Google Cloud, part of Alphabet Inc. (Nasdaq: GOOGL), has outlined new infrastructure and Kubernetes capabilities intended to help enterprises train, serve and operate multi-step AI workloads at larger scale. The updates include A5X instances based on NVIDIA’s Vera Rubin platform, new scale-out networking and workload-management features in Google Kubernetes Engine.
The company said the A5X bare-metal GPU instances will become generally available next month. They are designed for training and inference workloads that need large clusters, while Google Cloud’s Virgo Network is intended to link those clusters across buildings and sites. Google Cloud said the network can support clusters of up to 80,000 Rubin GPUs at one site and 960,000 across multiple sites; those capacity figures are company specifications.
For IT teams, the practical change is not only more accelerator capacity. The provider is pairing the systems with tools to coordinate distributed work. Google Kubernetes Engine will support A5X deployments with Managed DRANET, which treats network interfaces as Kubernetes resources for RDMA management, and Kueue, a scheduler designed to place distributed workloads according to their network topology and resource needs.
That combination addresses a recurring operational issue in AI infrastructure: a model job can be delayed or underperform when its compute, networking and storage allocations are not coordinated. Google Cloud said its orchestration features are meant to simplify management of those distributed resources for teams running agentic workloads, whose multi-step requests can create sustained inference and context-memory demands.
Storage and memory changes are also part of the A5X design. Managed Lustre will support NVIDIA GPUDirect Storage on the instances, giving data a direct RDMA path from the parallel file system to GPU memory. A5X systems will also support KV caching that reuses context across storage tiers. Google Cloud said that capability is intended to reduce redundant computation and improve inference efficiency for long-context and multi-turn workloads.
The company also announced several controls for organizations operating sensitive AI workloads. Confidential G4 GPU instances are generally available, while confidential computing is available on A4 instances. Google Cloud said it is also introducing confidential bare-metal capability for selected A4X Max systems, using hardware security features intended to protect models and prompts during training and inference.
For on-premises and disconnected settings, Google Cloud said NVIDIA HGX B300 systems are generally available on Google Distributed Cloud, alongside Gemini 3.5 Flash support. The provider is also previewing an enterprise-search capability for air-gapped deployments. Those options are aimed at organizations that need to keep AI processing within a local security boundary rather than send sensitive data to a public cloud service.
The immediate benefit for platform teams is a more integrated set of compute, networking, storage and scheduling components for large AI workloads. The infrastructure will still require capacity planning and application-level controls, but the additions give organizations more tools to place distributed AI work and protect sensitive workloads without assembling every layer themselves. Google Cloud did not disclose pricing for the upcoming A5X instances in its post. The company said the availability schedule may vary by region and configuration.