Google Cloud has launched its new A5X bare-metal instances, which utilize NVIDIA Vera Rubin NVL72 systems. Each rack within this infrastructure integrates 72 Rubin GPUs and 36 Vera CPUs, connected via NVLink 6 technology. To manage large-scale deployments, Google employs its proprietary Virgo networking fabric alongside NVIDIA ConnectX-9 SuperNICs.
The company aims to reach a total capacity of 960,000 GPUs across multiple sites. According to Google, these instances provide up to 10 times lower inference costs per token and 10 times higher throughput per megawatt compared to previous generations. These figures represent a maximum capacity ceiling rather than currently active hardware.
The first A5X instances went live on October 21, 2026, with general availability scheduled for November 2026. London-based Ineffable Intelligence is the first customer, utilizing the infrastructure for reinforcement learning tasks. Google has also announced plans for a future expansion into Europe, though specific locations and timelines remain undisclosed.
