Google Cloud and NVIDIA announced Google Cloud A5X, bare-metal instances built on NVIDIA Vera Rubin NVL72 racks, at GTC Berlin on October 21, 2026. Despite the headline of NVIDIA’s post, the instances are not yet generally available: both companies say general availability is “coming next month”, with availability in Europe planned after that.
A5X puts NVIDIA’s rack-scale system, 72 Rubin GPUs and 36 Vera CPUs per NVL72 rack, inside Google’s own networking, storage and cooling, and Google attaches a deployment ceiling to it: as many as 960,000 Rubin GPUs across multiple sites.
On this page
What A5X combines
According to NVIDIA, each A5X deployment brings together four parts:
| Component | Supplier | Role, as described by the companies |
|---|---|---|
| Vera Rubin NVL72 | NVIDIA | Rack-scale system of Rubin GPUs and Vera CPUs |
| ConnectX-9 SuperNIC | NVIDIA | Network interface for each node |
| Virgo Network | Google Cloud | Non-blocking scale-out fabric within and between buildings |
| Fourth-generation liquid cooling | Google Cloud | Cooling for the racks |
Google Cloud’s own GTC Berlin post, by Mark Lohmeyer and Ankur Mehrotra, gives the scale numbers. Virgo Network uses a mesh topology that Google says lets a cluster grow to 80,000 Rubin GPUs within a single site and up to 960,000 Rubin GPUs across multi-site deployments. These are fabric limits and a stated plan, not a count of GPUs installed today; neither company gives a delivery schedule for the full figure.
Google lists three more pieces around the instances. Google Cloud Managed Lustre will support NVIDIA GPUDirect Storage on A5X, giving data a direct RDMA path from the file system into GPU memory without passing through the CPU. A5X will support KV caching that extends context memory beyond the GPUs’ HBM to other storage tiers, aimed at long-context and multi-turn agent workloads; what a KV cache stores is covered in our KV cache explainer. And Google Kubernetes Engine supports A5X with Managed DRANET, which exposes RDMA network cards as Kubernetes resources, and Kueue for topology-aware scheduling.
Availability: next month, then Europe
The general-availability date is given only as “next month” in both posts. Google says that alongside general availability it is expanding its footprint to meet EMEA data-residency and sovereignty requirements with new local regions; NVIDIA describes this as “planned expansion of availability to Europe”. Neither post names the regions or a date for them, and neither publishes A5X pricing.
Early users named
NVIDIA names two companies that have used A5X before general availability:
- Ineffable Intelligence, the London company founded by AlphaGo creator David Silver, is described as the first customer to access A5X. It uses the instances for large-scale reinforcement learning research.
- WPP demonstrated its creative-agent reference architecture on A5X with NVIDIA Nemotron models and NVIDIA OpenShell. NVIDIA says that in WPP’s initial tests, workloads that usually needed a multi-node cluster ran on a single A5X instance; that is WPP’s and NVIDIA’s own result, with no published test detail.
Confidential computing on Blackwell instances
The same announcement covers confidential computing on Google’s current NVIDIA instances. NVIDIA says Google Cloud will be among the first cloud providers to offer NVIDIA Confidential Computing at rack scale on bare-metal instances such as A4X Max, which run on GB300 NVL72. Separately, Google has made Confidential Computing generally available on G4 VMs with NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. The companies also say they are exploring NVIDIA DSX MaxLPS for Vera Rubin infrastructure to raise performance per watt within data-center power budgets.
Where this fits
CoreWeave announced on September 30 that Vera Rubin NVL72 was available on its cloud, with Cognition as its first production customer; our CoreWeave report has the details. The first independent agent-workload results for the rack came from Artificial Analysis on October 15, covered in our AgentPerf report. Vera Rubin NVL72’s published specifications are on our Vera Rubin NVL72 page in the AI chips table, and the platform’s background is in our Nvidia Vera Rubin explainer. Venues and the programme of the event where A5X was announced are in our GTC Berlin preview.
