Google Cloud says its A5X Vera Rubin instances reach general availability next month

What Google Cloud and NVIDIA announced

Google Cloud's first instances built on NVIDIA's Vera Rubin platform will be generally available next month. In a post on NVIDIA's blog on 21 October 2026, published during GTC Berlin, the two companies said A5X instances, built on NVIDIA's Vera Rubin NVL72 rack-scale systems, will reach general availability in November, with availability planned to expand to Europe afterwards.

The post's headline says Google Cloud and NVIDIA "announce general availability", but the text itself is more careful: general availability "is coming next month." Until then, access has been limited to early customers, the first of which was Ineffable Intelligence, the London reinforcement learning lab founded by AlphaGo creator David Silver.

The same post includes a longer-term number that is easy to miss: Google Cloud plans to scale its Vera Rubin-based infrastructure to as many as 960,000 Rubin GPUs across multiple sites. It also covers confidential computing on several Google Cloud GPU products and an agent demonstration by the advertising group WPP.

What an A5X instance is

Vera Rubin is NVIDIA's next-generation data-centre platform after Blackwell. Its NVL72 configuration joins 72 Rubin GPUs and NVIDIA's Vera CPUs into one liquid-cooled rack that behaves like a single large accelerator, the same rack-scale approach NVIDIA described in its open MGX rack specification and in this week's containerised DSX POD design.

Google Cloud's A5X instance wraps that rack in its own infrastructure. According to the NVIDIA post, A5X combines Vera Rubin NVL72 with NVIDIA ConnectX-9 SuperNIC network cards, Google Cloud's Virgo network and its fourth-generation liquid cooling. It is offered as part of AI Hypercomputer, Google Cloud's bundle of compute, networking and storage for large AI jobs.

The two companies pitch A5X at agentic inference, running AI agents at scale, and at reinforcement learning. Reinforcement learning is the training method in which a model generates attempts, is scored, and is updated, again and again. That loop needs fast connections between the machines producing experience and the ones training on it, which is why NVIDIA and Google stress bandwidth between the Vera CPU and the Rubin GPUs and across the rack.

Ineffable Intelligence goes first

Ineffable Intelligence was the first customer to access A5X. Silver's company, which Computer Weekly reported raised a $1.1 billion seed round in April at a $5.1 billion valuation with backing from Sequoia, Lightspeed and the UK's Sovereign AI Fund, picked Google Cloud as its infrastructure partner in June. DatacenterDynamics reported that it plans to deploy "one of the largest" clusters of A5X instances, without disclosing a GPU count.

Ineffable is building what it calls "superlearners", systems meant to learn continuously from their own experience rather than from fixed datasets. "That demands both substantial compute and a tight connection between generating experience and training on it, so each improvement in the agent can quickly shape its next experience," Silver said in the NVIDIA post.

Silver is one of several AlphaGo veterans now running their own labs. His former colleague Thore Graepel launched Metis Reasoning this week. Ineffable has said little in public about its models, so the A5X announcement offers one of the few concrete details about how it works: it is betting on large-scale reinforcement learning on NVIDIA hardware rented from Google.

WPP's agent test on a single instance

The post also describes a test by WPP, the London-based marketing group. WPP ran its creative agent reference architecture on A5X instances using NVIDIA's Nemotron open models and NVIDIA OpenShell, the sandboxing layer NVIDIA promotes for keeping AI agents within set permissions. The agents turn conversational requests into multistep creative jobs such as generating, upscaling and restyling video, and wait for human approval before running.

According to NVIDIA, in WPP's initial tests workloads that usually needed a multi-node cluster ran on a single A5X instance. NVIDIA did not say which workloads, how long they took or what they cost, so this is best read as an early vendor-reported result rather than a benchmark.

Confidential computing for model builders

A second part of the announcement is about protecting models and data on rented hardware. Google Cloud says it will be among the first cloud providers to offer NVIDIA Confidential Computing at rack scale on bare-metal instances, starting with A4X Max instances built on NVIDIA GB300 NVL72 systems.

Confidential computing uses hardware encryption so that data and model weights stay protected even while being processed, including from the cloud operator. In a separate NVIDIA post the same day, Google Cloud engineer Andrés Lagar-Cavilla said its Confidential Bare Metal offering gives AI labs "hardware-isolated, non-virtualized compute environments and end-to-end cryptographic protection." That matters for model builders that want to train or serve proprietary models on someone else's hardware.

Google Cloud is also making confidential computing generally available on its G4 virtual machines, which use NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. NVIDIA names Vertebrae, Polaris Cloud and Confidential Core AI as customers using Confidential G4 for private conversations, verifiable agent runs and policy-governed enterprise AI.

The 960,000-GPU plan

The largest number in the post is a plan, not a deployment. Google Cloud intends to scale its Vera Rubin infrastructure to as many as 960,000 Rubin GPUs across multiple sites. That works out to more than 13,000 NVL72 racks if all of it is in that configuration. The post gives no timeline, no list of sites and no split between A5X and other products.

The figure is a reminder that Google, despite building its own TPU chips and opening up TorchTPU for PyTorch users, still plans to buy NVIDIA hardware at very large scale for customers who want it. Google Cloud and NVIDIA also say they are exploring NVIDIA's DSX MaxLPS power tools for Vera Rubin, to get more performance per watt within the power limits of existing data centres.

What is missing from the announcement

Several details that customers will need are absent. The post gives no price for A5X, no list of launch regions beyond the plan to expand to Europe, no exact November date and no minimum commitment. It also does not say how many A5X racks Google Cloud has running today or how many Ineffable is using.

There is also a race for bragging rights. DatacenterDynamics notes that Microsoft and CoreWeave have each claimed to be first to deploy Vera Rubin. Being first to general availability in a public cloud is a different milestone, and November leaves room for rivals to announce their own dates.

Why it matters

For AI labs and large enterprises, general availability changes who can get Vera Rubin capacity. Early access has gone to a few hand-picked customers with very large budgets. Once A5X is generally available, any Google Cloud customer with the quota and budget can request it, which widens access to NVIDIA's newest platform beyond the handful of labs that negotiate directly with NVIDIA and the big clouds.

The rest of the announcement points to where demand is heading. Ineffable's interest is in reinforcement learning at scale, WPP's in agents that need sandboxes and approvals, and the confidential computing products in protecting model weights on shared hardware. Whether A5X delivers will depend on numbers Google Cloud has not yet shared: price, regions and how quickly the 960,000-GPU plan turns into real capacity.