Google Cloud and NVIDIA Launch Vera Rubin-Powered A5X Instances for Scalable AI Training and Inference

Image: NVIDIA · Source

Google Cloud and NVIDIA announce the upcoming general availability of A5X instances based on NVIDIA Vera Rubin infrastructure, delivering integrated high-performance computing for training, inference, and reinforcement learning workloads with advanced liquid cooling and secure AI agent support.

Google Cloud and NVIDIA have jointly announced the imminent general availability of the A5X cloud instances powered by NVIDIA's Vera Rubin NVL72 architecture, designed to transform large-scale AI workloads. Scheduled for release next month, these A5X instances combine NVIDIA Vera Rubin GPUs, ConnectX-9 SuperNICs for high-speed networking, Google Cloud's Virgo Network fabric, and a fourth-generation advanced liquid-cooling system. This infrastructure is crafted to meet the demands of agentic inference, reinforcement learning, and AI training at unprecedented scale and efficiency.

The Vera Rubin-based platform will initially be available in select regions, with planned expansion to Europe, aiming to scale up to 960,000 Rubin GPUs across multiple data centers. Ineffable Intelligence, the London-based company founded by AlphaGo creator David Silver, is the first adopter, leveraging the A5X instances to develop continuous learning "superlearner" AI systems. These systems require vast compute resources to generate experiences, evaluate outcomes, and iteratively update models, a cycle supported end-to-end by the integrated infrastructure of computing, networking, and storage.

In addition to high compute performance, the A5X instances incorporate NVIDIA's OpenShell Reference Architecture, which offers governance, isolation, and security for AI agents. WPP, a global marketing firm, has demonstrated creative workflows on the platform that manage conversational AI requests through multi-step video generation and editing, utilizing NVIDIA Nemotron models with OpenShell's sandboxed environment to protect sensitive assets. Such secure execution is vital for enterprises deploying proprietary models and data in shared cloud infrastructures.

Google Cloud is also advancing confidential computing capabilities with NVIDIA's Confidential Computing offerings on bare metal A4X Max instances and G4 VMs featuring RTX PRO 6000 Blackwell GPUs. This technology allows customers to run AI workloads with enhanced data privacy and security, supporting confidential inference and fine-tuning operations.

Complementing these developments, NVIDIA and Google Cloud are jointly exploring NVIDIA DSX MaxLPS software solutions to optimize energy efficiency and maximize AI throughput within data center power constraints. These efforts align with broader initiatives to enable sustainable AI factory operations that conserve energy and water through innovative cooling solutions.

Together, the launch of NVIDIA Vera Rubin infrastructure in Google Cloud's A5X instances presents a comprehensive, high-performance, and secure platform for organizations pushing the frontier of AI training, inference, and agentic learning at scale.