NVIDIA's Vera Rubin superchip is here, and it's seriously fast
NVIDIA recently announced its Vera Rubin platform, first unveiled by CEO Jensen Huang at GTC Washington in October 2025 and detailed further at CES 2026.
This new superchip is set to deliver five times the AI performance of previous models used for chatbots and smart apps.
It packs two powerful Rubin GPUs and an 88-core CPU, with full production began in early 2026.
What makes Vera Rubin special?
Built on TSMC's advanced 3nm process, Vera Rubin uses HBM4 memory for ultra-fast data speeds.
Each GPU handles up to 50 petaFLOPS of AI inference or ~16 petaFLOPS for training, with a massive total of 576GB HBM4 memory per board (two GPUs)—so it can tackle huge AI workloads without breaking a sweat.
How does it stack up?
The NVL144 racks powered by this chip hit 3.6 exaFLOPS FP4 inference and offer blazing-fast bandwidths up to 260TB/s via NVLink.
That means they're over three times faster than NVIDIA's last-gen Blackwell setups in certain AI tasks.
And if you love future tech: NVIDIA says an even more powerful "Rubin Ultra" is coming in 2027, which is expected to compete with AMD and Google's custom chips.