Google Plans Up to 960,000 Rubin GPUs; AlphaGo Creator First

Google Cloud plans to scale its Nvidia Vera Rubin infrastructure “to as many as 960,000 Rubin GPUs across multiple sites,” Nvidia said in a blog post at GTC Berlin on 21 October 2026. The post launches Google Cloud’s A5X instances, built on Nvidia’s Vera Rubin NVL72 systems, and says the first customer to access them was Ineffable Intelligence, the London company founded by AlphaGo creator David Silver.

The post’s headline announces general availability, but its text says “General availability of the instances is coming next month, with planned expansion of availability to Europe.” In plain terms, general availability is the point at which any paying cloud customer can rent a product, rather than a small group given early access; Nvidia’s post dates that to next month.

Business Pill · GPUS

A short Business Engineer explainer on why AI runs on GPUs. It teaches the general idea only and says nothing about Google Cloud or Nvidia’s A5X.

Watch · The Keynote

NVIDIA GTC Berlin 2026 keynote with CEO Jensen Huang, 21 October, on NVIDIA’s YouTube channel. Facts and quotations in this piece come from Nvidia’s blog post, not from the spoken keynote.

What Google Cloud Is Offering

• The machine. Nvidia says A5X instances “bring together NVIDIA Vera Rubin NVL72, NVIDIA ConnectX-9 SuperNICs, Google Cloud Virgo Network and Google Cloud’s fourth-generation advanced liquid-cooling technology for demanding agentic inference and reinforcement learning workloads at scale.”
• The scale. As many as 960,000 Rubin GPUs, across multiple sites, according to the post. It gives no timeline for reaching that number and no list of sites.
• The timing. General availability next month, with Europe to follow, per the post.
• The price. The post gives no prices for A5X.

Nvidia says A5X brings Vera Rubin into Google Cloud’s AI Hypercomputer “as an integrated system, combining compute, networking and storage to support experience-driven learning at scale.”

Who Is Using It First

Ineffable Intelligence “was the first customer to access A5X,” according to Nvidia, and is using it to advance large-scale reinforcement learning and shape “superlearner” systems that learn continuously from experience.

“We’re building superlearners that continually improve through their own experience,” Silver, Ineffable’s founder and chief executive, said in the post. “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.”

WPP, the London-based marketing and communications company, has demonstrated its creative-agent reference architecture on A5X using Nvidia’s Nemotron models and OpenShell, the post says, with agents that turn requests into multistep creative workflows such as video generation “with human approval before execution.” Nvidia adds: “In WPP’s initial tests, workloads that typically required a multi-node cluster ran on a single A5X instance.”

The Security and Power Pieces

The post 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 instances powered by NVIDIA GB300 NVL72.” The post describes the need it answers as protecting proprietary models and sensitive data “even when running on infrastructure they don’t own.”

Separately, Nvidia says Google Cloud is making Confidential Computing generally available on G4 VMs powered by Nvidia RTX PRO 6000 Blackwell Server Edition GPUs, with customers including Vertebrae, Polaris Cloud and Confidential Core AI. On power, the post says Google Cloud and Nvidia “are exploring” Nvidia’s DSX MaxLPS capabilities for Vera Rubin-based infrastructure, to help expand capacity within available data-centre power budgets.

Connecting the Coverage

Our Vera Rubin piece reported earlier today that Nvidia’s release puts Vera Rubin NVL72 at “30x higher throughput per megawatt” for agentic workloads; A5X is Google Cloud’s product built on those systems. Our DSX POD piece covered Nvidia’s claim the same day that factory-built Vera Rubin units cut deployment at a prepared site “from six months to as little as 14 days.”

Our supply-chain piece reported that Nvidia is giving KYEC a $1.1 billion capacity reservation advance, under a release framed by Jensen Huang’s line “Demand is extraordinary. Capacity is the constraint.” Our Pichai piece reported on 21 October that Sundar Pichai said Google is “capacity constrained” serving its newest Gemini model.

The A5X post came a day after Google pushed its own chips. Our TorchTPU piece reported on 20 October that Google open-sourced TorchTPU, its native PyTorch stack for its own TPU chips, a move TechCrunch framed as Google joining Amazon in trying to break Nvidia’s hold on AI developers. And our UK startups piece reported on 15 October that Nvidia’s own post listed Ineffable Intelligence among the UK startups it has supported.

The Structural Read

As we read it, A5X is sold for a specific kind of work: Nvidia describes it as built for agentic inference and reinforcement learning, and both named customers, Ineffable and WPP, are running agent or learning-from-experience workloads rather than plain chatbots.

As we read it, the post is as much about trust and power as about chips: it pairs A5X with Confidential Computing on Google Cloud’s bare-metal GB300 instances and with Nvidia’s DSX MaxLPS capabilities, which the post says the two companies are exploring.

As we read it, Google is backing both its own chips and Nvidia’s: Nvidia’s post describes Google Cloud’s plan for Vera Rubin at scale a day after Google open-sourced TorchTPU for its own TPU chips, and the post gives no price for A5X or value for the build.

Three Implications

FOR AI LABS RENTING COMPUTE Nvidia says A5X general availability is coming next month, with Europe to follow, and describes the instances as built for agentic inference and reinforcement learning; the post does not give prices.

FOR ENTERPRISES BUILDING AGENTS Nvidia says WPP ran on a single A5X instance workloads that typically required a multi-node cluster, and that Google Cloud will be among the first clouds to offer Confidential Computing at rack scale on bare metal.

FOR NVIDIA AND ITS RIVALS Nvidia’s post puts a figure of as many as 960,000 Rubin GPUs on one cloud’s plan, while Google keeps building its own TPU chips; the post does not say how Google will split future capacity between the two.

The Business Engineer Lens

This story maps onto the Business Engineer framework The AI Capex Map.

The framework argues: “This is not a continuation of the cloud capex cycle. This is a different cycle, with different physics.” It describes an AI stack “governed from the bottom — energy, silicon, compute capacity.”

As we read it, a plan for as many as 960,000 Rubin GPUs is that bottom layer in a single cloud’s build, and the post’s focus on power budgets and cooling fits the same frame. The framework was published in May 2026 and does not discuss this announcement.

What Is Not Established

Nvidia’s post, as we read it on 21 October, gives no price for A5X, no date or site list for the 960,000-GPU plan, no number of GPUs Ineffable is using and no value for any customer agreement. It does not say whether Google Cloud is the first cloud to offer Vera Rubin. Google Cloud’s own post, linked from Nvidia’s, returned a not-found page when we checked, so this piece relies on Nvidia’s account. The icon count in our chart is our arithmetic.

The Bottom Line

Nvidia used GTC Berlin to put a number on Google Cloud’s Vera Rubin build: as many as 960,000 Rubin GPUs across multiple sites, sold as A5X instances from next month, with David Silver’s Ineffable Intelligence as the first customer. As we read it, Google is lining up Nvidia’s newest systems at scale a day after it opened its own TPU chips to PyTorch developers.