NVIDIA unveils Vera Rubin AI factories
NVIDIA rolled out Vera Rubin — a bundled platform of CPUs, GPUs, networking and orchestration aimed at building end‑to‑end “AI factories” for agentic, industrial AI; the reveal also surfaced hardware pushes for orbital AI data centers. The talk centered on composable pipelines, liquid cooling and enterprise control, signaling that AI is now sold as an integrated systems play, not just chips. ( )
NVIDIA described Vera Rubin as a multi‑chip platform that now includes seven new co‑designed parts — Rubin GPUs, Vera CPUs, an NVLink 6 switch, ConnectX‑9 SuperNICs, BlueField‑4 DPUs, Spectrum‑6 switching, and an integrated Groq 3 LPU. (nvidianews.nvidia.com) The NVL72 rack‑scale unit—marketed as a single accelerator—bundles 72 Rubin GPUs and 36 Vera CPUs inside a coherent rack design rather than a collection of discrete servers. (nvidia.com) NVIDIA’s NVLink 6 fabric is specified at ~3.6 TB/s bidirectional bandwidth per GPU and the company cites roughly 260 TB/s of scale‑up GPU‑to‑GPU bandwidth inside an NVL72 rack. (videocardz.com) Following a December 2025 licensing deal, NVIDIA integrated Groq’s low‑latency LPU into Vera Rubin and outlined Groq 3 LPX inference racks that target very large‑context, real‑time agentic inference with hundreds of LPUs per POD. (datacenterknowledge.com) At GTC CEO Jensen Huang raised NVIDIA’s long‑range revenue outlook to about $1 trillion through 2027, explicitly linking that target to a surge in inference economics and demand for production‑grade agentic deployments. (datacenterknowledge.com) NVIDIA previewed a Space‑1 Vera Rubin module for orbital data centers, claiming up to ~25× more space‑qualified inferencing compute versus an H100‑class baseline and naming partners including Aetherflux, Axiom Space, Kepler, Planet Labs, Sophia Space and Starcloud. (investor.nvidia.com) NVIDIA said the Vera Rubin rollout will be delivered via an MGX ecosystem of more than 80 partners, and OEMs have already announced manufacturing scale‑up and fully liquid‑cooled 2U Rubin HGX NVL8 systems to serve as fieldable building blocks for rack‑scale deployments. (nvidianews.nvidia.com; prnewswire.com; asus.com)
