Google Cloud’s Vera Rubin A5X instances arrive next month, Europe to follow

Google Cloud and Nvidia said in an announcement published on October 21 that A5X instances, built on Nvidia Vera Rubin NVL72, will become generally available next month, with expansion to Europe planned.

The post is tied to Nvidia’s GTC Berlin event. It describes A5X as an integrated platform for training, inference and large-scale reinforcement learning, aimed at agentic workloads.

What an A5X instance contains and where it will run

According to Nvidia, A5X combines four elements: Vera Rubin NVL72 racks, ConnectX-9 SuperNICs, Google Cloud’s Virgo Network and Google Cloud’s fourth-generation liquid cooling. The instances sit inside Google Cloud’s AI Hypercomputer, which bundles compute, networking and storage.

Looking further out, Google Cloud intends to grow its Vera Rubin fleet to a ceiling of 960,000 Rubin GPUs spread over several sites. The announcement gives no price, no region list and no exact date beyond “next month”, and says only that availability is planned to expand to Europe.

The first customer to get access was Ineffable Intelligence, a London-based company founded by AlphaGo creator David Silver. It is using A5X for reinforcement learning research on what Silver calls “superlearner” systems. In the announcement he said the work demands “both substantial compute and a tight connection between generating experience and training on it”.

WPP, the London-based marketing group, also demonstrated a creative agent architecture on A5X using Nemotron models and OpenShell, Nvidia’s sandboxing reference for safe agents. The agents turn conversational requests into multistep jobs such as video generation, upscaling and restyling, with human approval before anything runs. In WPP’s initial tests, work that normally needed a multi-node cluster ran on one A5X instance. That figure comes from the announcement and has not been independently checked.

Confidential computing moves to rack scale

The announcement also covers security. Google Cloud says it will be one of the earliest clouds to run Nvidia’s confidential computing across a whole rack on bare-metal machines, with A4X Max, which runs GB300 NVL72 systems, as the example. Separately, Confidential Computing is now generally available on G4 VMs powered by RTX PRO 6000 Blackwell Server Edition GPUs, aimed at confidential inference and fine-tuning. Vertebrae, Polaris Cloud and Confidential Core AI are named as users.

For readers who can use it today, the practical split is clear: the Blackwell-based G4 option is available now, A4X Max rack-scale confidentiality is described as coming, and A5X itself arrives next month.

Google Cloud and Nvidia are also exploring Nvidia DSX MaxLPS for Vera Rubin infrastructure, meant to improve performance per watt within a data center’s power budget. No results are given.

The Rubin timeline is now concrete

A5X is not new as a name. According to ITWire, which covered the April 23 announcement ahead of Google Cloud Next, Google presented A5X as bare-metal instances on Vera Rubin NVL72 and said a single-site cluster could reach up to 80,000 Rubin GPUs, with multisite clusters reaching 960,000. That report also repeated Nvidia’s claim of up to 10x lower inference cost per token and up to 10x higher token throughput per megawatt than the prior generation. Those are vendor figures, not benchmarks.

What the October post adds is a timetable and a first named customer. Six months after the preview, Google is committing to general availability within weeks, and to Europe afterward, which matters for customers with data residency needs. Nvidia is pursuing the same European audience elsewhere at GTC Berlin: its DSX POD offer for building AI factories in containers targets sovereign capacity, and its OpenShell blueprint is the agent-safety layer WPP used here.

The pattern is that Google keeps offering Nvidia’s newest racks alongside its own TPUs. Last October, Google said its A4X Max instances with GB300 NVL72 were shipping in production, according to DatacenterDynamics. That platform carried 72 Blackwell Ultra GPUs and 36 Grace CPUs per rack. A5X continues that rack-scale line one generation later.

What the announcement leaves open

Nvidia’s post does not give pricing, instance sizes, regions, or the European launch date. It also does not say how many of the 960,000 GPUs are installed or when they will be, and calls it a plan. The WPP result is a demonstration on a reference architecture, not a production benchmark. Buyers will need Google Cloud’s own documentation at launch to compare cost per token against A4X Max.