Nvidia and Google Cloud plan November availability of Vera Rubin A5X

Nvidia and Google Cloud said general availability of A5X instances built on Nvidia’s Vera Rubin NVL72 will come next month, with Europe to follow. For AI buyers, the release sets a first public date for Rubin-generation capacity on a hyperscale cloud and a ceiling of 960,000 GPUs across multiple sites.

• A5X instances combine Vera Rubin NVL72, ConnectX-9 SuperNICs, Google’s Virgo Network and fourth-generation liquid cooling.
• Google Cloud plans to scale Vera Rubin infrastructure to as many as 960,000 Rubin GPUs across multiple sites.
• Watch for regional availability, pricing and a European launch date, none of which the release gives.

• 960,000 Rubin GPUs Google Cloud plans across multiple sites; no timeline given
• Nov 2026 planned general availability, one month after today’s October 21 announcement
• 1 A5X instance ran WPP workloads that usually need a multi-node cluster
• 3 named Confidential G4 customers, versus one named A5X launch customer

Nvidia published the release. Its opening line says “General availability of the instances is coming next month, with planned expansion of availability to Europe”. The text places general availability in the future, so the instances are not yet generally available.

A5X is positioned for agentic inference and large-scale reinforcement learning, and it joins Google Cloud’s AI Hypercomputer as an integrated system of compute, networking and storage.

First customers and what WPP’s test does show

Ineffable Intelligence, a London-based company founded by AlphaGo creator David Silver, was the first customer to access A5X. It is using the platform for reinforcement learning research aimed at “superlearner” systems that learn continuously from experience. Silver said Vera Rubin’s connectivity between the Vera CPU and Rubin GPUs, and between GPUs within the rack, lets his team push its research further.

WPP, the London-based marketing group, demonstrated a creative agent reference architecture on A5X using Nvidia Nemotron models and Nvidia OpenShell. In its initial tests, workloads that typically required a multi-node cluster ran on a single A5X instance. The release gives no throughput, cost or latency figures, and it does not say how large the original cluster was.

Confidential computing extends to Blackwell hardware

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 runs on GB300 NVL72. It is also making Confidential Computing generally available on G4 VMs powered by Nvidia RTX PRO 6000 Blackwell Server Edition GPUs.

Vertebrae, Polaris Cloud and Confidential Core AI are using Confidential G4 for conversational privacy, verifiable agent execution and policy-governed enterprise intelligence. The release does not give a date for rack-scale confidential computing on A4X Max.

What changes for buyers, and what is missing

For AI teams, the practical change is a dated path to Rubin-class rack-scale capacity inside a major cloud, rather than a roadmap slide. The WPP result, if it holds beyond initial tests, would mean some agent workloads need fewer nodes to coordinate. The confidential computing options matter to buyers who, in the release’s words, run on infrastructure they do not own.

Google Cloud plans to scale its Vera Rubin-based infrastructure to as many as 960,000 Rubin GPUs across multiple sites.

Nvidia release, October 21, 2026

The 960,000 figure is a plan with no date attached. If each rack follows the NVL72 name and holds 72 GPUs, it implies about 13,333 racks, though the release does not state a per-rack count for A5X. The release also does not say how many GPUs are available at launch, in which regions, at what price, or when Europe goes live.

On power, the companies say only that they are “exploring” Nvidia DSX MaxLPS capabilities to improve performance per watt and expand capacity within data-center power budgets. That wording describes work in progress, not a delivered feature.

The BlockWest read. We read the 960,000 figure as a statement about demand and power, not a delivery schedule. The release gives a launch month, a ceiling and no ramp, so the November instances are the only capacity customers can plan around. Power budgets, which the DSX MaxLPS work targets, will decide how fast the ceiling is approached. Buyers outside early research accounts should ask for allocation terms, not benchmark claims, before committing to Rubin-based training plans.

Next: general availability of A5X is due in November 2026, and Google Cloud and Nvidia have yet to state regions, pricing or a date for the European expansion.

BlockWest is a news publication. Nothing here is investment advice. Read our disclaimer and editorial policy.