NVIDIA has released details of this year's Rubin architecture.

NVIDIA CEO Jensen Huang officially unveiled the architecture details of this year's upcoming Rubin chip at CES 2026. The chip is currently in production and mass shipments are expected to begin in the second half of this year.

The Rubin platform comprises six components: a Vera CPU, a Rubin GPU, an NVLink 6 switch, a ConnectX-9 SuperNIC, a BlueField-4 DPU, and a Spectrum-6 Ethernet switch. It supports AI training and processing with reduced total cost. Huang stated that the current challenge in the AI ​​hardware world is the demand for both training and inference of AI models increasing faster than hardware capabilities can support, and Rubin addresses this problem. With five times better performance compared to the Blackwell architecture,

a single rack of Vera Rubin NVL72 will contain 72 Rubin GPUs and 32 Vera CPUs, all connected via NVLink 6, with a total of 20.7TB of HBM memory and 54TB of LPDDR5x memory. It supports inference runs at 3.6 EFLOPS and training runs at 2.5 EFLOPS. The cost per token is ten times lower than Blackwell.

Rubin will begin shipping through partners in the second half of this year, while major cloud providers will begin deployment in the first half.

💬 AI Content Assistant

Ask me anything about this article. No data is stored for your question.

NVIDIA CEO Jensen Huang officially unveiled the architecture details of this year's upcoming Rubin chip at CES 2026. The chip is currently in production and mass shipments are expected to begin in the second half of this year.

The Rubin platform comprises six components: a Vera CPU, a Rubin GPU, an NVLink 6 switch, a ConnectX-9 SuperNIC, a BlueField-4 DPU, and a Spectrum-6 Ethernet switch. It supports AI training and processing with reduced total cost. Huang stated that the current challenge in the AI ​​hardware world is the demand for both training and inference of AI models increasing faster than hardware capabilities can support, and Rubin addresses this problem. With five times better performance compared to the Blackwell architecture,

a single rack of Vera Rubin NVL72 will contain 72 Rubin GPUs and 32 Vera CPUs, all connected via NVLink 6, with a total of 20.7TB of HBM memory and 54TB of LPDDR5x memory. It supports inference runs at 3.6 EFLOPS and training runs at 2.5 EFLOPS. The cost per token is ten times lower than Blackwell.

Rubin will begin shipping through partners in the second half of this year, while major cloud providers will begin deployment in the first half.