What does Google Cloud's Vera Rubin infrastructure mean for AI training at scale?

What the A5X instance actually contains

Each A5X instance combines NVIDIA Vera Rubin NVL72 systems with NVIDIA ConnectX-9 SuperNICs. Google Cloud's Virgo Network links the racks. The fourth-generation advanced liquid-cooling technology handles the thermal load. This integrated platform targets training and inference workloads. It also supports large-scale reinforcement learning.

General availability starts next month. Google Cloud plans to scale Vera Rubin-based infrastructure to as many as 960,000 Rubin GPUs across multiple sites. The European expansion follows the initial launch.

How reinforcement learning changes with this hardware

Ineffable Intelligence was the first customer to access A5X. The London-based technology company was founded by AlphaGo creator David Silver. He serves as its CEO. The company advances large-scale reinforcement learning to shape what it calls superlearner systems that learn continuously from experience.

Learning through interaction requires a continuous cycle. The cycle involves generating experience, then evaluating outcomes, then updating models. A5X brings Vera Rubin into Google Cloud's AI Hypercomputer as an integrated system. It combines compute and networking. Storage is also part of the package to support experience-driven learning at scale.

"We're building superlearners that continually improve through their own experience," said David Silver. "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. 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."

WPP has also demonstrated its creative agent reference architecture on A5X instances. The London-based marketing and communications company used NVIDIA Nemotron models and NVIDIA OpenShell. The agents turn conversational requests into multistep creative workflows. These workflows include video generation and upscaling. Restyling happens with human approval before execution. OpenShell provides governed sandboxes for agent execution. It helps control access to sensitive brand assets and creative tools. In WPP's initial tests, workloads that typically required a multi-node cluster ran on a single A5X instance.

Confidential computing at rack scale

As organizations deploy agentic systems, they need to protect proprietary models and sensitive data even when running on infrastructure they don't own. Google Cloud will be among the first cloud providers to offer NVIDIA Confidential Computing at rack scale on bare metal instances. The A4X Max instances are powered by NVIDIA GB300 NVL72.

Additionally, Google Cloud is making Confidential Computing generally available on G4 VMs powered by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. This offers a flexible platform for confidential inference and fine-tuning. Customers include Vertebrae and Polaris Cloud. Confidential Core AI is also a customer. They use it for conversational privacy and verifiable agent execution. Policy-governed enterprise intelligence is another use case.

Energy efficiency and deployment limits

To support energy-efficient deployment at scale, Google Cloud and NVIDIA are exploring NVIDIA DSX MaxLPS capabilities for Vera Rubin-based infrastructure. The goal is to optimize performance per watt. Another goal is to expand capacity within available data-center power budgets. The announcement does not specify what power budget constraints exist at the planned European sites. It also does not say when the 960,000 GPU target might be reached.