Google Cloud and NVIDIA announce A5X instances with Vera Rubin NVL72 infrastructure
Google Cloud and NVIDIA announced on October 21, 2026, at NVIDIA GTC Berlin that Google Cloud A5X instances powered by NVIDIA Vera Rubin NVL72 rack-scale systems will achieve general availability in November 2026. The announcement marks the commercial cloud debut of the Vera Rubin GPU architecture, designed specifically to address the massive compute and low-latency interconnect demands of agentic multi-step inference, continuous reinforcement learning loops, and foundational world models.
The A5X infrastructure combines Vera Rubin NVL72 compute racks, NVIDIA ConnectX-9 SuperNICs, Google Cloud's high-throughput Virgo Network fabric, and Google's fourth-generation facility-level liquid cooling systems. Google Cloud outlined architectural plans to scale the deployment across multiple data center regions to encompass up to 960,000 Rubin GPUs.
The integrated architecture targets experience-driven learning, where autonomous agent swarms iteratively interact with software environments, evaluate execution feedback, and update underlying weights in real time. Early pilot deployments include Ineffable Intelligence, an AI research lab founded by AlphaGo creator David Silver utilizing A5X for large-scale reinforcement learning, and advertising conglomerate WPP, which demonstrated creative video generation and agentic editing pipelines on a single A5X instance using NVIDIA Nemotron models within governed OpenShell sandboxes.
For enterprise CIOs, cloud infrastructure architects, and frontier AI research directors, the launch signals an inflection in cloud compute density. The NVL72 rack-level interconnect treats 72 GPUs as a unified computational fabric, drastically reducing inter-node communication latency during complex multi-agent simulations and agent fine-tuning compared to traditional multi-chassis clusters. Google Cloud also confirmed concurrent support for NVIDIA Confidential Computing across bare-metal A4X Max instances and general availability on G4 virtual machines.
However, technology procurement teams must consider significant commercial and operational caveats. The announcement establishes a forward-looking November 2026 availability window rather than immediate access, with initial deployments centered in North America before expanding to European regions.
Furthermore, neither Google Cloud nor NVIDIA has published hourly instance pricing, minimum reserved capacity commitment terms, or AI Hypercomputer networking overage rates. Given the extreme power density and specialized liquid-cooling requirements of NVL72 deployments, production access will likely require negotiated enterprise contracts with significant annual spending thresholds.