Six chips form an AI supercomputer
The Vera Rubin platform, named after the American astronomer, consists of six different chips: the Vera CPU, Rubin GPU, NVLink-6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU, and Spectrum-6 Ethernet Switch. Nvidia says the system is already in "full production" and will be available through partners starting in the second half of 2026.
The performance claims are impressive: The Rubin GPU is said to deliver three times the AI training compute and five times the AI inference compute of its predecessor, Blackwell. That's specifically for NVFP4. CEO Huang emphasized that the third-generation Transformer Engine with hardware-accelerated adaptive compression plays a major role in these gains. Inference token costs should drop by a factor of ten, and training large mixture-of-experts models will require only a quarter of the GPUs.
The complete Vera Rubin NVL72 rack reaches 260 terabytes per second of bandwidth, according to Nvidia. The sixth generation of NVLink delivers 3.6 terabytes per second per GPU.
What "full production" actually means
