Google Cloud and NVIDIA launch Vera Rubin infrastructure for AI

The A5X Infrastructure

Google Cloud and NVIDIA have announced the launch of A5X instances, a new hardware configuration designed for heavy AI workloads. These instances integrate the NVIDIA Vera Rubin architecture into Google Cloud’s AI Hypercomputer, which is a unified system that combines compute, networking, and storage.

The hardware is built to handle reinforcement learning, a type of machine learning where an AI learns to make decisions by interacting with an environment to achieve a goal. To do this effectively at scale, the system requires a tight connection between generating new experiences and training on them. The A5X architecture provides the high-bandwidth connectivity needed to keep these cycles moving quickly.

To manage the intense heat generated by these chips, Google Cloud is using its fourth-generation advanced liquid-cooling technology. This cooling is essential for maintaining performance in data centers as AI models become more demanding. The setup also includes NVIDIA ConnectX-9 SuperNICs and Google Cloud Virgo Network to ensure data moves quickly between components.

Real-world AI applications

Two different companies are already using the A5X platform to test different types of AI capabilities. Ineffable Intelligence, a London-based company founded by David Silver (the creator of AlphaGo), was the first customer to access the hardware. They are using it to develop 'superlearners'—systems that learn continuously from experience rather than just being trained on a static dataset.

David Silver noted that the Vera Rubin hardware allows them to push the scale of their research because of the performance between the CPU and GPUs. This allows the AI to quickly shape its next experience based on the improvements it makes in the current one.

On the creative side, the marketing firm WPP has demonstrated a reference architecture for safe AI agents. These agents can take a conversational request and turn it into a multi-step creative workflow, such as upscaling or restyling video. These workflows include a human approval step before the final execution. WPP reported that workloads which usually required a multi-node cluster were able to run on a single A5X instance during their initial tests.

Security and scaling

As companies move toward using these autonomous agents, protecting sensitive data becomes a priority. Google Cloud plans to be among the first to offer NVIDIA Confidential Computing at rack scale on bare metal instances. This technology allows organizations to run proprietary models on infrastructure they do not own while keeping the data private.

The company is also making Confidential Computing available on G4 VMs, which are powered by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. This provides a flexible way for businesses to perform private inference and fine-tuning. Several companies, including Vertebrae and Polaris Cloud, are already using these for conversational privacy and policy-governed intelligence.

Looking ahead, Google Cloud and NVIDIA are exploring DSX MaxLPS capabilities to optimize performance per watt. This focus on energy efficiency is critical for expanding capacity within existing data-center power budgets. The goal is to make it possible to deploy Vera Rubin-based infrastructure at a massive scale without exceeding power limits.

Availability and expansion

The A5X instances are scheduled for general availability next month. Google Cloud plans to expand the availability of these instances to Europe, providing an integrated platform for training, inference, and large-scale reinforcement learning to its cloud customers.

The infrastructure is designed to be highly scalable. Google Cloud intends to scale the Vera Rubin-based systems to as many as 960,000 GPUs across multiple sites. This massive scale is intended to support the next generation of agentic systems that require constant interaction and evaluation to improve over time.

Technical specifications

The A5X configuration is a sophisticated stack of hardware components. It relies on the Vera Rubin architecture, which is NVIDIA's next-generation GPU platform. By combining these with Google's proprietary networking and cooling, the system can handle agentic inference—where an AI agent performs a series of actions to complete a goal—at a much larger scale than previous hardware allowed.

The integration ensures that the bottleneck of data movement between the GPU and the rest of the system is minimized. This is particularly important for the 'experience-driven learning' mentioned by Ineffable Intelligence, where the speed of the feedback loop determines how quickly the AI can improve. By providing a single instance that can handle multi-node workloads, Google and NVIDIA are making it easier for developers to deploy these complex systems without managing complex distributed hardware setups.