Google Cloud expected its A5X instances, built around NVIDIA Vera Rubin NVL72, to reach general availability in November 2026, according to NVIDIA’s October 21 announcement. The announcement also described plans to expand availability to Europe.

What the A5X system combines

System areaNamed componentAnnounced role
ComputeNVIDIA Vera Rubin NVL72Compute system
NetworkingNVIDIA ConnectX-9 SuperNICs and Google Cloud Virgo NetworkNetworking
CoolingGoogle Cloud fourth-generation advanced liquid coolingLiquid cooling

NVIDIA presented A5X for training, inference and large-scale reinforcement learning. The announced system brings NVIDIA’s compute and networking components together with Google Cloud’s network and cooling technology.

Early A5X work spans research and creative agents

NVIDIA identified Ineffable Intelligence, founded by CEO David Silver, as A5X’s first customer to gain access. The company is using the platform for large-scale reinforcement-learning research.

WPP demonstrated a creative-agent system on A5X using NVIDIA Nemotron models and OpenShell. WPP said initial tests ran workloads that typically required a multi-node cluster on a single A5X instance. Its workflow covered video generation, upscaling and restyling, with human approval before execution. OpenShell provides governed sandboxes to control agents’ access to creative tools and sensitive brand assets.

What the Vera Rubin performance figures measure

An earlier NVIDIA comparison for Vera Rubin NVL72 claimed 10× lower cost per token for mixture-of-experts (MoE) inference and 4× fewer GPUs to train MoE models than the prior Grace-Blackwell NVL72 generation. Those are platform comparisons, not Google Cloud A5X benchmarks. Using fewer GPUs does not by itself establish a proportional reduction in total system cost.

Google Cloud’s scale plan and NVIDIA’s DSX claim

NVIDIA’s DSX reference architecture for AI factories, a broader context than Google Cloud A5X

Google Cloud said it plans to scale Vera Rubin infrastructure to as many as 960,000 GPUs across multiple sites. The announcement also said Google Cloud and NVIDIA are exploring DSX MaxLPS for Vera Rubin infrastructure.

Separately, NVIDIA says its DSX Max LPS approach uses 30% less power for the same output. That claim concerns NVIDIA’s broader AI-factory architecture, not A5X performance.