Google Cloud and NVIDIA expand co-engineering collaboration to scale infrastructure for the agentic era

Google Cloud and NVIDIA unveiled new advancements to support agentic AI workflows, addressing performance bottlenecks and operational sovereignty. The collaboration introduces A5X instances based on NVIDIA’s Vera Rubin platform, designed for bare-metal GPU acceleration and general availability next month. New infrastructure will expand to EMEA regions to meet data residency and sovereignty requirements, ensuring compliance with local regulations.

The updates include scale-out networking via Virgo Network, enabling clusters to expand to 80,000 Rubin GPUs within a single site and up to 960,000 across multi-site deployments. Google Cloud Managed Lustre will support NVIDIA GPUDirect Storage on A5X instances, providing a direct RDMA path for training and inference data. Advanced KV-caching capabilities will also be introduced to scale context memory for long-context and multi-turn agentic workloads.

Google Kubernetes Engine (GKE) will now support A5X instances, offering features like GKE Managed DRANET for simplified RDMA management and Kueue for topology-aware scheduling. This combination aims to enhance workload orchestration for distributed AI tasks. The updates also include support for open models like Gemma and Mistral in the Sovereign Cloud portfolio, allowing organizations to deploy and govern models while avoiding vendor lock-in.

New sovereign AI capabilities for on-premises environments and advancements in Confidential Computing for public cloud were announced. Google Distributed Cloud (GDC) now offers NVIDIA Blackwell Ultra GPUs and Gemini Flash models in air-gapped deployments. Confidential G4 instances and Confidential Bare Metal for A4X Max instances were also introduced to provide hardware-enforced security for sensitive AI workloads.