Nvidia said on Jan. 6 it launched the Rubin platform for next-generation AI at CES 2026.

The Rubin platform consists of six new chips designed to build advanced AI supercomputers. The six chips include the Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU and Spectrum-6 Ethernet Switch.

The Rubin platform applies advanced co-design across the six chips to reduce training time and inference token costs. It cuts cost per token by 10 times compared with the Blackwell platform and reduces by 4 times the number of GPUs required to train MoE models compared with before. The company stressed this was the result of introducing five technologies, including the latest NVLink interconnect technology, Transformer Engine, Confidential Computing, the RAS Engine and the Vera CPU.

Rubin is currently in mass production, and Rubin-based products are expected to be launched through partners in the second half of 2026. Cloud providers set to be the first to deploy Vera Rubin-based instances in 2026 include AWS, Google Cloud, Microsoft and OCI, as well as cloud partners CoreWeave, Lambda, Nebius and Nscale. Microsoft plans to deploy the Vera Rubin NVL72 rack-scale system in next-generation AI data centres, including the Fairwater AI Superfactory.

Sixth-generation NVLink supports bandwidth of 3.6 TB per second per GPU, and the Vera Rubin NVL72 rack provides bandwidth of 260 TB per second, larger than the entire internet. The Vera CPU has 88 custom Olympus cores, Armv9.2 compatibility and high-speed NVLink-C2C connectivity. The Rubin GPU includes a third-generation Transformer Engine that supports hardware-accelerated adaptive compression technology and provides 50 petaflops of NVFP4 computing for AI inference.

BlueField-4 introduces Advanced Secure Trust Resource Architecture, or ASTRA. It is a system-level trust architecture that provides a single trusted point of control so AI infrastructure builders can securely provision, isolate and operate large-scale AI environments without performance degradation. Spectrum-X Ethernet photonics co-packaged optical switch systems provide 10 times improved reliability and 5 times longer uptime for AI applications.

Nvidia founder and CEO Jensen Huang said, "As computing demand for AI training and inference is surging more than ever, Rubin's arrival is very timely." He said, "Nvidia is unveiling next-generation AI supercomputers every year. Against that backdrop, Rubin, which applies advanced co-design across six chips, will make a leap forward toward a new horizon for AI."