NVIDIA debuts Rubin supercomputer at CES 2026 cutting inference costs
NVIDIA just introduced its Rubin supercomputing platform at CES 2026, and it's all about making powerful AI more affordable and accessible.
Rubin can deliver up to a 10x reduction in inference token costs and requires four times fewer graphics cards to train mixture-of-experts (MoE) models compared to NVIDIA's previous Blackwell platform.
The goal? To help accelerate mainstream adoption of advanced AI models, especially in the consumer space, and make large-scale AI deployment more practical.
Six-chip Rubin includes Vera CPU
Rubin packs 6 chips into 1 supercomputer, including an NVIDIA Vera CPU (with 88 custom Olympus cores) and the NVIDIA Rubin GPU capable of delivering up to 50 petaflops of NVFP4 computational power.
Thanks to upgraded networking tech, everything runs smoother and faster.
Starting later this year, major cloud services like AWS and Google Cloud will get access, potentially making high-end AI tools way more common in our daily lives.