Google Cloud’s A5X instances, built on NVIDIA Vera Rubin NVL72, are announced but not yet generally available. Google’s GTC Berlin post of 21 October 2026 says general availability is “coming next month,” which points to November 2026 but gives no exact date.
The scale figures are plans, not deployed capacity. ClueTrove read both posts on 22 October 2026 and tested nothing.
The short answer
- What it is: bare-metal GPU-accelerated instances on the NVIDIA Vera Rubin NVL72 platform.
- Availability: general availability “coming next month,” with no exact date and no regions named.
- Early access: Ineffable Intelligence was the first customer to access A5X.
- Scale (planned): up to 80,000 Rubin GPUs in a single site and up to 960,000 across multi-site deployments.
- Networking: Virgo Network, a non-blocking fabric within and between buildings, built on a mesh topology.
- Europe: Google says it is expanding its footprint with new local regions for EMEA data residency and sovereignty requirements, and it names none of them.
Why can the headline mislead?
NVIDIA’s blog post carries “General Availability” in its headline. Its body says general availability of the instances is coming next month.
Having a first customer with access is not the same as general availability. If you plan capacity, wait for Google’s own announcement with dates and regions.
What does the A5X stack include?
- Storage: Google Cloud Managed Lustre will support NVIDIA GPUDirect Storage on A5X, according to Google. That gives a direct RDMA path from the parallel file system into GPU memory, bypassing the CPU and system memory.
- KV caching: A5X will support advanced KV-caching to extend context memory beyond GPU memory. Google says this is meant to reduce redundant computation and lower time-to-first-token, without giving figures.
- Orchestration: GKE features such as Managed DRANET and Kueue with topology-aware scheduling.
How should you read the GPU numbers?
The 80,000 and 960,000 GPU figures are what Google says the network “will enable” and that it plans to scale to. They are a ceiling, not what is running today. Google’s post gives no benchmarks and no per-instance specifications.
The only performance statement in NVIDIA’s post comes from WPP. In initial tests, WPP reported that workloads that typically needed a multi-node cluster ran on a single A5X instance. That is a partner’s preliminary result.
Google’s post adds a WPP use case: its CTO, Stephan Pretorius, says WPP’s Creative and Production teams will run agents for frame-accurate quality checks and brand asset adaptation. That shows the intended workload, not a measured outcome.
What is still unknown?
- The exact general availability date and the regions: neither post gives them.
- Prices: neither post lists any.
- Claims of 10x lower inference cost that appear in some outlets: ClueTrove did not find them in the two primary posts, so we do not repeat them.
Who should act now, and who can wait?
Teams that train or serve very large models and need EMEA data residency should watch this most closely. Everyone else can wait for the general availability post.
- Watch Google Cloud’s own announcement for the general availability date, regions and prices.
- Treat the GPU scale figures as plans when you size a project.
- Do not rely on outlet claims of 10x lower inference cost until a primary post confirms them.
Questions and answers about Google Cloud A5X
Is A5X available to everyone today?
No. Google says general availability is coming next month. Ineffable Intelligence has early access.
Which GPUs does A5X use?
NVIDIA Vera Rubin NVL72, according to both posts.
How much does A5X cost?
Neither post lists prices, so no budget can be built from them yet.
Does it run on Kubernetes?
Yes. Google says A5X runs on GKE, with Managed DRANET and Kueue topology-aware scheduling.
Terms in plain language
- General availability: the stage when a service is open to all customers, not only early testers.
- Bare-metal: a server you use directly, without a virtualization layer in between.
- KV cache: stored intermediate data that lets a model reuse earlier work instead of recomputing it.
- Time-to-first-token: how long a model takes to start answering.
Sources
- Google Cloud: infrastructure at NVIDIA GTC Berlin 2026 (read 22 October 2026)
- NVIDIA blog: Google Cloud and NVIDIA announce Vera Rubin infrastructure (read 22 October 2026)
Facts read on 22 October 2026.
