Google Cloud A5X Vera Rubin Instances Go Live: 960,000 GPU Clusters Reshape Agentic AI Infrastructure

On October 21, 2026, Google Cloud officially announced the launch of its A5X bare-metal instances, powered by NVIDIA Vera Rubin NVL72 rack-scale systems, with general availability (GA) planned for November 2026. This infrastructure breakthrough supports multi-site deployments of up to 960,000 Rubin GPUs, designed specifically for large-scale agentic AI inference and reinforcement learning workloads, marking a new era in cloud computing infrastructure.

Technical Specifications: New Benchmark for Rack-Scale AI Computing

A5X instances represent a core component of Google Cloud's AI Hypercomputer, integrating compute, storage, and networking as a unified system.

Hardware Configuration

Each A5X instance utilizes the NVIDIA Vera Rubin NVL72 rack-scale system, containing:

• 72 Rubin GPUs: NVIDIA's latest generation AI accelerators
• 36 Vera CPUs: ARM-architecture processors optimized for AI workloads
• NVLink 6 Interconnects: Providing ultra-high bandwidth GPU-to-GPU communication
• ConnectX-9 SuperNICs: High-performance network interface cards
• BlueField-4 DPUs: Data processing units offloading network and storage tasks

Network Architecture: Virgo Network

The Google Cloud Virgo Network uses a non-blocking mesh topology supporting:

• Single site: Clusters of up to 80,000 GPUs
• Multi-site: Expansion to 960,000 GPUs for hyperscale deployments

This network architecture makes A5X one of the largest-scale AI computing platforms available in the cloud.

Storage and Memory

• Google Cloud Managed Lustre: With NVIDIA GPUDirect Storage, providing an RDMA path that bypasses the CPU
• Advanced KV-caching: Expanding context memory beyond GPU high-bandwidth memory (HBM) for long-context workloads
• Fourth-generation advanced liquid cooling: Thermal management for high-performance environments

First Customers: Top AI Research Institutions' Choice

Ineffable Intelligence: AlphaGo Creator's New Journey

Ineffable Intelligence was founded by David Silver, creator of AlphaGo and a leading researcher in reinforcement learning. This London-based startup became the first customer of the A5X platform, using its infrastructure for large-scale reinforcement learning to develop "superlearner" systems that improve continuously through experience.

David Silver's choice of A5X centers on its optimization for reinforcement learning workloads: the 960,000 GPU scaling capability makes training ultra-large-scale reinforcement learning models possible, which is critical for developing truly general artificial intelligence.

WPP: Enterprise-Grade Creative Agent Applications

WPP, the world's largest advertising and communications group, demonstrated a creative agent reference architecture on A5X instances. Using NVIDIA Nemotron models and the NVIDIA OpenShell architecture, WPP's agents automate multi-step creative workflows — including video generation and upscaling — within governed, secure sandboxes.

This use case demonstrates A5X's practical value for enterprise-grade agentic AI workloads: not only providing raw computing power, but also protecting creative asset intellectual property through confidential computing.

Security and Sovereign AI: New Standards for Confidential Computing

Google Cloud positions the A5X platform as a secure foundation for mission-critical agentic workloads:

Confidential Computing

Google Cloud is among the first to offer rack-scale NVIDIA Confidential Computing on bare-metal instances, including:

• Hardware-rooted secure boot: Ensuring system integrity
• Remote attestation: Verifying the trustworthiness of computing environments
• Encryption in transit: Protecting data and model intellectual property

European Sovereign AI

Google Cloud is expanding its European footprint to support data residency and sovereignty requirements, including the use of Google Cloud Data Boundary and local independent partners to maintain operational control over sensitive AI deployments.

Comparison with Competitors

A5X has clear advantages in cluster scale and confidential computing, particularly suited for workloads requiring ultra-large-scale training and high security.

Asia-Pacific Perspective: Strategic Significance of AI Infrastructure

For AI enterprises and research institutions across Asia-Pacific, the launch of Google Cloud A5X has important strategic significance:

Research Institutions: Top AI research institutions in Asia-Pacific (such as NUS, University of Tokyo, Tsinghua University) can access world-class AI computing resources through Google Cloud without building expensive GPU clusters.

Enterprise AI Deployment: For Asia-Pacific enterprises requiring large-scale agentic AI inference, A5X provides on-demand scalable computing, particularly suited for security-sensitive industries like financial services, manufacturing, and healthcare.

Sovereign AI Considerations: As Asia-Pacific governments strengthen oversight of AI infrastructure, Google Cloud's confidential computing and data sovereignty features will become important factors in enterprise cloud provider selection.

Cost Efficiency: Compared to building GPU clusters in-house, using A5X bare-metal instances can significantly reduce upfront capital expenditure while gaining the performance advantages of the latest generation hardware.

Outlook: The Future of Agentic AI Infrastructure

The launch of A5X marks cloud computing infrastructure entering the "agentic AI era." As AI agents evolve from simple Q&A assistants to autonomous systems capable of executing long-running, multi-step workflows, demands on computing infrastructure are rapidly escalating:

• Long-running tasks: Agent tasks may run for hours or even days, requiring stable large-scale computing resources
• Reinforcement learning: Training agents capable of autonomous improvement requires massive computation
• Multi-agent collaboration: Coordinating large-scale agent clusters requires high-bandwidth, low-latency network infrastructure

Google Cloud A5X's 960,000 GPU scaling capability and Virgo network architecture are purpose-built solutions for these requirements.