NVIDIA unveils Vera Rubin
NVIDIA used GTC 2026 to introduce Vera Rubin — an ‘AI factory’ platform that pairs new Vera CPUs, Rubin GPUs, and advanced networking to orchestrate industrial‑scale AI rather than just sell chips. The announcement refocuses the conversation on end‑to‑end orchestration, security, and fleet management as the enterprise metric of value. (finance.yahoo.com)
NVIDIA says the Vera Rubin program now comprises seven new chips that are “in full production” to support large-scale AI factory deployments. (investor.nvidia.com)) The flagship NVL72 rack pairs 72 Rubin GPUs with 36 Vera CPUs and packs rack-level memory figures including HBM4 and multi‑TB LPDDR capacities as a single liquid‑cooled unit. (videocardz.com)) NVIDIA’s technical briefings peg NVL72 at up to 3.6 exaFLOPS for NVFP4 inference and about 2.5 exaFLOPS for training at the rack level. (developer.nvidia.com)) The new NVLink 6 fabric boosts per‑GPU bi‑directional bandwidth to roughly 3.6 TB/s and scales to an aggregate ~260 TB/s within an NVL72 rack. (convergedigest.com)) NVIDIA’s rack diagrams show NVL72 using multiple NVLink switches (nine in the flagship configuration) to achieve that all‑to‑all, low‑latency topology. (videocardz.com)) NVIDIA is pairing the hardware with factory‑level software and operational designs — including a DSX AI Factory reference with dynamic power provisioning and “Max‑Q” strategies that NVIDIA says can yield roughly 30% more deployable AI infrastructure per fixed power budget. (siliconangle.com)) Operators such as CoreWeave intend to treat an NVL72 rack as a single programmable entity via Rack Lifecycle Controllers and Mission Control software to validate power, cooling and network readiness before live workloads. (convergedigest.com)) Hyperscalers and cloud providers listed by NVIDIA and partners expect NVL72 availability in the second half of 2026, with Microsoft reporting it has already “brought up” an NVL72 system in a lab and AWS committing to deploy over one million NVIDIA GPUs across Blackwell and Rubin families in the coming 12 months. (nvidianews.nvidia.com)) NVIDIA and partners say Rubin silicon entered full production in early 2026 while the NVL72 rack is fully liquid‑cooled and uses cable‑free modular trays that NVIDIA claims cut field installation time from about two hours to roughly five minutes. (datacenterdynamics.com)) For pod‑scale customers, NVIDIA’s DGX Vera Rubin SuperPOD blueprint stitches 14 NVL72 racks into systems topping ~50.4 exaFLOPS FP4 and more than 1,000 Rubin GPUs, and NVIDIA has announced integration of third‑party inference accelerators such as Groq 3 LPX into the Rubin ecosystem. (blogs.nvidia.com))
