Nvidia has unveiled Rubin, a next-generation GPU structure and tightly built-in AI computing platform that’s set to exchange the US chip large’s present flagship Blackwell structure.
Rubin is Nvidia’s first extreme-codesigned platform and can comprise six AI chips, alongside varied networking applied sciences and system software program, all working collectively as a single computing unit. It was launched by Nvidia founder and CEO Jensen Huang onstage at CES 2026 in Las Vegas, United States on Monday, January 5. Rubin computing models are already in full manufacturing, with services and products powered by these models anticipated to launch within the second half of 2026.
In his roughly two hour-long keynote speech, Huang stated that AI is scaling into each area and each gadget. With Rubin, NVIDIA goals to “push AI to the following frontier” whereas slashing the price of producing tokens to roughly one-tenth that of the earlier platform, making large-scale AI much more economical to deploy, he additional stated.
Nvidia on Monday additionally introduced its newest sequence of open-weight AI reasoning fashions known as Alpamayo that’s designed particularly for autonomous autos like self-driving vehicles.
The launch of Rubin comes months after Nvidia reported report excessive information centre income, up 66 per cent over the prior 12 months. This progress has been attributed to a rise in demand for Blackwell and Blackwell Extremely GPUs. They’ve served as a key indicator in whether or not the AI growth is sustainable or turning right into a bubble. The runaway success of Blackwell and chips primarily based on this structure have set a excessive bar each in efficiency and market demand for the newly unveiled Rubin platform.
“Computing has been basically reshaped because of accelerated computing, because of synthetic intelligence. What meaning is a few $10 trillion or so of the final decade of computing is now being modernized to this new manner of doing computing,” Huang stated in his keynote speech.
“The quicker you practice AI fashions, the quicker you will get the following frontier out to the world. That is your time to market. That is know-how management,” he added. Emphasising Nvidia’s efforts to roll out open-weight AI fashions throughout domains, Huang stated that they’ve fashioned a worldwide ecosystem of intelligence that builders and enterprises can construct on.
Story continues beneath this advert
“Each single six months, a brand new mannequin is rising, and these fashions are getting smarter and smarter. Due to that, you can see the variety of downloads has exploded,” he additional stated.
What’s Rubin?
Nvidia’s next-generation computing platform has been named after Vera Rubin, an American astronomer recognized for her analysis on galaxy rotation charges which served as proof that darkish matter exists.
Rubin has been described as an AI supercomputer made up of six chips. The elements which might be a part of the computing platform are:
-Rubin GPUs with 50 petaflops of NVFP4 inference
-Vera CPUs engineered for information motion and agentic processing
-NVLink 6 scale‑up networking
-Spectrum‑X Ethernet Photonics scale‑out networking
-ConnectX‑9 SuperNICs
-BlueField‑4 DPUs
Story continues beneath this advert
All these elements have been designed collectively (excessive codesign) which is essential as a result of scaling AI to gigascale requires tightly built-in innovation throughout chips, trays, racks, networking, storage and software program to get rid of bottlenecks and dramatically scale back the prices of coaching and inference, Huang stated.
Rubin helps third-generation confidential computing and would be the first rack-scale trusted computing platform, as per the corporate.
By way of efficiency, Nvidia claimed that Rubin GPUs had been able to delivering 5 occasions as a lot AI coaching compute as Blackwell. Total, the Rubin structure can be utilized to coach a big ‘mixture-of-experts (MoE)’ AI mannequin in the identical period of time as Blackwell whereas utilizing 1 / 4 of the GPUs and at one-seventh the token price.
Nvidia additionally launched an Inference Context Reminiscence Storage platform for AI-native storage. It contains an AI‑native KV‑cache layer that reinforces lengthy‑context inference with 5x increased token throughput, higher efficiency per greenback of complete price of possession (TCO), and 5 occasions higher energy effectivity.
Story continues beneath this advert
What’s Alpamayo?
At CES 2026, Nvidia additionally introduced the newest addition to its household of open-weight AI fashions designed for particular domains. Through the years, the chipmaker has additionally steadily emerged as a frontier AI mannequin developer. These open-weight AI fashions have been educated utilizing Nvidia’s personal supercomputers.
Its present portfolio contains Clara, open-weight AI fashions designed for the healthcare sector together with Earth-2 for local weather science, Nemotron for reasoning and multimodal AI, Cosmos for robotics and simulation, and GR00T for powering embodied intelligence.
On Monday, Nvidia unveiled its Alpamayo fashions for autonomous driving. “Now on prime of this platform, NVIDIA is a frontier AI mannequin builder, and we construct it in a really particular manner. We construct it fully within the open in order that we will allow each firm, each trade, each nation, to be a part of this AI revolution,” Huang stated.
“These fashions are open to the world. You possibly can create the mannequin, consider it, guardrail it and deploy it,” he added.