Google Cloud scheduled A5X bare-metal instances, based on NVIDIA Vera Rubin NVL72, for general availability in November.

Google Cloud said A5X bare-metal GPU-accelerated instances would become generally available in November 2026. The platform combines NVIDIA Vera Rubin NVL72, ConnectX-9 SuperNICs, Google’s Virgo Network, and Google Cloud’s fourth-generation advanced liquid cooling, with regional expansion intended to support European and broader EMEA—Europe, Middle East, and Africa—data-residency requirements.

Google describes Virgo Network as a non-blocking mesh fabric spanning both intra-building and inter-building connections. It says the architecture will support clusters of 80,000 Rubin GPUs within one site and up to 960,000 Rubin GPUs across multiple sites; these are announced scaling capabilities, not disclosed installed GPU counts or customer commitments.

Managed Lustre is scheduled to support NVIDIA GPUDirect Storage on A5X, providing a remote direct memory access, or RDMA, path from the parallel file system into GPU memory while bypassing the central processing unit and system memory. A5X is also scheduled to support key-value caching beyond GPU high-bandwidth memory, with context offloaded and reused across storage tiers for long-context and multi-turn workloads.

Google Kubernetes Engine, or GKE, supplies orchestration features including Managed DRANET, which makes network interface cards native Kubernetes resources for RDMA management. The announcement also identifies Kueue for topology-aware scheduling and resource management, but provides no customer-specific deployment schedule, price, or benchmark for the complete A5X configuration.