Google Cloud unveils A5X on NVIDIA Vera Rubin, with general availability set for November

At NVIDIA's GTC Berlin event on October 21, Google Cloud and NVIDIA announced A5X, Google Cloud's first instances built on the NVIDIA Vera Rubin NVL72 rack-scale platform. They are not broadly available yet. Both companies' posts say general availability is "coming next month", which puts it in November 2026. NVIDIA's blog post uses "General Availability" in its headline, but the body text describes it as upcoming. One customer, London-based Ineffable Intelligence, already has access.

A5X matters less as a single SKU than as a signal of timing. It is the clearest date yet for when Vera Rubin capacity reaches a hyperscaler's public catalogue, and it comes with a networking roadmap that Google says can stretch a single training or reinforcement-learning job far beyond one data center.

What was announced

According to Google Cloud's post, signed by Mark Lohmeyer and Ankur Mehrotra, A5X instances are bare-metal and GPU-accelerated, built on the Vera Rubin NVL72 platform. NVIDIA adds that they combine Vera Rubin NVL72 with ConnectX-9 SuperNICs, Google's Virgo Network and Google Cloud's fourth-generation liquid cooling, and that they sit inside Google's AI Hypercomputer stack.

Neither company published pricing, per-instance memory or bandwidth figures, or a list of launch regions. Google says availability will expand into new local regions in EMEA to meet data residency and sovereignty requirements, but it did not name them.

The 960,000-GPU figure

The headline number is about networking, not shipping capacity. Google says its Virgo Network supports a non-blocking scale-out fabric, within and between buildings, that will let clusters grow to 80,000 Rubin GPUs at a single site and up to 960,000 Rubin GPUs across multi-site deployments. NVIDIA describes this as Google Cloud planning to scale its Vera Rubin infrastructure to "as many as 960,000 Rubin GPUs across multiple sites".

This is a stated plan and a capability of the fabric, not an installed fleet. Neither post gives a timeline for reaching that scale or says which sites are involved.

Storage, memory and orchestration

Google listed several pieces of software that will accompany A5X, all of which are aimed at keeping expensive GPUs busy rather than waiting on data:

• Managed Lustre with GPUDirect Storage on A5X, giving a direct RDMA path from the parallel file system into GPU memory that bypasses the CPU and system memory.
• KV-cache offloading that extends context memory beyond GPU high-bandwidth memory by storing and reusing it across storage tiers. Google says this cuts redundant computation and time-to-first-token, but it gave no figures.
• GKE Managed DRANET, which exposes network cards as native Kubernetes resources, and Kueue for topology-aware scheduling.

Early users

NVIDIA says Ineffable Intelligence, the company founded by AlphaGo creator David Silver, was the first customer to access A5X and is using it for large-scale reinforcement learning. Silver said the work needs "a tight connection between generating experience and training on it", and pointed to the bandwidth between the Vera CPU and Rubin GPUs and among GPUs in the rack.

WPP demonstrated a creative-agent reference architecture on A5X using NVIDIA Nemotron models and NVIDIA OpenShell, the secure agent runtime that NVIDIA also expanded at GTC Berlin. In NVIDIA's account, workloads that typically needed a multi-node cluster ran on a single A5X instance in WPP's initial tests. That is a vendor-reported result from a single customer, with no published methodology.

Confidential computing and on-premises

The same announcement covered confidential computing across several Google Cloud instance families:

• G4 VMs with NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs now have Confidential Computing generally available. NVIDIA names Vertebrae, Polaris Cloud and Confidential Core AI as users.
• A4X Max bare-metal instances on GB300 NVL72 will get NVIDIA Confidential Bare Metal, which adds encrypted memory and NVLink interconnects plus hardware-rooted attestation. Google says it is "introducing" this but gave no availability date.
• Google Distributed Cloud now has NVIDIA HGX B300 generally available, and Google says Gemini 3.5 Flash runs on it inside air-gapped deployments.

The companies also said they are "exploring" NVIDIA DSX MaxLPS for Vera Rubin deployments to improve performance per watt within data-center power budgets. That is not a committed feature.

What to watch

The practical questions for buyers are the ones neither post answers: price per GPU-hour against Blackwell-generation A4X instances, which regions open first, and how much capacity is available to customers who are not anchor tenants. November's general availability announcement should be the place to look for those details.