debut comes months forward of the late-2026 timeline Nvidia had beforehand projected.
Named after astronomer Vera Rubin, who found the existence of darkish matter, Nvidia mentioned in a press launch that the structure is designed to help extra advanced, agent-style AI workloads, in addition to extra networking and knowledge motion.
The Rubin methods are already lined up for deployment throughout a lot of the cloud trade. Nvidia mentioned companions, together with Amazon Web Services, OpenAI, Anthropic, alongside the upcoming Doudna system at Lawrence Berkeley Nationwide Laboratory, all plan to make use of the brand new platform.
The accelerated launch comes shortly after Nvidia reported document knowledge middle income, up 66% from a 12 months earlier, pushed largely by demand for Blackwell and Blackwell Extremely GPUs. These chips have grow to be a benchmark for the present AI increase are extensively seen as a take a look at of whether or not spending on AI infrastructure is sustainable.
Huang has beforehand estimated that between $3 trillion and $4 trillion could possibly be spent globally on AI infrastructure over the following 5 years. Nvidia mentioned services constructed on the Rubin platform will start rolling out from companions within the second half of 2026.
