Google Cloud plans to deploy up to 960,000 GPUs through a new family of A5X instances built on NVIDIA Vera Rubin NVL72 systems. Each bare-metal instance provides direct access to a complete rack containing 72 Rubin GPUs and 36 Vera CPUs, connected with NVLink 6 and linked across multiple sites through Google’s Virgo networking fabric and NVIDIA ConnectX-9 SuperNICs. Google claims up to ten times lower inference cost per token and up to ten times higher throughput per megawatt than the previous generation. The first A5X instances launched on October 21, 2026, while general availability is planned for November 2026. Ineffable Intelligence, a London-based company, is the first customer and is using them to run complex reinforcement learning systems, with a European expansion yet to be given specific sites or dates.
Source: Read the original article
