The headline and the first sentence disagree
Nvidia published a post at GTC Berlin on 21 October titled “Google Cloud and NVIDIA Announce General Availability of NVIDIA Vera Rubin Infrastructure”. Its opening sentence reads: “General availability of the instances is coming next month, with planned expansion of availability to Europe.”
So the instances are not generally available. They will be, in November, with Europe to follow. That is a normal pre-announcement, and it is worth noting only because the title says otherwise, and because titles are what get quoted.
The product is the Google Cloud A5X instance. Nvidia describes it as combining Vera Rubin NVL72 systems, Nvidia ConnectX-9 SuperNICs, Google Cloud’s Virgo Network and Google Cloud’s fourth-generation liquid cooling, aimed at agentic inference and large-scale reinforcement learning. A5X brings Vera Rubin into Google Cloud’s AI Hypercomputer as an integrated system.
The 960,000 figure
The number that will travel furthest from this announcement is this one: “Google Cloud plans to scale its Vera Rubin-based infrastructure to as many as 960,000 Rubin GPUs across multiple sites.”
Read it carefully. It is a plan, it is a ceiling — “as many as” — and it is spread across multiple sites rather than concentrated in one cluster. Nvidia gives no date for reaching it and no figure for how many Rubin GPUs Google Cloud has installed today. A reader who takes 960,000 as a current deployment is reading a sentence that does not say that.
Who is using it
The named first customer is Ineffable Intelligence, a London company founded by David Silver, whom Nvidia identifies as the creator of AlphaGo. The company is using A5X for large-scale reinforcement learning and what it calls “superlearner” systems that learn continuously from experience.
“We’re building superlearners that continually improve through their own experience,” Silver said in the announcement. “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.”
The second named user is WPP, the London marketing group, which demonstrated a creative agent reference architecture on A5X using Nvidia’s Nemotron models and Nvidia OpenShell. The agents turn conversational requests into multistep creative workflows — video generation, upscaling, restyling — with human approval before execution. The result Nvidia reports is specific: “In WPP’s initial tests, workloads that typically required a multi-node cluster ran on a single A5X instance.”
That is a vendor-reported figure from an early test by a partner, not an independent benchmark, and Nvidia does not say which workloads or what the multi-node baseline was.
The confidential computing half
The same post carries a second announcement that got less attention. 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, powered by Nvidia GB300 NVL72. Separately, Confidential Computing is now generally available — properly, this time — on Google Cloud G4 VMs powered by Nvidia RTX PRO 6000 Blackwell Server Edition GPUs. Nvidia names Vertebrae, Polaris Cloud and Confidential Core AI as customers using Confidential G4.
This matters for the same reason the capacity numbers do. Companies renting frontier compute increasingly run proprietary weights on hardware they do not own, and rack-scale confidential computing is the mechanism being built to make that acceptable to a legal department.
Nvidia also says it and Google Cloud are exploring Nvidia DSX MaxLPS for Vera Rubin infrastructure, to improve performance per watt and fit more capacity inside existing data-centre power budgets — the constraint that now governs most of these deployments.
What to watch is the November availability date, and whether Google Cloud puts a number on installed Rubin capacity rather than planned capacity.
