Google Cloud Deploys Million-Scale Vera Rubin GPUs, A5X Instances Launch Next Month

Google Cloud and NVIDIA announced the expansion of their infrastructure based on the Vera Rubin platform across multiple sites, with plans to deploy up to 960,000 Rubin GPUs. The A5X compute instances, built on NVL72 rack-scale systems, will become generally available to all customers next month, marking a significant transition from planning to large-scale deployment of next-generation AI infrastructure.

Cross-verified reports indicate that A5X instances are optimized for agentic inference and reinforcement learning workloads, integrating ConnectX-9 SuperNICs, Virgo networking, and fourth-generation liquid cooling. Early tests show this architecture accelerates the "generate-score-retrain" loop via low-latency CPU-GPU links and supports massive cross-site scaling. Additionally, Google Cloud plans to offer rack-level confidential computing on bare-metal instances, becoming one of the first providers to do so, with DeepMind co-founder David Silver’s new company as the initial customer.

This deployment solidifies Google Cloud's leadership in next-gen AI infrastructure and establishes an end-to-end ecosystem through the full-stack agentic system architecture (NVL72, MGX, CMX, RTX). It lowers barriers for developers building complex AI agents, shifting industry focus from pure model training to high-frequency interactive agent applications and reshaping the global AI compute landscape.