By Hamza Faisal
Nvidia's Vera Rubin chips are about to go mainstream on Google Cloud. The two companies said Wednesday that A5X instances built on Nvidia's Vera Rubin NVL72 systems will be generally available next month, with a planned expansion to Europe — and that Google Cloud intends to scale its Rubin infrastructure to as many as 960,000 GPUs across multiple sites. The announcement was timed to Nvidia's GTC Berlin keynote.
Each A5X instance packs Vera Rubin NVL72 racks together with Nvidia ConnectX-9 SuperNICs, Google Cloud's Virgo Network fabric and Google's fourth-generation liquid-cooling technology. The systems slot directly into Google Cloud's AI Hypercomputer as an integrated stack of compute, networking and storage, aimed at AI training, agentic inference and large-scale reinforcement learning.
The first customer is an AlphaGo creator's new lab
The first customer through the door is Ineffable Intelligence, the London startup founded by AlphaGo creator David Silver, which raised a $1.1 billion seed round in April to build AI systems that learn from experience rather than human data. It is using A5X to develop what it calls "superlearner" systems that improve continuously through interaction.
"We're building superlearners that continually improve through their own experience," said Silver, founder and CEO of Ineffable Intelligence. "That demands both substantial compute and a tight connection between generating experience and training on it. Vera Rubin's combination of compute performance and high-bandwidth connectivity — between the Vera CPU and Rubin GPUs and between the Rubin GPUs within the rack — allow us to push the scale and ambition of our reinforcement learning research further."
Reinforcement learning is one of the most infrastructure-hungry workloads in AI: a model generates experience, scores the results and updates itself in a loop, and every turn of that loop needs the link between experience generation and training to be as fast as possible.
One instance did the work of a whole cluster
The second early customer is a very different kind of company. Marketing giant WPP tested its creative AI agents on A5X using Nvidia's Nemotron models and the OpenShell reference architecture, and said workloads that normally needed a multi-node cluster ran on a single A5X instance — speeding up creative-agent pipelines such as video generation and upscaling.
Google Cloud is also pushing security as a differentiator. It will be among the first providers to offer rack-scale Nvidia Confidential Computing on bare-metal A4X Max instances powered by GB300 NVL72 systems, with the technology already generally available on G4 virtual machines running RTX PRO 6000 Blackwell server GPUs. Customers named for the confidential-computing push include Vertebrae, Polaris Cloud and Confidential Core AI.
The bigger cloud GPU race
The announcement underscores how quickly Rubin is moving from launch to broad cloud availability — CoreWeave rolled out its own Vera Rubin systems earlier this month, and Google's 960,000-GPU target would be one of the largest single-architecture cloud GPU commitments disclosed by any hyperscaler. Both companies are also looking at Nvidia's DSX MaxLPS tooling to keep performance-per-watt in check as the infrastructure scales.
Markets took the news calmly: Nvidia shares gained nearly 2% in pre-market trading after closing at $216.47 on Tuesday, while Alphabet slipped about 1% after closing at $336.40, according to dpa-AFX. With the general availability window set for November and European expansion to follow, the real test will be how fast those 960,000 GPUs actually come online. By Hamza Faisal



