NVIDIA and Google Cloud have expanded their long-running collaboration to extend a co-engineered full-stack AI platform that spans performance libraries, frameworks, and enterprise cloud services. The update, announced at Google Cloud Next in Las Vegas, focuses on expanding Google Cloud AI Hypercomputer capabilities to support production-scale agentic AI and physical AI workloads, including robotics, simulation, and industrial digital twins.
The release introduces new infrastructure and software layers aimed at moving agentic systems from experimentation into deployment. Key additions include NVIDIA Vera Rubin-powered A5X bare-metal instances based on NVL72 rack-scale systems, preview availability of Google Gemini models running on Google Distributed Cloud with NVIDIA Blackwell and Blackwell Ultra GPUs, confidential virtual machines using NVIDIA Blackwell GPUs, and expanded agentic AI capabilities on the Gemini Enterprise Agent Platform integrating NVIDIA Nemotron open models and the NVIDIA NeMo framework.
At the infrastructure layer, the A5X systems are positioned as a next-generation AI factory foundation. Built on tightly co-designed hardware and software, they are designed to deliver up to 10x lower inference cost per token and up to 10x higher token throughput per megawatt compared to prior-generation systems. The platform scales across large clusters using NVIDIA ConnectX-9 SuperNICs and Google Virgo networking, supporting deployments ranging from single-rack systems to multi-site clusters with hundreds of thousands of GPUs.
Google Cloud’s NVIDIA Blackwell portfolio spans multiple compute tiers, from large-scale NVL72 systems to fractional GPU instances, enabling workload-specific scaling across training, inference, and simulation. This includes support for mixture-of-experts models, multimodal inference, and physically grounded simulations used in robotics and industrial AI workflows.
Security and sovereignty capabilities are extended through confidential computing. Gemini models running on NVIDIA Blackwell infrastructure are now available in preview on Google Distributed Cloud, while Confidential G4 VMs bring encrypted execution environments to multi-tenant cloud deployments, ensuring data, prompts, and models remain protected even from infrastructure operators.
On the software and model layer, NVIDIA Nemotron 3 Super is integrated into the Gemini Enterprise Agent Platform alongside NVIDIA NeMo tools, enabling managed reinforcement learning workflows for training and customizing agentic systems at scale. Managed Training Clusters automate infrastructure provisioning and recovery, while customers such as CrowdStrike apply NeMo libraries for synthetic data generation and domain-specific model tuning in cybersecurity workflows.
Industrial and physical AI capabilities are expanded through NVIDIA Omniverse libraries, Isaac Sim, and Cosmos Reason models available via Google Cloud services, including Vertex AI and Google Kubernetes Engine. These tools support digital twins, robotics simulation, and real-world automation pipelines used by industrial software providers.