Google Cloud and NVIDIA set general availability for A5X, a Vera Rubin-based platform for training and agentic reinforcement learning

NVIDIA and Google Cloud said they are moving NVIDIA Vera Rubin infrastructure into general availability through a new cloud offering called A5X instances. The announcement frames A5X as a tightly integrated compute, networking, and storage platform meant for the full loop of agent development, including model training, inference, and large-scale reinforcement learning, where systems must repeatedly generate experience, evaluate outcomes, and update models.

The companies said A5X instances combine NVIDIA Vera Rubin NVL72 GPUs with NVIDIA ConnectX-9 SuperNICs, Google Cloud’s Virgo Network, and Google Cloud’s fourth-generation advanced liquid cooling technology. NVIDIA and Google Cloud positioned that combination as particularly important for “agentic inference” and reinforcement learning workloads that require high bandwidth between components rather than isolated bursts of compute.

In the same release, Google Cloud said it plans to scale its Vera Rubin-based infrastructure to as many as 960,000 Rubin GPUs across multiple sites. The stated ramp is intended to give cloud customers a single place to run training and inference while also supporting reinforcement learning at scale, where data produced during interaction with an environment must quickly feed back into training.

Ineffable Intelligence, a London-based firm founded by David Silver, the creator of AlphaGo, was described as the first customer with access to A5X. The company said it is using the platform to advance large-scale reinforcement learning and to build “superlearner” systems, which aim to continually improve through their own experience and adjust behavior based on what happens next.

Silver was quoted saying that building these systems requires both “substantial compute” and a “tight connection” between generating experience and training on it. He also pointed to Vera Rubin’s combination of compute performance and high-bandwidth connectivity between the Vera CPU and Rubin GPUs, and between Rubin GPUs within a rack, as a way to push the scale and ambition of reinforcement learning research further.

Beyond reinforcement learning, the release also highlighted use cases for enterprise agent workloads that connect conversational requests to multi-step actions. WPP, the London-based marketing and communications company, demonstrated a “creative agent reference architecture” on Vera Rubin-based A5X instances using NVIDIA Nemotron models and NVIDIA OpenShell, part of NVIDIA’s OpenShell Reference Architecture for safe AI agents.

According to the companies, WPP’s tests mapped conversational prompts to workflows such as video generation, upscaling, and restyling, with human approval before execution. The release said OpenShell provides “governed sandboxes” for agent execution, intended to help control access to sensitive brand assets and creative tools, and that some workloads typically requiring a multi-node cluster ran on a single A5X instance in initial testing.

The companies also connected the infrastructure push to security needs as organizations move from controlled pilots to production deployment. They said organizations deploying agentic systems must protect proprietary models and sensitive data even when models run on infrastructure they do not own, and they pointed to Confidential Computing as one approach.

Google Cloud said it 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, it said Confidential Computing is becoming generally available on G4 virtual machine instances powered by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, with named customers including Vertebrae, Polaris Cloud, and Confidential Core AI.

Google Cloud and NVIDIA added that they are exploring NVIDIA DSX MaxLPS capabilities for Vera Rubin-based infrastructure to optimize performance per watt, an attempt to expand capacity within data-center power constraints. Next month’s general availability and the planned expansion to Europe are the main timing markers provided, but pricing, detailed availability windows by region, and specific performance targets were not disclosed in the announcement.

From a business perspective, the move is a announcement that “agentic” AI systems are driving a shift from isolated GPU capacity toward integrated, tightly coupled cloud platforms. If reinforcement learning workloads remain bandwidth-hungry and iteration-heavy, customers may increasingly prefer infrastructure that reduces data movement and shortens the loop between experience generation and training.

For NVIDIA, the A5X launch extends its Vera Rubin platform beyond chip shipments into complete infrastructure bundles alongside networking and cooling. For Google Cloud, it positions its AI Hypercomputer offerings as a place where customers can combine training, inference, and agent learning under one operational model, while the Confidential Computing references aim to address enterprise security requirements as deployments scale.