Skip to content
SM Secure Blog

Google Cloud Launches A5X Instances with NVIDIA Vera Rubin NVL72 to Power Next-Generation Agentic AI Infrastructure

Sagar Majumdar

Sagar Majumdar

Google Cloud has officially launched the general availability of its A5X computing instances, powered by NVIDIA Vera Rubin NVL72 rack-scale systems. The release marks a major architectural milestone in hyperscale cloud computing, introducing the first publicly available cloud infrastructure specifically built around NVIDIA next-generation Rubin architecture. Engineered as an integral tier of the Google Cloud AI Hypercomputer platform, the new bare-metal instances are tailored to address the heavy computational and data delivery bottlenecks currently facing large-scale reinforcement learning and continuous agentic workloads. At the core of the A5X platform is a profound shift from individual accelerator boxes to fully unified rack-scale computing. Each Vera Rubin NVL72 system consolidates seventy-two graphics processing units and thirty-six central processing units into a singular liquid-cooled computing fabric. To interconnect these dense hardware clusters without throttling distributed training jobs, Google Cloud coupled the hardware with its proprietary Virgo scale-out network fabric and ConnectX-9 SuperNICs. The resulting network architecture is designed to scale up to eighty thousand Rubin accelerators within a single facility and up to nine hundred and sixty thousand across interconnected cloud campuses. Beyond raw compute density, the A5X infrastructure addresses the critical memory and input-output challenges that emerge when running long-context models and complex reasoning chains. A notable architectural advancement is the introduction of tiered key-value cache offloading, which enables systems to move active context data beyond physical high-bandwidth memory into ultra-fast secondary tiers. By caching and reusing conversational context across multi-turn reasoning loops, the platform significantly reduces redundant token computation, lowers latency to first token response, and maximises inference throughput per dollar for production deployments. Storage integration has also been fundamentally redesigned to eliminate input-output bottlenecks. Through Managed Lustre parallel storage supporting NVIDIA GPUDirect Storage, the architecture establishes a direct remote direct memory access path between the distributed file system and GPU memory. By completely bypassing the host processor and system memory during data transfers, data pipelines achieve sustained throughput required for massive checkpointing and real-time reinforcement learning parameter distribution. Orchestration is managed directly through Google Kubernetes Engine, which introduces native enterprise enhancements designed for high-density accelerator fabrics. Through Managed DRANET, network interface cards are treated as first-class native Kubernetes resources, streamlining the configuration of complex high-speed network topologies. Meanwhile, the integration of Kueue delivers topology-aware job scheduling, ensuring that distributed training and inference jobs are placed on physically adjacent nodes to minimise cross-rack latency and maximise hardware utilisation. The commercial rollout of A5X reflects how rapidly enterprise demand is transitioning from traditional batch model training toward persistent agentic systems that require continuous environment interaction. Early enterprise adopters, such as the Frontier Research Lab Ineffable Intelligence are already utilising A5X clusters to build self-improving foundation models that learn dynamically from ongoing experience. By co-designing specialised hardware, optical networking, high-speed storage, and container orchestration into a coherent operational fabric, Google Cloud is setting a new benchmark for scalable enterprise cloud infrastructure.

Comments

No comments yet. Be the first to add one.