Google Cloud details Vera Rubin systems and sovereign AI tools

Google Cloud said A5X instances are bare-metal systems built on NVIDIA’s Vera Rubin NVL72 platform.[1] The company expected the instances to reach general availability the month after its October 21 announcement.[1] NVIDIA said AI company Ineffable Intelligence was the first customer to access A5X and is using it for large-scale reinforcement learning.[3] WPP’s initial tests ran workloads that typically needed a multi-node cluster on a single A5X instance.[3] These examples offer early indications of how customers are evaluating the system, rather than a broad performance benchmark.

Google’s Virgo Network uses a mesh topology to provide a non-blocking scale-out fabric between and within buildings.[1] Google says it can connect clusters of up to 80,000 Rubin GPUs at one site, or up to 960,000 across multiple sites.[1] The company also described a direct RDMA path from a parallel file system into GPU memory.[1] Its high-speed caching layer is designed to reuse context across storage tiers and reduce redundant computation.[1] Google Kubernetes Engine is also adding features for distributed AI workloads.[1]

Google Cloud plans to expand its infrastructure footprint into new local EMEA regions to support data-residency and sovereignty requirements.[1] Its Cloud Data Boundary can enforce regional residency controls for sovereign AI workloads on the public cloud.[1] Google Cloud Dedicated provides an isolated cloud operated by local partners, including S3NS in France and Thales in Germany.[1] For air-gapped Google Distributed Cloud, organizations can run high-throughput multimodal reasoning within their perimeter.[1] Enterprise search for those deployments is in preview; Google says indexing, embeddings and interactions stay within the sovereign security boundary.[1]

Confidential Computing on Google Cloud G4 instances with NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs is generally available across 36 zones.[2] Google says its tests found less than 3.5% inference-throughput degradation compared with standard G4 instances.[2] The company says Confidential Computing is available to select customers on A4 instances powered by NVIDIA HGX B200 and Intel TDX.[1] Google also says bare-metal AI workloads use a trusted execution boundary to protect prompts and models during inference and training.[1] Customers can run and govern open-weight models while retaining custody of model weights and runtimes.[1]