Google Cloud is preparing to put NVIDIA's next rack-scale system into service. In a post published for GTC Berlin on October 21, 2026, NVIDIA said Google Cloud's A5X instances, built on Vera Rubin NVL72, will reach general availability next month, with availability in Europe planned afterwards. The headline speaks of general availability, but the body places it in the coming month, so customers should treat A5X as imminent rather than open today.
Each A5X system combines a Vera Rubin NVL72 rack with NVIDIA ConnectX-9 SuperNICs, Google's Virgo data-center network and the fourth generation of Google's liquid cooling. Both companies pitch it at agentic inference and large-scale reinforcement learning, workloads that repeatedly generate experience, score it and update a model. NVIDIA says the instances slot into Google's AI Hypercomputer as one integrated system spanning compute, networking and storage.
The scale target is the striking number. According to NVIDIA, Google Cloud intends to expand its Vera Rubin fleet over several sites until it reaches up to 960,000 Rubin GPUs. Neither company gave a timetable for that build-out or said how much of it will sit in Europe, and the figure describes an intention rather than installed capacity.
Two early users are named. Ineffable Intelligence, the London company founded by David Silver, the researcher behind AlphaGo, was the first to get access to A5X and is using it for reinforcement learning research aimed at what Silver calls superlearners that improve from their own experience. He credits the bandwidth between the Vera CPU and the Rubin GPUs, and among GPUs inside the rack, for letting the team push its experiments further. WPP, the marketing group, ran a creative-agent reference design on A5X using Nemotron models and NVIDIA's OpenShell sandbox, and says jobs that usually needed a multi-node cluster fit on a single instance in its initial tests.
Security for models and data running on rented hardware is the other thread. NVIDIA says Google Cloud will be one of the earliest clouds to bring its confidential computing to whole racks on bare-metal machines such as A4X Max, which uses GB300 NVL72. Separately, Google announced that Confidential G4 virtual machines, powered by RTX PRO 6000 Blackwell Server Edition GPUs, are now generally available across 36 zones. Google's own tests show less than 3.5% throughput loss against standard G4 machines on models including Llama 3 and SDXL, a result it notes depends on configuration.
The companies also say they are exploring NVIDIA DSX MaxLPS for the Vera Rubin fleet, a way to squeeze more performance per watt out of fixed data-center power budgets. No results were given.
For AI labs and enterprises, the announcement signals when Vera Rubin capacity becomes rentable on a major cloud and how Google intends to compete for frontier training and reinforcement-learning work. Pricing, regional details and independent performance data were not published, and the customer results come from the companies and their partners.