Google Cloud and NVIDIA unveil Vera Rubin AI systems

Next-generation hardware for AI

Google Cloud and NVIDIA have announced a new suite of computing systems designed to handle increasingly complex enterprise AI workloads. A major highlight of the announcement is the upcoming A5X bare-metal instances, which will be powered by NVIDIA's Vera Rubin NVL72 platform. These systems are intended to provide the heavy lifting required for large-scale model training and inference.

The infrastructure combines GPU acceleration with specialized networking and storage technologies. Specifically, the partnership is integrating NVIDIA GPUDirect Storage through Google Cloud Managed Lustre. This allows data to move directly into GPU memory, which helps speed up the processing of massive datasets. These hardware improvements are essential for organizations moving toward agentic AI systems that can perform multi-step tasks.

The new Vera Rubin infrastructure is built for massive scale. Google's Virgo Network is designed to accommodate groupings of as many as 80,000 Rubin GPUs at one location. While these figures represent the intended scaling capacity rather than an immediate operational deployment, they signal the high ceiling for future AI computing power. The systems are scheduled for general availability in November 2026.

Sovereign AI and local control

A significant portion of the update focuses on sovereign AI capabilities. These tools are designed for organizations that must operate AI models within specific geographic jurisdictions or isolated computing environments. By offering these options, Google and NVIDIA are addressing the needs of entities with strict regional data residency and sovereignty requirements.

Google Cloud is expanding its Sovereign Cloud portfolio to include support for open-weight models. This includes popular models such as Gemma, Mistral, and NVIDIA Nemotron. By supporting open-weight models, the platform allows organizations to retain full custody of the model weights and the runtime environments where the AI actually operates.

The partnership also includes the availability of NVIDIA Blackwell Ultra GPUs through Google Distributed Cloud. This allows for the use of NVIDIA HGX B300 systems for computing deployments within an organization's own physical facilities. This setup ensures that multimodal AI processing, including text, audio, images, and video, happens entirely inside the user's controlled environment.

Security and confidential computing

To protect sensitive information during AI training and usage, the companies are introducing several security updates. Confidential computing is a method of protecting data while it is being processed by the CPU or GPU. Google Cloud is now offering Confidential G4 instances, which are powered by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs.

Furthermore, Confidential Bare Metal is being introduced on A4X Max instances. These instances utilize NVIDIA GB300 NVL72 technology and hardware-based security measures. These features are specifically designed to protect intellectual property, user prompts, and the underlying AI models from unauthorized access during active processing.

These security layers are aimed at organizations managing highly sensitive workloads. For these users, data protection and infrastructure isolation are non-negotiable requirements. By integrating these protections into the hardware level, Google and NVIDIA are attempting to make high-performance AI safer for enterprise and government use cases.

Enterprise search and robotics

The update also introduces enterprise search features powered by Gemini and multimodal retrieval-augmented generation, or RAG. This technology is currently available in preview for air-gapped Google Distributed Cloud deployments. It allows organizations to search, summarize, and examine internal documents and media while keeping all processing inside their own controlled environment.

Beyond standard text and data search, Google is adding compatibility for NVIDIA Cosmos 3 and Nemotron 3.5 SuperVL to its Gemini platform. These tools are specifically designed for model customization and the creation of simulated environments. These environments are frequently used in the development of robotics and autonomous vehicle technologies.

Industry leaders are already looking at how these tools can be applied to creative production. Stephan Pretorius, Chief Technology Officer at WPP, noted that the infrastructure could help teams direct intelligent agents to perform quality checks. These agents could adapt brand assets seamlessly, showing how the underlying hardware can be applied to specific industrial workflows.