NVIDIA unveils Rubin: a six‑chip platform powering a next‑generation AI supercomputer

Nvidia launched “Rubin,” a six-chip AI platform it says will sharply cut the cost of running and training frontier models, as competition to supply the world’s largest AI systems intensifies. Debuting at CES, Rubin integrates a new Vera CPU, Rubin GPU, NVLink 6 interconnect, ConnectX-9 SuperNIC, BlueField-4 DPU and Spectrum-6 Ethernet switch in a rack-scale design. Nvidia claims up to a 10x reduction in inference token costs and 4x fewer GPUs needed to train mixture-of-experts models versus its prior Blackwell platform. The company highlighted Microsoft’s plan to deploy Vera Rubin NVL72 systems in next-gen “Fairwater” AI superfactories and said AWS, Google Cloud, OCI and CoreWeave will roll out Rubin-based instances in the second half of 2026. Rubin also adds features aimed at reliability and security—confidential computing spanning CPU, GPU and NVLink, a second-generation RAS engine, and an AI-native storage layer using BlueField-4 to share inference context. Nvidia’s Spectrum-6 Ethernet and new Photonics switches promise better power efficiency and uptime through co-packaged optics. The release, backed by endorsements from OpenAI, Anthropic, Meta and others, underscores Nvidia’s push to keep annual cadence in AI systems, while cautioning that many statements are forward-looking and timelines may change.

Read more